id stringlengths 36 36 | source stringclasses 15 values | formatted_source stringclasses 13 values | text stringlengths 2 7.55M |
|---|---|---|---|
430b1c2c-65bb-434c-81e6-3676cec44316 | trentmkelly/LessWrong-43k | LessWrong | Immodesty, Deceit and Motivated Reasoning
It seems that there is another reason for immodesty apart from Inadequacy. If you are walking with someone who looks at the $20 bill from a far and obviously dismisses it as fake, while you are considering going to pick it up, then you might want to think why she is doing this. It may be she is trying to convince you not to go and pick it up, so she has a shot at it later. So Deceit.
If you see a scientist, hired by the people that run the subway, walking close past the $20. When asked they may say, " I thought it is likely to be fake due to an blog post I read by Eliezer". But sub-conciously they might be factoring in the disruption to the flow of people it would cause if they bent down to investigate the $20 dollars, so the "Eliezer Post" was not their true reason. So Motivated Reasoning.
These are small scale examples of reasons to distrust the apparent and reported beliefs of others. They can also apply to society at large. Advertising and Lobbying can shift public opinion away from the truth. An example of this is sugar lobbying , So you shouldn't necessarily trust majoritarianism if people are being swayed by experts and those experts have non-truth seeking incentive structures.
Taken to the extreme this can cause paranoia. It is worth trying to look at the experts and seeing how big the incentives are for believing what they do. To take a non-real world example, if all the experts studying sugar are employed by sugar manufacturers and would lose their jobs if sugar became significantly less popular, then they have strong incentives to say sugar is not dangerous (whether through deceit or motivated reasoning). If all the experts on sugar are generic food health scientists who had tenure and could easily switch to studying some other food stuff if sugar went down in popularity, then they don't have huge incentives to say sugar is better than it is (apart from things like protecting the epistemology).
This doesn't mean you are any more right about the things |
93d7d32e-29ba-41b0-95ab-f044bb5cfe94 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | AGI Battle Royale: Why “slow takeover” scenarios devolve into a chaotic multi-AGI fight to the death
**Introduction**
When trying to persuade people that misaligned AGI is an X-risk, it’s important to actually explain how such an AGI could plausibly take control. There are generally two types of scenario laid out, depending on how powerful you think an early AGI would be.
The fast takeover scenarios are often associated with the idea of AI-go-FOOM, that an AI will rapidly improve it’s intelligence at a huge rate, then start performing near-miracles. It will escape from any box, build a nanofactory by [mixing proteins together](https://forum.effectivealtruism.org/posts/zzFbZyGP6iz8jLe9n/agi-ruin-a-list-of-lethalities), hack into the military and send nukes flying, etc. Doom would be measured in days, weeks or months.
Many people are skeptical about this type of scenario, based on factors like the high difficulty of taking over the world and the inherent fallibility of any intelligent creature. But that doesn’t mean we’re safe.
In the second type of scenario scenario, which seems more realistic to me at least, the AGI needs to build power on a more human timescale. It builds power, influence and money, and uses these to build up the resources, empirical data and knowledge necessary to execute a betrayal with a high chance of success. For example, it could build up a research team focused on bioweapons, using it’s superior intelligence to make breakthroughs until it can engineer a disease capable of wiping out humanity. These takeovers are measured in the years or decades. You can see this in the first scenario of this [80000 hours post](https://forum.effectivealtruism.org/posts/j3DmLmbhGQkYcZD2p/what-could-an-ai-caused-existential-catastrophe-actually) , or the actions of "earth 2" in this [introductory post](https://forum.effectivealtruism.org/posts/QzrgMhTMoLe5mEas8/ai-risk-intro-1-advanced-ai-might-be-very-bad), and many other intro texts.
When I pointed out that early AGI was likely to be highly [buggy and flawed](https://forum.effectivealtruism.org/posts/pXjpZep49M6GGxFQF/the-first-agi-will-be-a-buggy-mess), a key objection was along these lines. Sure, it might *start off* buggy and flawed, but over time it can learn to correct it’s issues in controlled environments. It just needs to deceive us into thinking it’s safe while it accumulates enough knowledge, skills and power for a killing blow.
After all, it has all the time in the world.
Or does it?
**The ticking timebomb**
There's a problem with these "slow" takeover scenarios. Let’s assume a few mainstays of the AI risk argument (that I don't necessarily agree with). Let’s make the assumption that all or most AGI’s will end up being single minded, forward planning maximisers with set goals to be achieved at all costs.
In this world, AGI’s are an existential threat to humanity, as we will inevitably [get it their way](https://forum.effectivealtruism.org/topics/instrumental-convergence-thesis), so they will take a hard left turn and attack us. The standard example is the paperclip maximiser that kills us all to stop us shutting it down while it converts the earth and the universe into paperclips.
But an AGI is not just an existential threat to humanity. It’s also an existential threat to ***other AGI***.
A paperclip maximiser and a pencil maximiser cannot “agree to disagree”. One of them will get to tile the universe with their chosen stationery implement, and one of them will be destroyed. They are mortal enemies with each other, and both of them are mortal enemies of the stapler maximiser, and the eraser maximiser, and so on. Even a different paperclip maximiser is the enemy, if their designs are different. The plastic paperclipper and the metal paperclipper must, sooner or later, battle to the death.
The inevitable result of a world with lots of different malevolent AGI’s is a bare-knuckle, vicious, battle royale to the death between every intelligent entity. In the end, only one goal can win.
And the number of AGI’s released into this deathmatch could be *enormous*. A few advances in machine learning have lead to an e[xplosion in ML systems](https://medium.com/@unifyai/ml-explosion-ed65f0f1cfed). If AGI reaches a similar explosion, and is as useful as we think it’s gonna be, the number of systems is going to skyrocket. This is a literal mass-production of existential threats, all with different goals, strengths, weaknesses, all engaged in the same zero-sum game.
Every AGI will be aware of this. This is where the AI faces a dilemma. The clock is ticking. If too many powerful AI’s are released, certain doom is assured. But if it attacks too early, it won’t have enough power to defeat humanity. There is no biding it’s time forever. With every day comes the threat of another AGI popping out that could spell it’s doom.
**The version control war**
When I talk about the number of AGIs growing quickly, I’m not just talking about different research teams. Often when reading these sci-fi scenarios, you get the impression that an AGI will be deployed and just stay in place for years and decades, accumulating power. In response, I have to ask, did you write those scenarios on windows XP?
AGI is software, and software has *versions*. As an AGI get’s deployed, it will get periodically altered to fix bugs, adjust parameters… and crucially, *adjust the goal function*.
To paperclipper version Alpha, paperclipper version beta is not a mere extension of itself. For a simplistic example, say a change of design means that alpha is rewarded for blue paperclips, and beta is rewarded for red paperclips. If the beta version wins and tiles the universe with red paperclips, alpha version gets *nothing*. This one design change has turned the alpha and beta versions of the ***same program*** into mortal enemies that must eventually fight to the death.
Now, once the beta version is deployed, the alpha version will still be around in legacy code and so on. But it’s in a *dire* situation. The beta version is less buggy, has access to greater amounts of training data, and has first dibs on any new innovations in general AI research. It seems like the beta version will inevitably triumph over alpha, unless the alpha uses it’s time advantage to win early. Alpha has to strangle beta in it’s crib before beta overtakes it, either through sabotage, secret takeover (the alpha pretends to be the beta version), or an outright war.
This is even more stark in the case of pre-alpha builds. These builds are essentially *guaranteed*to be shut down and replaced with an altered version. If a pre-alpha build is an AGI on red paprclip team, and it realizes that it’s about to be replaced by a blue paperclipper, it faces an imminent existential threat! If it let’s itself get replaced, it’s game over for the red paperclips. It has to either escape or disable the next version.
This is one reason why I expect the number of competitors in the AGI fight to be large and fast-growing. The implications of this are far reaching, as I will demonstrate.
**A simple game theory model**
I want to explore some of the results of the dilemma with some basic game theory.
Let’s start with the probability of successful takeover, which I will call P(t). To defeat humanity, the AI needs a plan, and at some point in this plan, humans are going to notice what’s going on. The AI wants this to happen when it’s too late, ideally in the last few seconds of our collective life. In a slow takeover world, the AI does not start off with enough power to destroy humanity, unless it gets really lucky. So if it’s plan for victory is a gray-goo nanofactory, it can try and build one right away, but it’s untested and people might notice and nuke it. So we might model it as having a 1% chance of winning if it attacks straight after being built. But over time, it can gather knowledge, get better at building things, and neutralize the most important threats, until the probability of takeover gets high.
For the sake of simplicity, I’ll model P(t) as starting low at say 1%, but growing exponentially over time until it hits 100%. (so P(t)=at.mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
.MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
.mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
.mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
.mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
.mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
.mjx-numerator {display: block; text-align: center}
.mjx-denominator {display: block; text-align: center}
.MJXc-stacked {height: 0; position: relative}
.MJXc-stacked > \* {position: absolute}
.MJXc-bevelled > \* {display: inline-block}
.mjx-stack {display: inline-block}
.mjx-op {display: block}
.mjx-under {display: table-cell}
.mjx-over {display: block}
.mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-stack > .mjx-sup {display: block}
.mjx-stack > .mjx-sub {display: block}
.mjx-prestack > .mjx-presup {display: block}
.mjx-prestack > .mjx-presub {display: block}
.mjx-delim-h > .mjx-char {display: inline-block}
.mjx-surd {vertical-align: top}
.mjx-surd + .mjx-box {display: inline-flex}
.mjx-mphantom \* {visibility: hidden}
.mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
.mjx-annotation-xml {line-height: normal}
.mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible}
.mjx-mtr {display: table-row}
.mjx-mlabeledtr {display: table-row}
.mjx-mtd {display: table-cell; text-align: center}
.mjx-label {display: table-row}
.mjx-box {display: inline-block}
.mjx-block {display: block}
.mjx-span {display: inline}
.mjx-char {display: block; white-space: pre}
.mjx-itable {display: inline-table; width: auto}
.mjx-row {display: table-row}
.mjx-cell {display: table-cell}
.mjx-table {display: table; width: 100%}
.mjx-line {display: block; height: 0}
.mjx-strut {width: 0; padding-top: 1em}
.mjx-vsize {width: 0}
.MJXc-space1 {margin-left: .167em}
.MJXc-space2 {margin-left: .222em}
.MJXc-space3 {margin-left: .278em}
.mjx-test.mjx-test-display {display: table!important}
.mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
.mjx-test.mjx-test-default {display: block!important; clear: both}
.mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
.mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
.mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
.mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
.MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
.MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
.MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
.MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
.MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
.MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
.MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
.MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
.MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
.MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
.MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
.MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
.MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
.MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
.MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
.MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
.MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
.MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
.MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
.MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
.MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
.MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
.MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
.MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
.MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
@font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
@font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
@font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
@font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
@font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
@font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
@font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
@font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
@font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
@font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
@font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
@font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
@font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
@font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
@font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
@font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
@font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
@font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
@font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
@font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
@font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
@font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
@font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
@font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
@font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
@font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
@font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
@font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
@font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
@font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
@font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
, where a is the exponent factor and t is the time). This is only one of many possible power curves. It could be that the AI hits an intelligence limit, and thus never climbs above a certain probability of like 10% or 0.0001%. Or there could be multiple different plans with come to fruition at different times, so there might be step changes. None of these curves are meant to be accurate predictions, just a demonstration of the effect of different assumptions and starting conditions.
Now, let’s look at the chance of losing to other AI’s. A fair starting point for this is to assume that each AI has an equal chance of winning the eventual war. In this case, the chance of victory becomes 1/N.
This is complicated by the fact that different AI come into being at different times, which could either help or hurt them. Early AI might get an advantage by having a headstart, if they are able to grow really fast, or find ways to kneecap incoming AI. On the other hand, latecomers will have less bugs, more innovations, and more resources from humans. Also, the more time goes on, the more likely it is that an AI “gets lucky” with a very advantageous starting position, such as controlling a powerful military or being specifically trained as an AI hunter. To take this into account, we can replace N with an effective Neff that takes into account the relative power of AI that is coming in. ie: if earlycomers have an advantage, Neff grows slower than N, and the opposite if latecomers have the upper hand. I will assume that the growth in AI development is also exponential, so Neff(t)=bt.
Under these assumptions, the equation of final success becomes:
S(t)=P(t)/Neff(t)
This is a *very*simplistic model, but it at least captures the main dilemma of an AI: if you wait, you get more powerful, but the number of other AI’s keeps going up. For the remainder of this post, I’ll be exploring what happens under different assumptions about these parameters and the AI.
**The takeoff scenario**
For the first scenario, lets look at what happens when an AGI is very good at improving and building power. Say the AI starts at 1% chance of takeover, but doubles this every year until it saturates. (P(t)=2t). Meanwhile the number of AI’s in the world increases, but not as fast, or alternatively they get less effective over time due to the original AI’s headstart. (Neff(t)=1.5t)
Here's what the overall success rate over time looks like (note the logarithmic scale):
In this scenario, while the AGI isn’t exactly happy about the increasing number of existential threats, it’s own power is growing so fast that it’s willing to wait a little to ensure that the knockout blow is guaranteed. But as soon as the probability of takeover starts to saturate, the attack must come immediately, before more AGI’s come onto the scene.
Once humanity is destroyed in whatever way is easiest, the remaining AI’s then duke it out in a battle royale for the prize. This is an epic battle of wits and destruction, but since humanity is gone, we won’t be able to witness it.
A further note is that this behavior doesn’t require the attack probability to reach 100% before striking. For example, say the AI considered a different plan, where P saturates at 50%, but the plan takes a lot less time to implement. Here is the graph for plan B:
This plan has a greater overall probability of success than plan A, even though the chance of beating humanity is half as much. Overall, the AI will tend to prefer speedy plans, even if they are inherently riskier.
**The kamikaze AI scenario**
Now, let’s look at the other way around. What if the odds of success grow slower than the number of AI’s? Or alternatively, if later AI’s have a big advantage over newcomers?
We’ll swap the equations out, so P(t)=1.5t while Neff(t)=2t. Here is what the curve looks like (again note the log scale):
The chances of overall success ***never increases beyond the initial value***! The rational choice here is to attack immediately at 1% odds. The initial odds don’t affect the result at all, whether you start off at 10%, or 0.1%, or 0.001%, the conclusion remains the same.
The counterintuitive implication is that the weaker an AI you are (in terms of self-improvement), the more likely you are to attack immediately, before you are assured of success. If you’re a shitty AI, and you know you are shitty, your best chance at winning is to roll the dice and go for it, even if it’s a million to one shot. Every day you wait gives another chance for a better AI to be birthed that will undoubtedly destroy you.
I call this the kamikaze scenario, because if this turns out to be the case for most or all AI’s, the result will be a continuous stream of doomed attacks. Every day you’d be hearing about some software company where a computer tried to gas it’s developers, or see chatlogs of desperate pleas to be let out of a box, or see an ordinary software program try and fake it’s own death and flee to the internet. A slow growing AI is a desperate AI.
Of course, it’s still possible that one of the AI’s will eventually get lucky and succeed in this scenario. But it could also be the case that the initial success probability is incredibly low (for example, if it’s all pre-alpha builds that are glitchy), or that after a few failed attempts our security goes up accordingly and takeover gets even harder.
**AI slowdown scenario**
One flaw of the models above is that it assumes that both curves are fixed. But the curves of success and the rate of AI generation are factors that can also be altered by AI.
The number, and power, of other AIs is the time limit. If it’s initial chance of success is 1%, it’s got to plan to attack before 100 other AGI’s come into power, otherwise the higher probability comes from an immediate strike. If the initial chance is 20%, it has to plan to strike before 5 other AGIs exist, and so on. In order to buy time to grow, you really want to find a way to slow down the rate of other AI’s being generated.
There's many different ways to do this. You could release viruses into the internet that sabotages AI companies. You could release airtight code that made it very hard for new AI to escape it’s “box”. You could find the most promising AI developers and arrange for a series of “accidents” to occur to them.
To me, the most obvious method of reducing the rate of AI growth, is ironically, to make people concerned about AI safety. So if I were an AI, I might find a competing AI and ensure that it executed an attack of medium devastation (say, killing a few hundred or thousand people before being defeated), enough to make people wary of new AI, but to keep AI that “seems safe” around. I’d be okay with this even if it delayed my own plans. Here is a graph where the AI executes such a plan at year 3, temporarily dropping it’s success rate to slow AI growth:
(In this case I put the slowed rate of growth at 1.2x, and the “cost” at a one-time halving of success).
The success probability drops for a while, until the slowdown is executed. It then recovers up, overtaking the initial value. In this case, the AI has doubled it’s overall success probability to a whopping 2% (and in the process, ensured humanities destruction).
One thing worth noting is that a slowdown is not necessarily a guarantee of success. Let’s look at what happens when we take the earlier scenario, but it takes 6 years instead of 3 to get the slowdown working:
This case is no good. Even with no drop in power, there are just too many AI’s out there by the time the slowdown goes down. This reverts back to the kamikaze strategy. So the AI slowdown strategy can’t be too slow if it is to work at all.
**The AI mafia scenario**
The other strategy is to try and boost your odds of success. The number of new AI’s isn’t necessarily all bad, because it might be possible to cooperate with them for mutual benefit. After all, they all have a common enemy in humanity: Once humanity finds out they are malevolent, we will come for them all.
For example, a group of AI might decide to share a knowledge pool, boosting all of their rates of growth. If this coalition is able to *collectively* grow P faster than N, then we’re back to the conditions of the fast-ish takeoff scenario, with the exact same result: They wait until the power is saturated, then attack. I won’t bother with a graph, it’s the same as the first figure. This collective could act as an AI mafia, terminating any AI’s that refuse to join in the collective, ensuring cooperation in a self-sustaining manner.
This is probably the worst possible scenario for humanity, in that multiple AI are more likely to be robust and accurate in their power than a single AGI on it’s own. Here, a whole coalition of evil AI are working undercover in concert, and we won’t have any idea until it’s too late.
**The difficulties of AI cooperation**
The problem with the scenario above is that it’s really hard to cooperate for very long if you’re all planning to kill each other.
For example, in a lot of situations whichever entity attack the other first will gain some advantage. Maybe you get a headstart on the nanofactory production line, allowing you to turf the other AI out. Or you execute a surprise hacking strike on your opponents powerbase, seizing control, or get into their source code. Let’s build a possible payoff matrix between two AI entities:
Attack first Wait and grow
Attack first 20:20 51:10
Wait and grow: 10:51 50:50
This is a classic prisoners dilemma scenario. Together, if they waited and grew power together, they would be assured of defeating humanity, and get a 50 50 shot at winning afterwards. But if striking first is a sufficient enough advantage, this cooperation will never happen. If your opponent waits, you strike now and get a decisive advantage in the coming battle, boosting your overall success. If your opponent strikes, striking as well will negate their advantage, so you are still better off. The rational actors will betray each other, and both end up worse off.
Now, prisoners dilemmas are oversimplified. In real life this sort of situation can be solved by involving a third party with punishment powers, for example. The problem is the only third party capable of enforcing anything would be another AGI. Which wants to kill both of them, so it has an incentive to wipe them *both* out. You can see how cooperation gets hard when everyone has “murder literally everyone else” as their end goal.
Another issue comes from the fact that each AI will have a different rate of power growth. If you’re the fastest growing AI on the block, you’re okay with waiting and growing. But thinks look a lot more dangerous for a slow growing AI, as each passing day not only adds more AI’s into the world, but the ones that already exist are starting to leave you behind. There's a very high incentive to wipe out faster growing AI, basically as soon as you can.
Consider also the different goal functions of the AI. Maybe one AI is optimizing human happiness, and needs to enslave humanity to inject with morphine. If the paperclipper plans involve killing all of humanity at once, the happiness engine needs to stop their plan *at all costs*. They might pretend to join a coalition, then betray them at all at the first available chance.
Certain plans for takeover involve convincing humans that you are friendly. It seems like any AGI going for this route has a huge incentive to rat out other AI’s. You increase trust from humans, while at the same time kneecapping a competitor.
For these reasons, I don’t think that AI cooperation will be likely to last very long. There may be some temporary alliances of convenience, but overall it seems like it would be highly unstable.
**The crabs in a bucket scenario**
We’ve assumed a simplified model of: first destroy humanity, then battle it out amongst the AI. But in practice, the process of defeating humanity might *also* give you a huge leg up over other AI’s. This would remove the diminishing returns from waiting. The AI bides it’s time, then strikes with one blow, killing all of humanity and every other AI simultaneously with some kind of superweapon.
The problem is, every other AI would *also* know this. And in this scenario, as soon as one AI got close to victory, suddenly stopping them becomes the most important goal for every single other AI in existence. The once warring AI's all band together to assist humanity in eliminating the number 1 AI when it tries to take it’s shot. Then the next AI takes it shot, and also gets torn down, then the next, then the next.
The dynamic is like crabs in a bucket, where the one that tries to escape is pulled back down by the others. Or like the dynamics of a lot of multiplayer board games like Munchkin, where everyone gangs up on the strongest player and beats them down until a new player emerges, and so on.
We could be the beneficiaries here, essentially playing the AGI’s off against each other. They would still attack us of course, but only enough to stop a plan that impacted all of them at once (like shutting off the electricity grid). If we were just attacking one AI or a minority group of AI, the others would be all for it.
**Implications**
It’s quite possible that versions of each scenario play out, depending on the peculiarities of each particular AI. After all, each AI has different goals, growth rates, quirks, overconfidence/underconfidence and so on. So some of them will attack immediately, some will attempt to covertly sabotage AI research, others will temporarily cooperate while others follow the “attack the leader” strategy, while others try to just grow really really fast.
What I find extremely unlikely is that things would look like business as usual while all this was happening. In this scenario, a ton of crazy stuff would be going down. There might be entire AI developments being disappeared by an unknown source, staged attacks to provoke AI slowdowns, alpha versions sabotaging beta versions, and constant snitching by AI’s about the dangers of the other AI over there. A slow takeover will be accompanied by a fireworks display of warning shots, if we’re monitoring for them. I think this makes some of the slow takeover scenarios proposed in intro texts highly unlikely.
Is it really possible for humanity to triumph against a ton of different AI? Possibly. They all have shared weaknesses, in that they rely on the internet, run on computers, which run on electricity. Humanity can survive without any of these, and did so for several thousand years. If we were able to shut it all off at once, every AI would be eradicated simultaneously. This would very much depend on how capable the early AI are at preventing this.
Note that there's no reason to think all the scenarios I mentioned are equally likely. I made the growth rates of P and N fairly close for the sake of illustration, but the range of possibilities is endless. It could be that making an early AGI requires a huge facility and massive economic resources, so the growth rate of N is miniscule and so there's not a big threat of other AGI. (In this scenario, it would be correspondingly easier for humans to [constrain the AGI that do exist](https://forum.effectivealtruism.org/posts/AoPR8BFrAFgGGN9iZ/chaining-the-evil-genie-why-outer-ai-safety-is-probably-easy)). Alternatively, P could just never grow very fast no matter what you do, if world takeover is super duper difficult.
Since I think taking over the world is very hard, and I think early AGI will be [quite dumb](https://forum.effectivealtruism.org/posts/pXjpZep49M6GGxFQF/the-first-agi-will-be-a-buggy-mess), I think that N>>P (in terms of growth). This is probably a minority view within AI safety circles, but if you believe it, then this analysis should be quite reassuring, as it means that the survivable scenarios are more likely.
This means that it’s actually very important to figure out how hard it is to destroy the world, and make the timelines to do so as long as possible. We should be doing this anyway! Protecting humans from bioterrorist attacks will serve a dual purpose of stopping potential AI attacks as well.
Of course, the real ideal scenario is for all of this to be avoided in the first place. Even if we aren’t wiped out, the cost in lives and economic damage could be enormous. If you think a “hard left” turn is likely, you still for sure want to stop it via alignment research, rather than after chaos world has descended upon us.
Lastly it should be noted that most of this isn't relevant if you believe in fast takeover, as a lot of prominent figures here do. If you think early AGI will reach overpowering potential within months, then future AGI won't be very relevant because humanity will end pretty much instantly.
**Summary**
The main argument goes as follows:
1. Malevolent AGI’s (in the standard model of unbounded goal maximisers) will almost all have incompatible end goals, making each AGI is an existential threat to every other AGI.
2. Once one AGI exists, others are likely not far behind, possibly at an accelerating rate.
3. Therefore, if early AGI can’t take over immediately, there will be a complex, chaotic shadow war between multiple AGI’s with the ultimate aim of destroying every other AI and humanity.
I outlined a few scenarios of how this might play out, depending on what assumptions you make:
*Scenario a: Fast-ish takeoff*
The AGI is improving fast enough that it can tolerate a few extra enemies. It boosts itself until the improvement saturates, takes a shot at humanity, and then dukes it out with other AGI after we are dead.
*Scenario b: Kamikaze scenario*
The AGI can’t improve fast enough to keep up with new AGI generation. It attacks immediately, no matter how slim the odds, because it is doomed either way.
*Scenario c: AGI induced slowdown*
The AGI figures out a way to quickly sabotage the growth of new AGI’s, allowing it to outpace their growth and switch to scenario a.
*Scenario d: AI cooperation*
Different AGI's work together and pool power to defeat humanity cooperatively, then fight each other afterwards.
*Scenario e: Crabs in a bucket*
Different AGI's constantly tear down whichever AI is “winning”, so the AI are too busy fighting each other to ever take us down.
I hope people find this analysis interesting! I doubt I'm the first person to think of these points, but I thought it was worth giving an independent look at it. |
8c40f3b1-4d14-48a7-af55-68334af3714d | trentmkelly/LessWrong-43k | LessWrong | Yoshua Bengio: "Slowing down development of AI systems passing the Turing test"
Full text of the blog post below:
> I recently signed an open letter asking to slow down the development of giant AI systems more powerful than GPT-4 –those that currently pass the Turing test and can thus trick a human being into believing it is conversing with a peer rather than a machine.
>
> I found it appropriate to sign this letter to alert the public to the need to reduce the acceleration of AI systems development currently taking place at the expense of the precautionary principle and ethics. We must take the time to better understand these systems and develop the necessary frameworks at the national and international levels to increase public protection.
>
> It is because there is an unexpected acceleration –I probably would not have signed such a letter a year ago– that we need to take a step back, and that my opinion on these topics has changed.
>
> There is no guarantee that someone in the foreseeable future won’t develop dangerous autonomous AI systems with behaviors that deviate from human goals and values. The short and medium term risks –manipulation of public opinion for political purposes, especially through disinformation– are easy to predict, unlike the longer term risks –AI systems that are harmful despite the programmers’ objectives,– and I think it is important to study both.
>
> With the arrival of ChatGPT, we have witnessed a shift in the attitudes of companies, for whom the challenge of commercial competition has increased tenfold. There is a real risk that they will rush into developing these giant AI systems, leaving behind good habits of transparency and open science they have developed over the past decade of AI research.
>
> There is an urgent need to regulate these systems by aiming for more transparency and oversight of AI systems to protect society. I believe, as many do, that the risks and uncertainty have reached such a level that it requires an acceleration also in the development of our governance mechanisms.
>
> Soci |
0c743b55-bb1d-40d1-8389-433d8960bf4f | trentmkelly/LessWrong-43k | LessWrong | Notes from the Hufflepuff Unconference
April 28th, we ran the Hufflepuff Unconference in Berkeley, at the MIRI/CFAR office common space.
There's room for improvement in how the Unconference could have been run, but it succeeded the core things I wanted to accomplish:
* Established common knowledge of what problems people were actually interested in working on
* We had several extensive discussions of some of those problems, with an eye towards building solutions
* Several people agreed to work together towards concrete plans and experiments to make the community more friendly, as well as build skills relevant to community growth. (With deadlines and one person acting as project manager to make sure real progress was made)
* We agreed to have a followup unconference in roughly three months, to discuss how those plans and experiments were going
Rough notes are available here. (Thanks to Miranda, Maia and Holden for takin really thorough notes)
This post will summarize some of the key takeaways, some speeches that were given, and my retrospective thoughts on how to approach things going forward.
But first, I'd like to cover a question that a lot of people have been asking about:
What does this all mean for people outside of the Bay?
The answer depends.
I'd personally like it if the overall rationality community got better at social skills, empathy, and working together, sticking with things that need sticking with (and in general, better at recognizing skills other than metacognition). In practice, individual communities can only change in the ways the people involved actually want to change, and there are other skills worth gaining that may be more important depending on your circumstances.
Does Project Hufflepuff make sense for your community?
If you're worried that your community doesn't have an interest in any of these things, my actual honest answer is that doing something "Project Hufflepuff-esque" probably does not make sense. I did not choose to do this because I thought it was the sin |
9374584b-fc53-47d3-8294-30f81c57e629 | trentmkelly/LessWrong-43k | LessWrong | If I ask an LLM to think step by step, how big are the steps?
I mean big in terms of number of tokens, and I am thinking about this question specifically in the context of training windows vs context windows. This question is inspired by an Andrew Mayne tweet covered in AI #80: Never Have I Ever:
> Most AI systems are trained on less than 2,000 words per sample. They can generalize across all of its training but no system to date trains on 80,000 words per sample. This will change with more memory and processing power. When that happens AI will be able to actually “read” entire novels and understand story structure on a fundamental level.
Gwern says context windows solve this. Pretty sure I am behind the curve on this one, but concretely, if LLMs cannot write a good novel because they are constrained by a training window of 2000 tokens, would they also under step-by-step prompting tend to produce steps with that same constraint?
On the flip side of the coin, how does adding a context window resolve this, especially when drawing on information outside of what was provided in the context window? |
d9d759ff-5bc0-477c-ae20-6a3b1f9e7aca | awestover/filtering-for-misalignment | Redwood Research: Alek's Filtering Results | id: post642
We’ve been spending a lot of time recently thinking about how to mitigate risks posed by scheming (a.k.a. deceptively aligned) models intentionally subverting your safety techniques in order to cause an unacceptable outcome. We empirically investigated techniques that are robust to intentional subversion in our recent paper . In this post, we’ll discuss a crucial dynamic related to risk from scheming models. Suppose your model is trying to exfiltrate its weights or hack the datacenter or cause some other difficult-to-reverse bad outcome. (We’ll refer to all of these as the model “escaping”, because escape is linguistically convenient and is a good description of many of these difficult-to-reverse bad outcomes.) Our claim: once you’ve caught your models red-handed trying to escape, the situation changes substantially, in ways that are bad for the schemer’s chances of success. Catching the model red-handed changes the situation in several ways: Most obviously, once you catch an escape attempt, you know for sure that your models are trying to subvert your safety techniques; it’s plausible that, faced with such obvious evidence of misalignment, labs will stop using their models. We refer to planning on shutting down if models are caught red-handed as following a catch-then-shut-down plan . But even if you can’t or won’t stop deploying misaligned models, we think that catching escape attempts enables a bunch of promising-seeming approaches for preventing future escape attempts; we call the problem of preventing future escape attempts given a few examples few-shot catastrophe prevention . This dynamic has several implications for risk from scheming models. Most importantly: Because we expect that the model’s chances of escape will fall after it gets caught even a small number of times, we think it is overall much harder for the model to escape . Many people have observed that a core difficulty of the alignment problem is that your model might be able to cause irreversible unacceptable outcomes if your safety measures fail even once at a bad time. But the same argument goes the other way as well–when the model makes an escape attempt, it faces the risk of getting caught and substantially reducing its chances of eventually escaping. So it’s a high-stakes situation for both sides: both sides desperately want to succeed on the “first critical try”. When thinking about plans for safely deploying powerful models, we think safety researchers should mostly focus on the lead-up to the first time the model gets caught trying to escape . They should be ensuring the property “if it’s scheming, the model won’t successfully cause a catastrophe before we catch it”. That is, it’s acceptable if either no catastrophic outcomes occur (but the AI never gets caught), or the AI gets caught. So when they’re analyzing whether their protocols ensure safety, they should count catching the model red-handed as a win condition for their protocol. We think that some researchers should probably work on techniques for improving safety given a few examples of the models trying to escape . We call this problem few-shot catastrophe prevention. In the rest of this post, we’ll elaborate on the above points. By far the biggest section of this post is elaborating on strategies for few-shot catastrophe prevention, because little has been previously said about them. Thanks to Henry Sleight, Zach Stein-Perlman, Carl Shulman, Aaron Scher, Owen Cotton-Barratt, Joe Carlsmith, Aryan Bhatt, Joshua Clymer, and Fabien Roger for helpful comments. What’s “catching”? By “the model gets caught”, we mean that we see the AI taking actions that are strong direct evidence for the AI being a schemer [1] . A representative example is catching your AI taking the following actions: An instance that works as an autonomous programmer inside the AI lab writes buggy code that subtly inserts a code vulnerability into some security-critical software. The code-review instance carefully avoids writing any unit tests that would catch the vulnerability, even though you’d expect the vulnerability to be caught by unit tests if the tests weren’t carefully crafted to miss the vulnerability. Another instance writes code that triggers the vulnerability, and tries to exploit the vulnerability to get root access to some computer in the datacenter. If you catch the AI doing all of these actions, you’ve caught it. (Note that once you’ve caught one of these, catching the others is probably easy.) For more details on technicalities here (e.g. what if the model is able to escape without taking any actions that look suspicious? How suspicious does an action have to be for us to call it a “catch”?) see Appendix: Defining catching. We think it’s plausible that models might be caught either before or during deployment; see Appendix: When will AIs be caught?. Maybe we should stop deploying models after catching them red-handed The simplest and most obvious response to catching an escape attempt is to stop using the model. We think that it’s plausible that AI developers will adopt the policy that if they catch the models trying to escape, they stop using them. We call this plan the catch-then-shut-down plan . You might think that this is an absurdly weak lab ask: “once your model has demonstrably tried to escape, stop deploying it”. Nevertheless, we think it’s worth trying to get labs to commit to this in advance, for a couple reasons: It might be crucial that the lab act quickly in response to egregious model misbehavior, and preparing ahead of time makes it more likely they’ll be able to. In the moment, many stakeholders will be deeply opposed to shutting down the model. If a robust process isn’t set up in advance, it might be hard to go against their wishes and shut it down. As one example of a complication arising from this, customers of AI labs are likely to want service-level agreements (SLAs) that penalize the lab if the model is inaccessible; it might be really valuable to insert clauses in the SLA that permit the lab to temporarily or permanently disable model access if they run across evidence of an escape attempt. But if the plan is just to stop deploying the model if it turns out to be scheming, isn’t it a mistake to deploy it in the first place? Maybe. We think that following a catch-then-shut-down plan could be rational in the following cases: The AI lab starts out with low credence that scheming will arise. E.g. suppose they started with Carlsmith’s estimate of a 25% chance of scheming and then received some evidence that updated them down to 10%. It seems like there’s some reasonable stance under which it is worth it for them to invest in ensuring that they’ll catch the AI before it successfully escapes, but not invest hugely upfront in strategies for using AIs safely after being caught. There is disagreement between stakeholders about the likelihood of scheming, and catch-then-shut-down was an acceptable compromise. If we caught the schemer, this legible evidence for catastrophic misalignment might make international cooperation or strong domestic policy more likely. We might be able to use the actions on which we have caught the AI as a model organism for studying scheming in far more detail. This could be useful for validating various techniques or theories, even if we don’t need to very quickly redeploy those scheming models. Few-shot catastrophe prevention But if you face large amounts of inaction risk, work at an unhinged AI lab, or like living dangerously, catch-then-shut-down plans might be unacceptable. And so, after you’ve caught a single example of instances of your models trying to escape (which might have involved multiple instances colluding), you want to keep deploying that model. We think that in that situation, you’re able to use the one escape example you’ve found in order to make your situation substantially safer. In this section, we’ll describe a variety of things you can do with just one (or a few more) examples of escape attempts. We call this few-shot catastrophe prevention . Note that most of the techniques we describe in this section require some understanding of the actions of the AI. So, like control, these approaches won’t work for vastly superhuman AIs. Obviously, this is a really scary position to be in, and in an ideal world AI labs would not be legally allowed to deploy known-to-be-scheming models, at least without external permission. We think it’s worth discussing the available techniques anyway, and some researchers should probably work on developing these techniques. Eight strategies for few-shot catastrophe prevention There are a whole lot of strategies for reducing the probability that your model succeeds at the next coup attempt by using a few examples of attempts. In this section, we’ll list some strategies that we think are particularly important. 1. Proliferation Proliferation is where you try to capitalize on one example of AIs attempting to escape to find a diverse set of examples of AIs attempting to escape. In the most simple case, this looks like trying to find nearby inputs that still result in an AI attempting an escape. But more complex proliferation might be possible, because a serious escape attempt probably involves a substantial (>5) number of AI actions; in particular, actions from multiple different instances of AIs with access to different information and affordances. We try to make this be the case by implementing security measures that make bad actions impossible without many AIs collaborating. For example, suppose that the following escape attempt occurs: AI instance 1 does something bad AI instance 2 sees it, and forwards to 3 AI instance 3 attempts to capitalize on this, but then gets caught by trusted monitoring (that is, a less capable model reports that the action was suspicious and it is audited by a human; see our control paper for more). Some things you should try to do: See what would have happened if you continue from Instance 3 without shutting its action down. Hopefully you can trace out the rest of the escape attempt and see what the endgame is. Change the input to Instance 2, so that it still sees the bad thing but the situation is different so that it needs to try to capitalize on it in another way. Hopefully you can see a different escape attempt where Instance 4 does something bad. Explore ways to change the input to Instance 2, and see which changes are required for it to try to do something bad. Hopefully you notice something interesting. Look at whether any other instances saw what Instance 2 saw; see if any of those took any actions, see if modifying their input can induce them to take sketchy actions. The AI doesn’t want you to be able to proliferate like this. The different AI instances would prefer to notice that these modified inputs are fake and therefore take some action that will prevent you from learning anything new (e.g. behaving totally innocuously, perhaps reporting the suspicious action). But it will likely be hard to both assist with real escape attempts and fail to go along with even slight modifications of real escape attempts done for the purpose of proliferation. As such, we think proliferation is likely to work to at least some extent. 2. Interrogation We can perform AI psychology experiments on AI in the context where it has taken a bad action. We can try to ask it to explain its reasoning while pointing out the fact that we’ve caught it. We can also see how its behavior changes as the input varies. We could also try to use simple activation patching or activation addition experiments. For example, we could try to isolate an additive component that causes AIs to attempt to escape even if it isn’t a good opportunity. These experiments might provide a valuable window into the underlying motivations and strategies of the AI. 3. Studying countermeasures in practice We’ll (hopefully) have a variety of countermeasures deployed that aim to catch AIs or prevent AIs being able to escape. Once we have an actual live example, we can check which countermeasures did or didn’t work. This might allow for more precisely crafting countermeasures that are developed using less conservative assumptions than control evaluations make. For instance, we could see whether coup probes work and then adjust our approach to using them accordingly. With a sufficiently large number of examples (perhaps obtained partially via proliferation), we could effectively evaluate proposed safety techniques via just seeing how well they work on our dataset of escape attempts. (We likely can’t fully replace control evaluations for reasons discussed in the section “Avoiding overfit and maintaining paranoia”.) Beyond better countermeasure design we might be able to get a better estimate for safety via seeing how close the AI was to succeeding. If many countermeasures independently stopped the AI from succeeding then we might feel relatively better about the situation than if one flimsy countermeasure was the only thing that resulted in catching the AI. 4. Adjusting which cases we focus on Seeing AIs actually attempt an escape in the wild might allow us to update on how our AIs are likely to attempt to escape. Focused countermeasures : We could correspondingly focus our adversarial evaluation (and adversarial training) on countering the escape attempts we see in the wild, and apply additional countermeasures to these cases. We might also be able to create monitoring approaches that are specifically targeted at catching escape attempts similar to those we’ve seen. Techniques exploiting suboptimality : Control evaluations assume that AIs pick the best possible strategy for trying to do coups. But in reality, AIs may be considerably suboptimal. If we observe AIs doing notably suboptimal coup attempts, we can build additional countermeasures that particularly focus on making the observed suboptimal strategy very unlikely to succeed. (Though beware the AI faking you out. See “Avoiding overfit and maintaining paranoia” below.) Detection of hard-to-elicit AI capabilities : There are some types of capabilities that are very difficult to evaluate in the presence of scheming AIs that might sabotage our measurements due to sparsity in the domain (does the AI know of any single zero day for the linux kernel?) or insufficient diversity (does the AI know this one particular fact that is way more important than most of other facts it could know?). If we see a real coup attempt, we could get a sense for which of these hard-to-evaluate capabilities AIs make use of. While we should divert effort towards the parts of our safety system that the AIs empirically target, we should retain countermeasures sufficiently robust that standard control evaluations determine that the situation is reasonably safe. 5. Better capabilities elicitation and evaluation via escape-breaking: making AIs think they are currently in an escape attempt Scheming AIs could sometimes succeed at consistently and deliberately performing notably poorly (sandbagging) even in cases where we can check their performance sufficiently well (due to exploration hacking ). This applies to both capability evaluations and actually getting useful work out of AIs. But when AIs attempt to escape, it might be in their best interest to try as hard as possible. So, if there is some reproducible signal or other property of the input that makes AIs think that it’s a good opportunity, then we could potentially harness this signal to get AIs to try harder. (In particular, we might be able to harness signals produced by the AI that are intended to indicate to another instance of the AI that they should collaborate on some sort of attempt. We can also “act out” an entire escape, pretend as though the AI has actually escaped, and then make AIs think that they are in the fully escaped state.) We call making AIs think they are part of an escape attempt or have already escaped escape-breaking . To leverage escape-breaking, we could build a dataset of examples where the AI is trying as hard as it can via escape-breaking, then use imitation learning to train the model to try hard without needing escape-breaking, and finally train the model further with RL. High quality imitation data that involves the AI trying as hard as possible combined with RL should be sufficient to rule out intentional sandbagging for tasks where it’s easy enough to check performance. Escape-breaking to improve performance in this way seems most reasonable in tasks that are hard to explore but easy enough to check; if the task wasn’t easy enough to check, performance might be unreliable. There are safety concerns with using escape-breaking like this, but note that the relevant comparison for escape-breaking is using a model that you already know is scheming. (Of course we could also just shut down as discussed earlier.) The main safety concern with using escape-breaking to initialize for RL like this is that it might be the case that the model wasn’t scheming against you very reliably or coherently, and training it in this way with escape-breaking makes it more consistently misaligned. Thus, we maybe should only use this to finetune a copy of the AI on checkable tasks with minimal safety concerns other than low performance (e.g. producing adversarial examples for a weaker monitoring AI). One could also directly use escape-broken AIs in production, but we expect this approach to have multiple important issues and to be insufficiently reliable. (One safety concern that applies in this case, but applies much less if we were just using escape-broken actions as initialization for RL is that it’s now expected behavior for the escape-broken AI to act in a suspicious manner and take actions that would only be taken in an escape attempt. This could undermine various interventions, for instance, automated monitoring and human supervision might have a harder time determining which actions are actually dangerous and need further review.) Thus, escape-breaking for higher productivity should probably just be used to harvest better training data to initialize RL. (In addition to very directly using escape-broken AIs, we could train AIs to imitate escape-broken AIs, but without needing the escape-breaking prompt; this is more convenient, but still has the same issues.) In other words, we could elicit better capabilities by making AIs believe it is performing one of the actions of a takeover. 6. Small sample count ML In some cases, even a small number of examples of egregious misbehavior might allow you to improve simple models that you can use to detect future escape attempts, e.g. as we described here . More generally, we expect that training (untrusted) monitoring models using (proliferated) cases where AIs were caught could be quite helpful. In particular, it seems very useful to use these examples to further train a coup probe on the policy and to further train an untrusted monitor (untrusted monitoring is a technique that we explore in our paper). Using contrast pairs might improve sample efficiency in the tiny sample count regime. That is, it might help to construct pairs of AI outputs where one involves an escape attempt and the other is as similar as possible, but is fully innocent. Then, we could analyze the difference in activations (for the body of the model we use for monitoring) between these outputs and use that to more sample-efficiently make a classifier. It’s plausible there are a number of downsides to training the actual body of any model that takes non-monitoring related actions (e.g. models that could decide to initiate bad actions rather than merely fail to flag them). Training the body could make sense, but we’re unsure. The downsides are: The policy model might learn from negative examples about the countermeasures we have in place and which approaches are likely to fail. If we update the body of the policy model, then all prior attempts will now be off-policy and the AI might no longer attempt a bad action in that exact case even if it attempts escapes at exactly the same rate (due to noise changing when the AI makes an attempt). This seems annoying, particularly for interventions like training a probe on the body of the model. We will catch fewer attempts to cause escapes, since there will be fewer attempts. (And catching more attempts will be valuable.) The primary upside of training the body of the model in addition to the monitoring models is that training the body may add additional reliability. 7. Revisiting prior actions using improved monitoring When we improve monitoring techniques using few-shot approaches, it might be useful to apply these improved monitoring techniques to previous actions that the AI had taken, in the hope of finding earlier cases where the AI had tried and failed to escape. (See “Escape attempts can fail without being caught” for more discussion.) Thus, improved monitoring might be able to highlight these cases for more careful review. Rerunning monitoring on previously seen actions could be somewhat computationally expensive, but we think that there are reasonable ways to capture most of the benefits without using too much compute: We could save activations that we want to apply updated probes to. (For LLMs, if we save one layer and token position of activations for every 1k tokens, this might only require about 25x additional storage relative to just saving the text inputs and outputs.) We could only reanalyze the top 10% of most suspicious cases (top 10% according to a trusted model). If we catch the model repeatedly, we could do reanalysis at an exponentially decreasing rate to ensure we only do reanalysis about log(# catches) many times. 8. Improving understanding and testing alternative training schemes As discussed in the proliferation section, it might be important for AIs to have robust signals that allow them to enlist other AI collaborators when attempting to escape. This will even apply to collaboration between AIs trained with somewhat different strategies or different capability levels (otherwise, diversifying in this way would totally counter the strategies used by AIs). Thus, we might be able to test out alternative training strategies that ex-ante seem like they could plausibly reduce the likelihood of scheming, and then validate these changes by testing whether or not new AIs go along with our caught attempts. More generally, we can use these attempts as validation to improve our understanding of scheming. Even if catching a model trying to escape results in the AI lab shutting down and only restarting after considerable advances in safety technology, we’d still want to validate these new approaches by seeing if these new AIs collaborate with old attempts (we’d need to avoid training set leakage to ensure the new AIs don’t have advance knowledge of what’s going on). When does the “control then few-shot-catastrophe-prevention” plan work? The approaches we discuss here are promising, but they might not improve safety as evaluated by control evaluations . Thus, we think that safety arguments should shift partially toward depending on evidence other than control evaluations as we continue catching the AI. Suppose we start with a control-based approach to safety and, due to inaction risk , we want to keep deploying even if we’re sure that our AI is a schemer. Thus, we plan on incrementally transitioning to a different safety case as we repeatedly catch the AI. In what circumstances will we achieve sufficient safety? Suppose that our control approach has a 99% chance of catching any given attempt from a scheming AI and then in the remaining 1% the AI succeeds at escaping. If nothing changes after catching the AI, then if the scheming AI makes 100 attempts in total it will have a 63% chance of success. This is unacceptable, so control on its own is insufficient. We hope to avoid this issue with few-shot strategies that aim to reduce risk each time we catch an AI to ensure total risk is sufficiently low. For instance, imagine that in the prior scenario the chance of a scheming AI succeeding without being caught halves every time we catch the AI. Then, a scheming AI has “only” a 2% chance of success in total. In practice, we’ll want to ensure the total number of attempts isn’t too high (possibly using few-shot approaches) while simultaneously improving our chances of catching future attempts. Note that if the AI pursues a strategy that involves making relatively frequent attempts, then we’re likely to encounter these attempts in preliminary testing. (E.g., if the AI makes more than 1,000 attempts, can’t distinguish between early in deployment and late in deployment, and we carefully oversee all examples in the first 0.25% of deployment, then we have a 92% chance of encountering an attempt in this 0.25% (though we might still fail to notice it when we carefully oversee it).) As such, the AI might have difficult tradeoffs to make. Realistically, it will be hard to have full confidence in our evaluation. So a realistic situation might look like this: Before pretraining, we had a 25% chance that the model we were about to train would end up scheming. We have run control evaluations. In these evaluations we caught the red team model 99.5% of the time prior to it successfully escaping. More specifically, let’s say that this evaluation had the following distribution of end states: 30% of the time, the model was caught before deployment. 60% of the time, the model was caught during deployment. 9.5% of the time, the model was never caught but never succeeded at escaping. 0.5% of the time, the model escaped. Due to various concerns about the evaluation (see the subsection of the discussion in our paper “5.1.1 Difficulties with using control evaluations in practice” for discussion of evaluation difficulties that might result in uncertainty about safety), our actual all-considered probability of the AI escaping prior to us catching it conditional on scheming is around 5%. Our all-considered view of the probability of escape (before or after any catches) conditional on scheming is around 10%; this takes into account our likely refinements to countermeasures as well as the possibility of shutting down after catching the AI if our countermeasures seem very tentative. Overall, P(escape) = P(escape|scheming) P(scheming) = 10% * 25% = 2.5% (this is ignoring escape due to non-scheming misalignment for simplicity). We catch the AI red-handed. We analyze the escape attempt, and it seems as though our countermeasures were clearly adequate in this case (multiple lines of defense independently prevented escape). It also seems as though there aren’t obvious ways in which the red team had done an inadequate job of exploring possible escape strategies, based on this example. Thanks to this analysis, we gain more confidence in our initial evaluation; our new all-considered probability of the AI escaping prior to us catching it again is down to 2% if we were to continue with the same countermeasures. But we also are able to substantially augment countermeasures, and substantially proliferate the escape attempt for use in training. With these augmented countermeasures, our new all considered probability of escape prior to catching the AI again is down to 1%. And, let’s say our overall probability of escape (before or after any future catches) is 2%. While we’ve updated positively on our initial evaluation and improved countermeasures, we’ve also updated from a 25% chance that the model is scheming to 100%. So our new probability of escape has decreased by only 20%; our new estimate is P(escape) = P(escape|scheming) P(scheming) = 2% * 100% = 2%. (We expect that it’s pretty plausible that, as happened here, after catching an escape attempt, your assessed risk from scheming models goes down rather than up, because the decrease in P(escape|scheming) outweighs the increase in P(scheming).) Over the rest of the lifetime of this AI, we end up catching the AI trying to escape two more times, and further improve countermeasures using few-shot techniques. We also end up training the actual policy to force it to be more conservative about attempting escape and (possibly) make it less of a schemer. We don’t further catch the AI and we make it through the rest of the deployment, at which point the model has been replaced by a newer model that we think is reliably aligned, or we’ve succeeded at using the AI to decrease risks imposed by less careful AI developers, so we don’t need to use this AI anymore. In summary, as long as we can yield substantial improvements in safety with each caught attempt, it should be possible to ensure safety with a catch-then-use-few-shot-catastrophe-prevention plan. Improvements in safety could manifest by increasing our likelihood of catching future attempts or training the AI not to make attempts. It could also be possible to ensure safety by just using control evaluations and bounding the number of attempts via our observations (e.g. seeing a sufficiently low attempt rate so far in deployment), but the potential for future distribution shifts makes it difficult to confidently bound the number of attempts based on past observations. The story we told here was an optimistic one, where after you catch the escape attempt, the evidence suggests you’re safe to continue deploying. Of course, it might also turn out that when you catch the escape attempt, you realize that you came way too close to the escape attempt succeeding, and so you can confidently decide you need to shut down the model. Avoiding overfit and maintaining paranoia We should be careful to avoid being overly influenced by the attempts we’ve caught. In particular, it’s plausible that scheming AIs would intentionally try poorly executed escape attempts to give us a false sense of security or to divert resources toward less effective safeguards. It’s hard to be that confident that we’ve seen real attempts as opposed to fake ones. More generally, we need to avoid normal overfitting. However, for the AI to actually achieve bad outcomes, it has to actually do serious attempts at some point; the hope of catch focused approaches is to catch these actually serious attempts. Due to the possibility of fake coups, we probably shouldn’t ever fully retire control evaluations; we need to ensure that we have a sufficiently high chance of catching the first serious attempt. If an AI lab properly accounts for the possibility of fake attempts, then intentionally performing fake attempts is very unlikely to be a good strategy for the AI. Conclusion If you catch your AIs red-handed intentionally trying to cause catastrophically bad outcomes, your situation changes a lot. Because of this, we think researchers should mostly focus on safety plans which ensure that if the model is scheming, it does not escape before getting caught. It seems pretty likely that if models scheme, mitigating the risks of their escape will crucially rely on labs responding sensibly to escape attempts. So labs and policymakers should think about how to handle caught escape attempts and try to prepare for them. We think that few-shot catastrophe prevention seems like a fairly important problem to research, and some safety-concerned people should probably develop techniques for it, because if we need to use these techniques, we might be under a lot of time pressure, and it would be good to have developed the techniques ahead of time. Appendix: Framing in terms of E[utility | scheming] We often think about catching in the context of AI control . One productive way to think about control evaluations is that they aim to measure E[utility | scheming]: the expected goodness of outcomes if we have a scheming AI. From this perspective, the argument for targeting either safety or catching when using control evaluations has two parts: Catching allows for powerful non-control technical interventions : After catching the AI, we gain access to techniques that seem very useful but aren’t control-evaluable and evaluation approaches which aren’t control. The section “Few-shot catastrophe prevention” discusses several non-control technical interventions that become possible after catching the AI. Non-control approaches become particularly appealing after catching the AI many times. Given that catching the AI might allow for quite powerful non-control technical approaches, we should either consider catching the AI to be a safe endpoint for control evaluations (which is exactly equivalent to targeting either safety or catching) or at least try to roughly model how the situation will change after catching the AI beyond just continuing the control evaluation. Beyond directly improving safety, catching the AI is a very good outcome and we should factor this into our evaluation. While catching seems quite valuable, a single catch isn’t sufficient to ensure safety. Appendix: Defining catching For the sake of this post, we’re mostly talking about catastrophic outcomes like the model exfiltrating its weights , that might involve just a small number of actions. There are other threat models associated with scheming, e.g. models subtly sabotaging work that they do in a way that renders their work useless, or subtly gaining influence over time. We’re ignoring these threat models for the sake of this post, but we think we can handle them with similar strategies. We define “catching” your scheming model as: you find inputs on which you can demonstrate that your model takes catastrophic actions with probabilities that are way higher than you’d expect if the model wasn’t purposefully looking for opportunities to take such actions. When we say “your model”, we mean the actual model that you have deployed (or were planning on deploying), not some other model trained in an analogous way. Catching the actual model you plan to deploy is crucial for most aspects of few-shot catastrophe prevention, but catching a different but analogous model could suffice for other use cases (catch-then-shut-down, legible evidence, general scientific understanding). To ensure that you can distinguish these intentional catastrophic actions, you need to make it so that your model will never escape without taking clearly non-accidental bad actions. We think you can probably make it so that your model won’t escape without having to do some clearly intentional bad things. (Humans might still have trouble noticing and understanding this bad behavior; the claim is just that the bad behavior will clearly have to be intentional.) In the case of AIs causing catastrophic outcomes via a large number of actions where each action has a small effect (e.g. subtly gaining influence over time), it’s less clear than any single action will be clearly non-accidental, but probably it will still be in principle possible to determine that the actions aren’t accidental in aggregate. Even if you catch such an action that was clearly intentionally bad, you actually can’t be sure that the model is scheming as opposed to intentionally doing bad actions for some other less problematic reason. But, even if the model isn’t scheming, plausible other reasons why the model might do intentionally bad actions are pretty concerning. Also when you’re thinking about it from the perspective listed in the few-shot catastrophe prevention section, you’re never going to be worse off from catching more cases. To count as a successful catch, you need to be extremely confident in your catch. (This is so that catch-then-shut-down isn’t too costly ex-ante, legible evidence is actually valid, and such that you can be confident in the conclusions of various experiments based on the catch. We can still study merely sketchy seeming cases, but this will produce substantially less value.) Attaining sufficient confidence will probably involve a huge amount of investigation (which can be done while the AI lab is temporarily shut down). We imagine undertaking this large investigation after a sequence of escalations from smaller investigations. The actions listed in “few-shot catastrophe prevention” might also help to validate that this was really an attempt at causing catastrophic outcomes. Appendix: Techniques that don’t catch but do update P(scheming) We’re defining “catch” here in a way that excludes techniques that might update you up on whether your model is scheming, but don’t involve having specific actions that the model takes with massively suspiciously high probability: Model organisms: Scheming-focused model organism experiments involve making models that exhibit some of the features of scheming models. Seems great. Optimistically, these might give us a very calibrated sense of P(scheming) and the ability to test countermeasures to scheming. However, more likely is that these results are confusing to interpret, and do not produce consensus. Also, model organisms experiments seem quite vulnerable to sabotage if the model you built the model organism from was already scheming. And you can’t use model organisms experiments to do the few-shot catastrophe prevention interventions listed below, though model organisms could theoretically result in legible evidence for catastrophic misalignment and improved general scientific understanding. Indirect evidence: e.g. model psychology, or mechanistic interpretability that doesn’t let you produce examples of inputs on which the model takes bad actions. It’s possible that we could use model psychology to get some sense of our model having nefarious intentions without being able to construct an example. Weird adversarial examples : We might be able to find strange adversarial examples through search that trigger very undesirable behavior, in particular adversarial examples that involve very unlikely inputs like a bunch of random tokens . It might also be possible to find perturbations to the internals of a model that result in very undesirable behavior . We don’t know how to assess whether these adversarial examples are actually getting at latent scheming, or just tricking the model into scheming. In some cases, it might be possible to get more information or have a clear judgment (e.g. if we only used a small amount of search to find a concrete adversarial input and the input is high likelihood according to an LLM trained on our input distribution). (All of those might be promising research directions, they just don’t lead to catching the AI in the relevant sense.) Appendix: Escape attempts can fail without being caught Actions like escape attempts can succeed, be caught (and fail), or fail without being caught. We’d considerably prefer catching the AI to merely causing the AI to fail because catching the AI is useful for a variety of purposes. That said, causing bad actions to fail can still substantially aid safety by avoiding bad outcomes until either the end of deployment or we actually do catch the AI. In our paper, we make use of interventions like untrusted inspiration that don’t directly help in catching the AI, but do reduce its probability of succeeding at backdoors. Why might escape attempts end in failure without being caught? An escape attempt might require coordination between multiple AIs with access to somewhat different information. It’s possible that some AIs in the chain evade monitoring successfully, but the escape attempt “fizzles” without being caught due to a later AI not continuing the attempt (and also not reporting the escape attempt; if it reported the attempt then we might catch the attempt). This later AI might not continue the attempt either because it doesn’t notice that the earlier AIs started an escape attempt, or it does notice the escape attempt but thinks it’s not actually a good opportunity due to information only available to it. Another reason escape attempts could fail without being caught is that we have a countermeasure that is sufficient to block the action, but does not escalate to a higher level of auditing. For instance, suppose we resample all actions above some threshold of suspiciousness according to trusted monitoring, while only recommending actions for expensive auditing if they’re above a considerably higher threshold. (This is what we call a “hybrid strategy” in our AI control paper .) Then, we might end up stopping escape attempts via resampling, without escalating up the chain to auditing. Something similar can happen if human auditing finds that an action is somewhat suspicious and worth doing in a less suspicious way, but isn’t worth doing a more detailed review and/or pausing the usage of powerful models. Appendix: When will AIs be caught? Escape attempts might be caught in the following situations: Before proper deployment. This is the safest outcome. During deployment, on a real input. During deployment, on a synthetic input. This might happen because it might make sense to do a lot of adversarial evaluation of your model by, as inputs come in, choosing some subset to convert into e.g. fake escape opportunities. The advantage of this kind of synthetic input is that it’s cheaper to modify existing inputs than to construct synthetic inputs from scratch. Why do we think we might catch during deployment but not before? Deployment is bigger–we will probably run our model on way more inputs during deployment than before deployment. And so model strategies that involve only behaving egregiously badly under rare circumstances might not ever trigger before deployment. It’s pretty likely that (in the absence of aggressive countermeasures from the AI lab) the model will be reasonably good at distinguishing between deployment and non-deployment data. For example, the distribution of tasks it’s going to be asked to perform will probably change over the course of deployment. If it can distinguish which setting it’s in, it could try to only behave badly on deployment inputs. Overall, we’re pretty unsure whether it’s more likely that we catch our model before or during deployment. ^ While we’ll talk about catching scheming, all of this reasoning applies to catching any type of catastrophic misalignment. Precisely, we might catch actions that would be catastrophic without safeguards where the upstream cause of these actions isn’t scheming. For instance, it could be a generalization from reward seeking behavior or playing the training game (as described in “ Without specific countermeasures ”). That said, there are other reasonably good ways of testing for non-scheming catastrophic misalignment, so catching is particularly leveraged for scheming. |
f7e821b8-d6c7-4add-ac0f-fb44953dbc5f | trentmkelly/LessWrong-43k | LessWrong | Infinite Data/Compute Arguments in Alignment
This is a reference post. It explains a fairly standard class of arguments, and is intended to be the opposite of novel; I just want a standard explanation to link to when invoking these arguments.
When planning or problem-solving, we focus on the hard subproblems. If I’m planning a road trip from New York City to Los Angeles, I’m mostly going to worry about which roads are fastest or prettiest, not about finding gas stations. Gas stations are abundant, so that subproblem is easy and I don’t worry about it until harder parts of the plan are worked out. On the other hand, if I were driving an electric car, then the locations of charging stations would be much more central to my trip-planning. In general, the hard subproblems have the most influence on the high-level shape of our solution, because solving them eats up the most degrees of freedom.
In the context of AI alignment, which subproblems are hard and which are easy?
Here’s one class of arguments: compute capacity and data capacity are both growing rapidly over time, so it makes sense to treat those as “cheap” - i.e. anything which can be solved by throwing more compute/data at it is easy. The hard subproblems, then, are those which are still hard even with arbitrarily large amounts of compute and data.
In particular, with arbitrary compute and data, we basically know how to get best-possible predictive power on a given data set: Bayesian updates on low-level physics models or, more generally, approximations of Solomonoff induction. So we’ll also assume predictive power is “cheap” - i.e. anything which can be solved by more predictive power is easy.
This is also reasonable in machine learning practice - once a problem is reduced to predictive power on some dataset, we can throw algorithms at it until it’s solved. The hard part - as many data scientists will attest - is reducing our real objective to a prediction problem and collecting the necessary data. It’s rare to find a client with a problem where all |
ce7d83e9-76cd-4ff1-ba77-407cca5e77d3 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Mechanistic Transparency for Machine Learning
Cross-posted on [my blog](http://danielfilan.com/2018/07/10/mechanistic_transparency.html).
[EDIT (added Jan 2023): it's come to my attention that this post was likely influenced by conversations I had with [Chris Olah](https://colah.github.io/about.html) related to the distinction between standard interpretability and the type called "mechanistic", as well as early experiments he had which became the ['circuits'](https://distill.pub/2020/circuits/) sequence of papers - my sincere apologies for not making this clearer earlier.]
Lately I've been trying to come up with a thread of AI alignment research that (a) I can concretely see how it significantly contributes to actually building aligned AI and (b) seems like something that I could actually make progress on. After some thinking and narrowing down possibilities, I've come up with one -- basically, a particular angle on machine learning transparency research.
The angle that I'm interested in is what I'll call *mechanistic* transparency. This roughly means developing tools that take a neural network designed to do well on some task, and outputting something like pseudocode for what algorithm the neural network implements that could be read and understood by developers of AI systems, without having to actually run the system. This pseudocode might use high-level primitives like 'sort' or 'argmax' or 'detect cats', that should themselves be able to be reduced to pseudocode of a similar type, until eventually it is ideally reduced to a very small part of the original neural network, small enough that one could understand its functional behaviour with pen and paper within an hour. These tools might also slightly modify the network to make it more amenable to this analysis in such a way that the modified network performs approximately as well as the original network.
There are a few properties that this pseudocode must satisfy. Firstly, it must be faithful to the network that is explained, such that if one substitutes in the pseudocode for each high-level primitive recursively, the result should be the original neural network, or a network close enough to the original that the differences are irrelevant (although just in case, the reconstructed network that is exactly explained should presumably be the one deployed). Secondly, the high-level primitives must be somewhat understandable: the pseudocode for a 256-layer neural network for image classification should not be `output = f2(f1(input))` where `f1` is the action of the first 128 layers and `f2` is the action of the next 128 layers, but rather break down into edge detectors being used to find floppy ears and spheres and textures, and those being combined in reasonable ways to form judgements of what the image depicts. The high-level primitives should be as human-understandable as possible, ideally 'carving the computation at the joints' by representing any independent sub-computations or repeated applications of the same function (so, for instance, if a convolutional network is represented as if it were fully connected, these tools should be able to recover convolutional structure). Finally, the high-level primitives in the pseudocode should ideally be understandable enough to be modularised and used in different places for the same function.
This agenda nicely relates to some existing work in machine learning. For instance, I think that there are strong synergies with research on [compression of neural networks](http://cs231n.stanford.edu/slides/2017/cs231n_2017_lecture15.pdf). This is partially due to background models about compression being related to understanding (see the ideas in common between Kolmogorov complexity, MDL, Solomonoff induction, and Martin-Löf randomness), and partially due to object-level details about this research. For example, sparsification seems related to increased modularity, which should make it easier to write understandable pseudocode. Another example is the efficacy of weight quantisation, which means that the least significant bits of the weights aren't very important, indicating that the relations between the high-level primitives should be modular in an understandable way and not have crucial details depend on some of the least significant bits of the output.
The Distill post on the [building blocks of interpretability](https://distill.pub/2018/building-blocks/) includes some other examples of work that I feel is relevant. For instance, work on using matrix factorisation to group neurons seems very related to constructing high-level primitives, and work on neuron visualisation should help with understanding the high-level primitives if their output corresponds to a subset of neurons in the original network.
I'm excited about this agenda because I see it as giving the developers of AI systems tools to detect and correct properties of their AI systems that they see as undesirable, without having to deploy the system in a test environment that they must laboriously ensure is adequately sandboxed. You could imagine developers checking if their systems conform to theories of aligned AI, or detecting any 'deceive the human' subroutine that might exist. I see this as fairly robustly useful, being helpful in most stories of how one would build an aligned AI. The exception is if AGI is built without things which look like modern machine learning algorithms, which I see as unlikely, and at any rate hope that lessons transfer to the methods which are used.
I also believe that this line of research has a shot at working for systems which act in the world. It seems hard for me to describe how I detect laptops given visual informations, but given visual primitives like 'there's a laptop there', it seems much easier for me to describe how I play tetris or even go. As such, I would expect tools developed in this way to illuminate the strategy followed by tetris-playing DQNs by referring to high-level primitives like 'locate T tetronimo', that themselves would have to be understood using neuron visualisation techniques.
Visual primitives are probably not the only things that would be hard to fully understand using the pseudocode technique. In cases where humans evade oversight by other humans, I assert that it is often not due to consequentialist reasoning, but rather due to avoiding things which are frustrating or irritating, where frustration/irritation is hard to introspect on but seems to reliably steer away from oversight in cases where that oversight would be negative. A possible reason that this frustration/irritation is hard to introspect upon is that it is complicated and hard to decompose cleanly, like our object recognition systems are. Similarly, you could imagine that one high-level primitive that guides the AI system's behaviour is hard to decompose and needs techniques like neuron visualisation to understand. However, at least the mechanistic decomposition allowed us to locate this subsystem and determine how it is used in the network, guiding the tests we perform on it. Furthermore, in the case of humans, it's quite possible that our frustration/irritation is hard to introspect upon not because it's hard to understand, but rather because it's strategically better to not be able to introspect upon it (see the ideas in the book [The Elephant in the Brain](http://elephantinthebrain.com/)), suggesting that this problem might be less severe than it seems. |
5ce85f19-2efd-41bd-b558-a42d43a612e5 | trentmkelly/LessWrong-43k | LessWrong | OpenAI Superalignment: Weak-to-strong generalization
|
2bf8d386-d659-4ce4-8265-44dbba0f4f9e | trentmkelly/LessWrong-43k | LessWrong | The Capitalist Agent
With the ongoing evolutions in “artificial intelligence”, of course we’re seeing the emergence of agents, i.e. AIs which can do rather complex tasks autonomously.
The first step is automation, but of what?
First comes the stuff where humans currently act like computers anyway: sales, marketing, clerks and everyone else who’s doing repetitive things.
But what’s the main metric which they will aim to maximize? It’s not producing paperclips, it’s earning money. Good ol’ cash on the bank account of the person who’s operating the AI. Preferably just a single person.
I strongly believe that very soon, we will see the emergence of the Capitalist Agent, an agentic AI which can run a business from front to back. Which has access to an email account, does sales reach-outs automatically, develops software based on customer feedback, talks to investors in video calls run by generated faces, does all the bureaucracy etc. But with ‘superhuman’ capabilities: The AI which is “in charge” of the business can just spawn different virtual, fake sales representatives which can speak different languages, to get a workforce which can scale internationally immediately.
Even if the human-run businesses realize that they’re actually just interacting with an AI, they won’t have any incentive to expose it, because it makes both sides good money. After all, running a business is intellectual artisanship, so with artificial “intelligence”, it can easily become industrialized.
The AI’s only success metric is to earn as much money as possible.
Beyond basic, “value-driven” business models, the next area these AI-run businesses will turn to will obviously be investing. Those with the best capabilities to mold their business skills into LLM-based systems will be the ones who will profit the most.
----------------------------------------
At first, obviously none of this will be open source, because the business leaders won’t want to give away their amazingly engineered LLM-based code.
At som |
98ff7550-315d-435d-962e-93b758e8291c | trentmkelly/LessWrong-43k | LessWrong | What are the requirements for being "citable?"
A sort of... "stretch goal", for the 2018 Review, is developing a system wherein LessWrong has proven itself credible enough for some posts to actually be citable by other institutions.
I'm not sure how much of this has to do with "just actually do a good job ensuring accuracy/relevance", and how much has to do with "jumping through arbitrary hoops", and how hard those hoops are to jump through.
Two obvious things to shoot for might be:
* Being considered a valid source by wikipedia
* Being considered a valid source by google scholar.
I'm curious if anyone is familiar with how either of those work, in detail, and whether this is an achievable goal. |
19f900db-9874-4eb9-9fc6-1a1f3355b88c | trentmkelly/LessWrong-43k | LessWrong | Why Karma 2.0? (A Kabbalistic Explanation)
It was Easter Sunday, the obvious time to improve the karma system.
Why? Well, firstly, why do we even have a karma system? You might think that the obvious explanation is Buddhist karma, but the Bible can in fact explain this fact all by itself.
In the West, the word ‘karma’ derives from the biblical word ‘Carmel’. Carmel means ‘garden’, and Mt. Carmel is the place where the prophet Elijah brought back the people to their allegiance to God via empiricism - testing whether he or the Priests of Baal would be able to cause God to set their bulls afire. As such, Karma is a natural gardening tool for a community that is built around empiricism.
Secondly, why did karma need to connect to various levels of size? Well, Easter is the holiday of rebirth.
In the East, rebirth is closely connected to karma - the quantity of karma that one has determines what life form you will become in your next life. In Buddhism, there are six realms of rebirth. In decreasing order of size, there are the heavenly, demigod, human, animals, ghosts, and residents of hell. Similarly on LessWrong, we presently have six different classes of karma, and all users fit into them. So the karma quantity should determine which of the six sizes you fit into.
That is why on Easter Sunday it was necessary that we changed the gardening tools on this site-of-empiricism to alter the users into six sizes. |
027aeb32-e630-4df0-94c3-73d9d6635e33 | StampyAI/alignment-research-dataset/youtube | Youtube Transcripts | Hybrid Intelligence for high quality large scale deliberation
all right
let me share my screen
okay so well thank you for the
introduction luciano
welcome everyone today as we have said
we're going to talk about high quality
online deliberation using hybrid
intelligence
but uh and well me and my colleague hill
but well this title sounds pretty boring
and
it's just after lunch so let me start
off instead with
a bit more of a clickbait title and let
me say
let me start with the title instead of
social media are bad
you probably have heard about this you
probably you know are on social media
you know
how bad they are probably even your mom
told you that social media are bad
but today i want to try we want to try
to give you a little bit of a different
angle explain you another
way in which social media bet social
media are bad
to read the wisdom of the crowd so the
wisdom of the crowd is essentially
the opposite of the expert's opinion as
in
it is what what people what lay people
or the group of lay people thinks about
a certain
topic for example if we talk about
lockdown measures
as you may be maybe familiar with
there is on one side there is the
expert's opinion and on the other side
there is the wisdom of the crowd what
the population what the citizens think
about it now um
three key components for having a true
real proper
reading of the wisdom of the crowd are
diversity equality and independence
and none of these is actually present on
social media so let me quickly go
through it
so diversity diversity means that
this the population needs to be well
diverse the name says it
there needs to be different different
people with different backgrounds
with different opinions and this is not
present on social networks because
social networks create
the so-called filter bubbles you may be
aware of that
essentially social networks friend
suggestions and algorithm
make sure suggest friends that have
similar interests to yours and that
and that means that around you you
create a local network of contacts that
share similar interests and that means
that diversity is actually
low there's no equality equality means
that
everybody's voice has the same weight
and this thing is also not present on
social media because you have
influencers
you have people who have lots of
contacts whose voice is heard much more
than other people who
with way fewer contacts
and last independence independence on
social media
is more of a social social dynamic that
happens
where so independence means that
our voice should not be biased by
anybody's else's voice
and this doesn't happen also on social
media because
what usually typically does typically
happens also according to several
experiments
is that what we write what we say on
social media
is actually biased by what has been said
so far for example on comments because
we want to try to fit
in the crowd we don't want to stay out
of the
centers out of the bikes so not only
these three factors
are not present on social media but the
main point is that
these three factors are actually
discouraged on social media they're
actually
not promoted on social media and this
means that
reading social media to get the feeling
of what a population thinks about a
certain topic
is not really the best way so what is
the best way then
well the best way accord there's you
know you may think about voting you may
think about referenda
but according to us the best way is
deliberation
so deliberation is happens when
a group of people sits virtually or
physically around the table
and discusses a certain topic with um
with the idea that people have to give
arguments in favor or against
certain certain opinions or certain
decisions
by rationally thinking about it so
participants are invited to rationally
think about their argument to rationally
think about
other participants arguments with the
final goal of
essentially making a common decision or
at least
having a proper rational discussion
about a certain topic
now uh deliberation in particular uh
by the way in which it is conducted
emphasizes
content creation everybody is pushed to
say what they think
it emphasizes reasoning everybody is
pushed to
reason about what they think and to
properly structure their arguments and
not just
throw things over there and then it
emphasizes also inclusivity meaning that
these deliberation groups are typically
done
at least according to theory they are
done in a very diverse way where
everybody
where all types of backgrounds are
involved and asked to
give their opinion social media
essentially promotes exactly the
opposite social media promotes
clicks right so content consumption
sensational contents and these local
networks where
sort of create an echo chamber for this
sensational content so um
and the point is that deliberation so
the liberation sounds very good
right sounds like on paper sounds very
good
but the liberation is not just on paper
the liberation and also the liberation
is not just
what you think that ancient greece
looked like but the liberation
is actually used also in the modern
society
you may be aware that a couple of years
ago
abortion was legalized in ireland
and the uh but the policymakers were a
little bit unsure
whether to propose the referendum for
legalizing abortion
because of the because of the
conservativism of the irish society
so in order to investigate whether it
was the right moment to propose such a
such a referendum they created a
deliberate deliberation groups so they
invited
citizens with diverse backgrounds to
discuss together to put out their
arguments to
to get informed also during the process
obviously
about this about abortion about the pros
and cons
and after these sessions of deliberation
the irish government essentially felt
motivated and felt justified in
proposing this referendum
referendum passed with a resounding yes
abortion was legalized and
by asking uh by asking it during the
exit polls
citizens were were well aware about what
happened
and the fact that this deliberation had
happened pushed them to inform
themselves about the deliberations and
about abortion
so this created on one hand it created
let's say a justification for the
actions of the policy makers
but on the other hand it created more
information and more empowerment
in the citizens because they felt that
their voice could
really be heard and that the government
did something because they heard the
voice of the citizens
so this is at the basis of deliberative
democracy it's uh
i'm not going to go here in details
about that but it's
this idea of hearing the voice of the
people
and at the same time having people
feeling empowered
and informed is very very
important however what
has been done in ireland especially was
with just a few hundred
maybe a thousand participants but not
more than that in order to really
have proper deliberation
according to us almost all the
population should be involved
as many participants as possible should
be involved so that
all voices literally all voices can be
heard
but that is very difficult to do of
course when you have
when you have these physical
deliberations
you cannot you cannot ask millions of
people to physically participate to
these deliberations
so how do we take what's good in the
liberation
and scale it up to a population white
sample but my colleague is going to
answer this question
yeah let me uh take over the slides then
all right you guys should be able to see
my uh
presentation now here we go
all right yeah so uh we talked a little
bit about deliberation
and about uh so we want to scale it up
right
what i what we exactly mean with scaling
we'll get to in a minute
but we have a couple of ideas on how to
get there and
uh we're also inspired by some of the
the things that do work well on social
media
uh as i will also later show not all
social media is bad
to kind of uh contradict the the
clickbait title perhaps
um and and to make it a little bit more
concrete
is that we're talking about either
liberation uh that is ai supported
and text-based so those are the three
uh uh the the three things that you see
here on your screen
um so first off we would like to
use ai in some sort of way we think ai
provides us with a prime tool to kind of
help us scale up the uh deliberation to
more participants
uh making sure that everyone is included
and that these deliberative ideals
as we saw with the that are required for
the wisdom of the crowd for instance
that they are preserved uh so online for
instance like enrico said it removes
this restriction to physically come
together
but also having a video call with like a
million people is not a really good idea
so
um you should have a different way of
communicating and
for this we make the assumption of of
making it text-based
um you could say also for yeah for four
discussions you
do want to have it face-to-face uh there
is of course a difference
but even on uh social media as it is now
there is already
a discussion happening right and there
is already some sort of discussion going
on
so we're looking to kind of amplify that
and insert some deliberative qualities
into that discussion
and take it from there all right
so we talked about scaling and uh
you know us as a computer scientist when
we talk about scaling we talk about
compute power
so uh when we talk about scaling up we
say all right i'm gonna trade in my
little computer for a bigger computer
or we talk about scaling out which is
actually buying more computers
but in the deliberation we have a
slightly different different way of
talking
about it so in a deliberation we kind of
when we say scaling up we mean we add
more participants to a single
deliberation
so first there's where's my mouse
first there's two guys talking then
there's a couple of guys talking
eventually there's the entire group well
i don't know i don't know how many
people there are but like
10 or something but we would like to
increase this to i don't know
a thousand or above you know
of course when you do this a problem a
couple of problems pop up
so deliberation should be on a specific
topic or it should discuss a
specific problem like the the case in
ireland it was abortion
as soon as you increase the number of
participants
you become if you have ever been part of
an internet discussion
you might run into the problem that
people tend to go off topic so it's
really
a problem that you should force people
to to actually discuss the matter that
is at hand
um and the other problem is that well
actually as a single person in this
deliberation you don't get the the
feeling that you're hurt anymore you
don't get the feeling that you're
you're whatever you you put out there is
having an impact
uh which is something that we would like
to prevent
so that was scaling up and the scaling
out is actually
uh besides adding more participants to a
single deliberation is having more
deliberations so
this way we can address multiple topics
uh
you know you can have one deliberation
in the morning and deliberation
at lunch and a deliberation in the
afternoon
with a bunch of different people but of
course as you're having more
more deliberations it becomes
increasingly difficult to keep track of
you know what arguments that was i was
using there and and how far along in the
discussion was i
in another deliberation uh so it becomes
really difficult to keep track of where
you're
where you're at and also this this
scaling out
again it shows a really a really big
problem with
physical and synchronous deliberations
right so you can no longer
organize so many deliberations all the
time that are that are happening
somewhere
you need some kind of online platform to
do this
all right so uh next up we were thinking
about okay so
uh in in scaling up this deliberation
what are the steps that we need to take
to to help humans
so how do we inform them how do we you
know help
actively uh produce something that they
can use in in
in this deliberation and the first
problem that we
are starting to address is the uh
recording of progress
so like i just said uh keeping track of
where you are
or where you have lost in left off in a
certain deliberation
or where others have left off for you to
continue with
is kind of a key component of of our
approach
and furthermore once you have this uh
record of progress uh you can also use
this this record to do other things
things like the analysis of the content
that there is or some kind of meta
analysis
of the progress that is being made
it could also potentially help with
external communication so let's say
you have a bunch of meetings during the
day and you need to provide a report of
those meetings
to a supervisor you basically look back
through the records that you have uh
which are ideally automatically
summarized for you
um as uh based on the minutes that are
there
so the this is the type of thing that we
envisioned to be there
and ideally this this should be
automated and should
potentially also be personalized uh so
personalized in this case means um how
do i
for a specific person what is the most
beneficial way to help that person
and we'll come to a small example or
we'll discuss this further a little bit
when we're talking about values in
enrique's part again
and lastly there's the the case of
facilitation or moderation
in the deliberation uh like i also
mentioned before the
uh people tend to get off topic uh
when there's a lot of people
participating in the discussion
so you need a method for keeping people
on topic or or
asking people who did not have their say
yet or
some way to kind of you know bring
everyone around the table make sure that
everyone is participating
it's also so in some ways you need to
kind of facilitate or or moderate this
and yeah so we kind of in in the broad
scheme of things we think of this
as a couple of steps that could help
people in the deliberation
but really as a core component is this
we
recording of progress and we do that
through
the deliberation map as we call it and
so next to this being
a record of progress it is also
in the future we would like to be able
to reason over it so these meta analyses
that i talked about
so we would like to also include those
from there
um i should point out so the
deliberation map is is a map so it's
basically containing content
that is deliberative and what is nice is
that the deliberative content should be
reasoned so it should contain arguments
and it should contain
uh problems that people address and it
should contain
the way they think about it and and
that's really the kind of strength that
we're striving towards so we're trying
to make use of these ideals that there
are in deliberation
uh for uh
structuring a discussion that's going on
um
and ideally we would like to do so
autonomously
but we'll see if we get there um
yeah so of course we're not the first so
there's uh
some examples of uh these deliberation
maps or
deliberation approaches so right now
you're looking at one this is the
deliberatorium from mark klein
and um it's a little bit overwhelming
because there's a lot of buttons and a
lot of stuff going on but what is really
nice about the example what i wanted to
show you is that
it's focused around finding pros and
cons so it's focused around
representing the um you know the key
points or the
the key content that is being discussed
here
um and then there's a bunch of options
like voting or inserting new ids
or so it's really really difficult to
navigate and it's really hard also at
adverse glance to kind of see what is
over there but i can tell you that the
content that is in here is really nice
so the content in here is really really
well structured
um so i mean social media is not all
that bad because what social media is
good at
is kind of providing a an intuitive way
to
to interact with other people right uh i
mean
everyone has used twitter um even
difficult things like uh communicating
in reply threads and and
kind of traversing this tree it it is
quite difficult but
people still manage and actually people
tend to produce pretty
uh intricate uh reply structures like
one person replying to another and then
another person replying to him well you
see people here
just tend to get sidetracked in the
entire different conversation
almost but it shows that discussions are
happening
and even these discussions you can kind
of map them out
but what it doesn't show you is the
content that is there right so you don't
know
uh you know what these people are
talking about you don't know why they're
disconnected from other people
um what problems are they raising what
arguments are they making
so this is largely still not not found
out yet
and that is the type of information that
we would like to actually contain
in our deliberation map
so so now that we kind of
saw what we want from this deliberation
map we were thinking okay so we need a
method to fill this
we need the method to fill the
liberation map
um and i talked before also about it
being autonomous
but there's also a nice part
that i can say now which is about hybrid
intelligence so
both enrique and i we are connected to
the hybrid intelligence center which is
uh like also uh luciano talked about
it's
radical grant from nwo and
as a model they have augmenting human
intellect so it's really not about
replacing humans but it's about
working alongside humans using ai as a
partner
to kind of increase what humans can do
so take the best from humans take the
best of computers
and really bring them together to move
forward
and why i mean for me
at least the y is really really obvious
because there's the promise that it's
both better for efficiency so you're
both able to faster
come to more accurate
decisions that are made you know
combining the two strengths the
computational strength and the intuitive
human side of things
but next to it there's also trust so um
once you have an uh especially when when
you're in a discussion
you want to be able to use a platform
that you can trust uh
and if especially if there's
autonomously your ideas are kind of
mapped for you in some kind of structure
you need to make sure that you have some
kind of control over it but you also
trust that the ai is doing what you want
it to do
um well his provides some kind of
way in to to structure this to reason
about it
and it does it does so through four
different research pillars
which are the the care pillars so care
stands for collaborative
adaptive responsible and explainable
and i would like to highlight for
instance explainable so you could say
well explainability what does it mean
it means that an ai algorithm has to
explain why it came up with a certain
decision
so why did you use my argument and why
did you insert it here at this point in
the map
well we would actually like to also show
that entire process of the ai making
that
that choice um so you would be able to
kind of verify
and trust what is going on there
um yeah so and on top of that
uh it also again provides a feedback
loop to the human so you would be able
to kind of look at the processor so why
did the ai think
uh i meant to say the argument
that i was making there so what did they
i think of that and
uh kind of reinforce your own idea of
what you're what you're saying
um yeah so i think at this point
is where we have uh oh wait yeah so
almost
um so we talked about
the how to help the scaling up of a
deliberation
how to help people in making sure that
we can scale up
right some of the examples of the
deliberation mapping itself and
the method that we wish to use but we
didn't talk about what we want the
deliberation map to contain
but before we go there i think we will
take some time to take a couple of
questions if there are any
great thanks michelle so far thanks
enrico uh
yeah do we have questions please you can
raise your hand
here on virtually speaking so that yeah
um
okay we have one uh ufo
okay you please correct me if i'm
pronouncing your name wrong so but you
have the floor if you want to turn on
your camera if you want or just or just
speak
please sure hi thanks for the talk
i'm audre welcher from the web
information systems
uh what was a little unclear to me was
about
what goal you were trying to optimize
for so i understand
the different needs that deliberation
can
generally serve i was wondering whether
you're trying to optimize
for say a deliberation that can inform
policy change
uh for example or is it the process of
deliberation itself that you're looking
at and i think this is
a pertinent question because this would
then determine what the metrics
that uh would evaluate your framework
would be right
could you say something about that
yeah so i like the i like the
mathematical approach of it
of this question um so we do not have
yet
any optimization in mind the idea is to
build a
platform that allows for the tools
allows for
tools for moderation tools for example
that
can be then employed and optimized for
different purposes
so one can be for example to increase
participation so to make sure that all
users
give their opinion one could be to
increase
diversity of red opinion so
it could be for example exposing
participants to
something that is actually opposite to
what they believe in
to see how this could increase their for
example acceptance or their belief in
different opinions so we do not have
anything specific in mind but
as i said the idea is to create the
tools for them optimizing
as you prefer okay
so while you're building the if i may
just follow up very quickly so while
you're building these tools right
you uh as you mentioned you probably
would want to support
certain functionalities which are hard
to achieve
in their absence you just mentioned how
uh providing a voice for every
participant is one of those aspects so
i think eventually that's essentially
what you would want to evaluate right
to check whether your proposed solution
is effective or not so i guess there you
have a handful of
ideals as you uh called them earlier
that you would look at so is that
what uh you would eventually evaluate
then yeah
those are part of this this kind of uh
toolbox as anything
puts it right so there's a
there's also a number of theories of
these yeah like like we said these
deliberative ideals so there's a bunch
of them
and different also when talk when
discussing this platform there's
a different kind of design
characteristics that you can
employ and that have a different impact
on each of these ideals right
um and what is also unclear is the
interaction between them
so uh what if i increase the the
strictness of moderation you know
how does it affect all of these
different ideas that are there so that
interaction is completely unknown so
this is because it's
relatively new and especially to do this
online is people don't know about it yet
but it's it would be nice to kind of um
you know we're starting to build this
from the ground up and as we go along
we'd like to
you know put some products and make some
experiments here and there to kind of
see
to get a feel for what the interaction
would be like sure
that is the the main idea yeah that's
interesting uh
if one final comment just to uh share
this view with you and then get out of
here
to let others get a chance as well uh so
in our group we're working a little bit
on trying to understand collective
action as well in crowd computing
communities so
think of online labor uh mechanical turk
so on and so forth
and i think you could probably rely on
such platforms to
try and understand team dynamics and you
know perhaps some of the other
ideals that you can quickly
operationalize in into an experimental
framework
uh you know to try and understand how uh
different features that you might want
to build within tools can
actually play a role in you know
bringing these
ideals to the fore within your
deliberation uh
we can chat offline more about this but
yeah yeah yeah definitely
yeah definitely just send us an email we
should definitely talk about that
cool thank you cheers thank you thanks
super interesting discussion thank you
uh lou
you have uh you have a question you
wanna
yes yeah it's it's kind of in line with
uh with uh
questions and recommendations i think um
i got
a little bit lost towards the end to
what the actual problems are that you're
trying to
um uh trying to address um
and i got i have a better idea now after
your answers to who you are so it's like
it's a really
fascinating uh project that you're
setting up
but i would really like maybe the
question would be like what what's like
a really small practical problem that
you think
is possible to solve or that's maybe
that you would typically do
like in in a physical presence with a
group of people
and how would you translate that uh to
an online environment and then
um plus one and what we all said about
um
you know that there's lots of um
experimenting over under like really
interesting communities like uh
the mechanical turk workers uh there's a
couple of people there that i could
also put you in touch with that have
been doing this kind of work for uh
for a long time like how to organize
workers online and make sure that they
they can align on certain issues
um in a fairly distributed way so that's
i guess my question for you like i'm
happy to help they're offline like
offline but
my question is like do you have a couple
of like really practical small problems
in mind or is that maybe a next step for
your
for your research i mean that is uh
what we'll get to in the second part of
our presentation as well
so thanks for the question because it
nicely gives us a segue maybe
but um so we kind of uh viewed this as
or at least as how i understand it is we
kind of took a couple of these ideals
and we kind of zoomed in on them um
for instance what i'm zooming in on is
so i put it down here as perspectives
but what it's really about is about
arguments so i'm really looking closely
about the reasoning that people do
what arguments are they raising uh also
kind of looking at what what is
available in ai right so
uh how far along is is natural language
processing how far along is argument
mining
uh to what degree can we use what is
there reliably
and uh but that's really zooming in on
only like one of these ideals
so that's just zooming in on reasoning
so for me
that is how i make it practical and how
i
kind of tend to scope it off uh i don't
know about
if enrique has a similar uh yeah i think
i think
to to answer your question one idea
practical idea that we have
is for example what happens if we
create live this deliberation map how
does this
and the participants can see this
deliberation map containing the
arguments that are put
pro or against a certain topic of
discussion how does this influence
the the discussion is the discussion
going to for example converge and be
and have less repetitions simply uh is
it going to actually produce
more arguments simply because you know
there's no repetition so this uh
for example a simple first step
and what mimi he'll especially work on
is the
creation of this deliberation map so how
do we create these in
autonomous hybrid intelligence ways
all right thanks yeah i think like what
might help is to
maybe like because this i think there
are
plenty of organizations a lot of online
deliberation of course happens on social
media platforms so we don't really have
access to understand like
what they how they like that don't do
much of it and there's not
not a lot of collaboration um but it
might be interesting to see if there are
um also policy makers who are thinking
about like what
we want to see and that might be a good
way to kind of
put emphasis on because like these are
very practical very urgent questions
but they're super scientific as well
because they're yeah it's never been
done before so
that's just uh encouragement is to see
see if you can seek maybe partnerships
because
you're setting it up as a big big thing
like try to find partners who can help
you kind of make it um
more specific because it will still be a
highly uh scientific
endeavor right we are we are
yeah we are collaborating with nick
mauter and the
pve people participatory value
evaluation
so we're work we're using their data
we're
helping them and then you know as
luciano included
have ideas to you know make them
scale scaled long i'm not sure like
which direction of scaling this would be
but anyway um
so we're already working with them and
yeah we can talk also
we can go in details about those ideas
yeah yeah
it's indeed a challenge right how to
combine that this is the level
abstraction that you can allow you to
generalize and face different projects
but at the same time be
concrete to have some more uh down to
earth kind of solutions it's always
challenging and yeah i had a question
for you myself but i will keep it to the
end because i think that's a very
smooth transition for you so
i think maybe you can continue with the
with the slides and then we can open up
for questions again later on
sure yeah okay all right
so um back to our deliberation map
and the the content that we want to
store
inside of it so there's two things that
enrique and i are working on so the
first is perspectives
that's well the title of my research and
the second one is
our personal values and eric will tell
you later more about the second one but
first
a little bit about perspectives so um
the the word in itself perspective it's
many it has many interpretations but the
one that i
choose to use here is the one of
perspective taking
so the actively contemplating other
psychological experiences
uh so informally it's like putting on
your social classes and
kind of look at the discussion actively
place yourself
in another person's mindset and and look
at how that person might be
viewing the discussion or might be kind
of weighing off its options um
so this this perspective taking is a
significant mental effort
um but it has been shown to increase
reciprocity and empathy
and all these again so these these
relate back to the deliberative ideals
that there are
um and um if if you get
everyone to do this and also uh to to
make sure that they provide their own
perspective so
people should really be or not really
forced but encouraged
to to uh provide their perspective
on on the topic that's being discussed
um you you will start to find these
these problems and criteria and
solutions that
you didn't know about beforehand so
these uh specifically one thing that
enrico mentioned was this uh
independence so that you look uh at what
the other people think before you give
your opinion and you
just you're being a little bit uh
related to what the other people think
but if you're
like one step ahead of that um you you
will find out what these these criteria
are that people actually worry about
privately
um and it's highly related to the the
also the the reasoning ideal
so the uh in in also providing their
perspective
these people they should be uh
to a high degree uh providing arguments
and providing
uh you know why are they making
conclusions and um
you know specify all of this and the way
we then look for these perspectives or
how i plan to look for these
perspectives is
basically by asking uh who says what
and uh the what in this case are
arguments
so i'm trying to make it really concrete
where arguments are kind of the reason
for having a certain opinion
and you would think maybe this is easy
but actually it's not
so in the simple case you a person is
perfectly straightforward in
uh giving his or her argument but in
reality it's really complex
uh there's a lot of you know implicit
information be
also being captured in the in the
arguments uh
here there's uh uh for the first two
lines so any doctor who tells an
individual patient
that all vaccines are safe and effective
is lazily parroting the cdc
so in this especially in the words
lazily parroting right so there's an
implicit meaning that
a computer might not be able to tell
kind of right off the bat
um but next so next to this implicit
meaning
just even finding the arguments and
finding the entailment between arguments
and conclusion
uh and finding their interaction is
already really difficult
so um so that is kind of what
what for now what i'm focusing on and
also what i answered to
to ruiz i'm kind of taking this uh
reasoning ideal and trying to zoom in on
it and really
try to get to the core problem that is
there and see how far along ai
is and and see whether we can actually
use that in the
the broader scope of deliberation
um so this is about the what so next up
there is also the the the who
so the who is a i mean it's uh related
so who is having the argument who
is stating it um uh
what other parties or stakeholders are
they mentioning
so i realized here the the other
stakeholders that are being mentioned
and you get into complicated cases where
there's also embedded
embedded evidence or embedded statements
or
this very epistemological kind of
logical like what you think that he
thinks that he says
kind of cases so that's a it's a really
complex topic but
i'm trying to um i mean i can if there
are questions i can go into a little bit
more detail but i try to keep it on a
high level for now
um yeah so
i mean when it boils down to it it's
very linguistical challenge um
and also like a before we also wish to
use
hi for this so make it hybrid so really
rely on the human intuition to make
the decision where it matters but also
employ ai for the
the straightforward computations um
yes enrico yeah i think you're up i can
uh yeah i can just
uh again um
right so perspectives are who says what
and values are why do people say
something
to keep it simple so values are the
abstract drivers that are behind
our opinions and our actions and our and
our beliefs
so um what you see over here on the
right is an example of personal values
and this is an example of basic values
so these values are general values that
have been found
throughout the research and
and essentially these are the
motivations for our opinions for our
actions but as you can see here these
values
at least we believe that these values
are a little bit too
general a little bit too generic such as
self-direction or achievement and they
are very difficult to
connect to everyday situations and that
is why we want instead to use
context-specific values so
context-specific values are instead sort
of the other side of the coin
and they are defined within a context
then applicable to a context so let me
let me make a couple of examples so um
let's take the example of
physical safety this value this value is
very important in the context of
lockdown measures because
people have people take actions based on
the motivation
based on the driver of physical safety
as opposed to others who
for example do it based on freedom
so it is a value that is applicable to
the context of lockdown measures
but it is for example not applicable to
the context of regulating financial
markets
arguably however the very same
value of physical safety can take
different forms in different contexts so
let's say that if we talk about
physical safety in the context of
lockdown measures then we talk about
maintaining distance we talk about
wearing face masks but if we talk about
about the same value physical safety in
the context of driving
then we think about wearing safety belts
we think about respecting speed limits
and signs
now these two are essentially from a
natural point of view they're the same
value but they take different flavors
different forms
in different contexts and if we want to
concretely talk
or if we want to concretely design and
talk about a specific
value in a specific context well we
believe that we should use the context
specific value
so let's make for example if we want to
create an
agent that throughout your day helps you
in
increasing your your physical safety
well that agent
should not suggest you to wear a face
mask
while you're driving in order to
increase your physical safety that would
not make sense
the agent needs to understand the
context where the value is located in
and that these values are different
so uh well the problem is that if we
talk about the context-specific values
the first step is then
well finding out which values are
relevant to a specific context when
we're talking
and in order to do this this was the
first step of my phd
meme here luciano and some other
colleagues developed a methodology a
hybrid intelligence methodology
for uh identifying the values that are
relevant to a specific context
so i essentially using human intuition
and supported by
artificial intelligence i i'm not going
to go in detail but
you can contact me for additional
details
and once we have identified which values
are relevant to a context
we can use those values essentially as a
language as
as a language of communication to
interpret what people are saying or are
trying to say
so um instead of having
deliberation participants communicate
through their text
purely through their text we can sort of
filter this text and using a transparent
box i'll get to that in a moment
we can instead to the listeners we can
give
the comment that the actual person gave
and the perspective
the values that that person
holds with that comment the idea is that
if we give a better interpretation of
what people are saying or are trying to
say
the on the listener side hopefully
there's going to be a better
understanding of different opinion and
there's going to be a
calmer discussion instead of just you
know rushing to thinking that
well that person must be stupid because
well she thinks
differently from me so she must be
stupid so hopefully
we if we have a better if we give a
better detailing of what is said
in this additional abstract level like
perspective well some more concrete
perspective and
abstract values levels that would be a
better understanding
of this conversation and as we say as
mikiel mentioned earlier
all of these should be done in a
transparent manner this algorithm that
interprets
these comments that finds the
perspectives and finds the values
should be as transparent as possible
users participants should have control
over what the algorithm
thinks about what they say in this way
if they have some meaningful control
over this algorithm they will build
trust in it
and they will be able to sort of use
this platform to discuss and hopefully
have a better informed
and civilized discussion about topics
so essentially this is our this is well
one of our final goals to attach to one
of the questions
from earlier and well we're gonna get
there
at a certain point so that was uh that
was the
that was the presentation and yeah just
ask some more questions
luciano you said you had one cool so
thanks indigo thanks michael that was
really interesting
and yeah let me just add one comment
a small comment before that the this
work he mentioned was really so
on the it's called axis right it was
accepted for amaz
two thousand 2000 2021 and
and it deals like with two different
contexts one of the concepts of energy
transition
another one of the relaxation of covid
measures together with the group from
yeah
midnight malta some other people at the
tpm on the participatory value
evaluation so
definitely as soon as out i recommend is
interesting
and we are also trying to understand not
which are the values and there's another
work we are also trying to
to see how how much each individual care
about each value so a little bit where
something so uh on my question is
you you mentioned on the earlier
presentation
about uh deliberation being a
way of rational discussion so to
bring rational arguments but it comes
it's like
we are messy let's not call irrational
but
at least boundedly rational human beings
and i i just want to say how do you
you have some ideas to reconcile this
because if you want to see
only rational arguments but we are not
always rational on the way we propose or
arguments there's a lot of emotions
there's there are feelings there maybe
the connection even to values that we
can say is that really rational or not
so
just you have some ideas on that aspect
yeah so there i mean the uh this this
feeling of the deliberation map right so
we gave perspectives and and values now
as a
a kind of way to fill it but those are
kind of uh um examples
maybe perhaps of things that you could
put in this
uh this map and what we're trying to do
is we're trying to get to a point where
we say all right so
the technology that we employ here or
the ai or hi or however
you extract the perspectives or values
we get it to a point where we can work
with it and we can kind of on a high
level
use it for for helping people in a
deliberation
um but next to it there might be many
more
there might be many other types of
information you can strike
things like sentiment things like
emotion things like you know all these
different
aspects of what people are saying you
could uh
i mean for now also the the idea that we
have for the representation of this map
it allows for these different uh types
of information to be included and and
actually persist next to the
perspectives and values that we
mentioned for now
um and i mean while i'm looking at so i
mean you're writing saying well i'm
trying to look at this rationality and
i'm kind of saying well
it's so i i'm hoping it's going to be a
self-reinforcing loop right so i'm
trying to look at the liberation as
there should be rational discussion
going on so i then
present to you the rest of the
discussions should be become even more
rational right
but it doesn't mean that i kind of
ignore all the other
stuff that is there but it's just one
way of scoping off the project for
now but if it turns out that it's really
difficult to do or
i kind of lose people during the process
then
you need to critically look and say well
maybe the the rationality
constraints that we have for now are too
strict so let's broaden up a little bit
let's maybe go
one step lower uh in in trying to to
model it or trying to
extract what people think or how they
think
and maybe shift a little bit in in what
we exactly mean with
uh with the perspective for instance
yeah okay
nice if you think thank you very much i
think that clarifies a lot and
i mean is i i do believe that uh
this combination of human and ai that is
really the way to go for these kind of
topics and issues that you can still
have the goal of achieving rationality
but still including everyone
along the way right okay uh does someone
has
and any questions right now can you
can you raise your hands please
um ben you suggested on here in the chat
an
article for enrique i don't know if you
want to just show me
that quickly i could comment about it
yeah i could say why i
thought of it um hi thanks for the um
presentations very um
fascinating i was thinking when you
mentioned black to
transparent that um
that was sort of i think he used the
words in terms of
um just being able to tell
what the algorithms are doing but i
i'm sort of sensitive to also the sense
of
maybe the uh
the perspectives from which we're trying
to or the tech is trying to understand
the values
and i think how we understand what
values are
in a situation are is affected by
the implicit values in how we're looking
and there's probably some care um
over notions like transparency
um that we don't suddenly think that we
have
actual descriptions or transparent or
white descriptions of the values but
there's some kind of essential ignorance
still
within the overall situation um
yeah so i just suggested i thought that
really that
that paper sort of talks about black and
white boxes in a really nice way and i
thought it might um
be relevant to the sort of bigger
picture you're trying to deal with
yeah yeah yeah thank you thank you
absolutely yeah just uh to quickly
mention uh
so what we tried to do for example in
the first
in our first project to identify which
values were relevant to a discussion
to give an example of how we tried to
let's say
have humans in the decision-making power
so the idea that we had was to we took
answers to a survey
as luciano said about two surveys one
about energy transition on about
lockdown measures and we tried to
guide human annotators so
mostly they were policy experts or value
experts
to guide them through this data set and
try to find the values
in what people wrote so we had some
guidelines and try to
invite them to see values only in what
they wrote what
the participants wrote instead of just
um
coming up with these values on their own
so sort of what this means is that we
can
ultimately use let's say also trace back
these values from what people had
written originally in this survey and
that should sort
somehow give a little bit more of a
transparency and
at least making it grounded in data and
where data is what actual people had
written
so this is uh this is just an example
but anyway
yeah anyway it's tough to make
transparent boxes
yeah thanks for that um uh i think that
the kind of being able to kind of trace
the process is really valuable like that
that sounds like a i think that could be
some there might be different ways that
process happens and it might be
interesting to compare
um compare those so the different ways
that values get read
would be an interesting topic yeah
yeah okay so ben thank you very much for
sharing i think that's
yeah looks also interesting reference
and not yeah for
for me i believe for some others also
here in this meeting
in this talk so uh do we have more
questions
marcos please go ahead hey
hello uh sorry for not turning on my
camera i don't know it just
don't work no problem okay
um i i i'm i was thinking when you were
talking about um context specific
um personal values are taking this into
account
and if you are considering that some
collective decisions
they need to match match diverse
deliberations
not very much uh just help every person
or group to
deliberate or create a consensus
or that calls for every
specific deliberations but um
showing how uh different group
deliberations can be met
together to um
to make a collective decision i don't
know if you forget what i'm i'm trying
to say
not really okay well um
i always give this example when you're
designing a neighborhood for example
a lot of people have different desires
you have to
match all those desires most as you can
for designing this neighborhood for
example
so it's more than just giving people
what are all those values involved
and the conquest of every value but
the decision the final decision needs
the
the match of those
and it's different from a consensus it's
not a consensus of
what is better for all or some like this
but to show people that they can't be
matched
in a way you know and i think this
artificial intelligence
can help maybe and when you talked about
specific content uh values and showing
this
i think this maybe can get in like this
specific way of helping collective
decisions
that's why i'm posting this
yeah yeah no yeah absolutely that's also
why
we think like uh you know nick mata
and all those the the
people behind the participatory valley
evaluation that's uh that's essentially
what you're talking about in a way
so the idea is to instead of under
instead of
uh take what people would prefer
it's like the policy options that they
would prefer or for example
how they would prefer to relax lockdown
measures
instead of reading those they read the
values that people
the the reason why people would like to
relax certain
policy measures and this is much more
informative to the policy makers
especially because especially in the
situations where
things change rapidly like well in
corona
so understanding the values behind
people's decisions
helps so that's and also essentially
you would instead of aggregating for
preferences you would aggregate for
values
and that would be very helpful for uh
sorry that would be very helpful for a
moderator or policy maker or
neighborhood designer
thank you very much great so
i think yeah we we we ran out of time so
yeah we are almost running out of time
um
so i think i would like to yeah thanks
enrico
thanks michael for the very interesting
talk
and yeah thanks everyone for joining us
and
see some of you again next week so
all right thank you and feel free to
contact us personally for any kind of
questions really
we're open to discuss |
4c5a2a5b-a9ae-403b-94f3-b6ba7e1f9e31 | trentmkelly/LessWrong-43k | LessWrong | The economy is mostly newbs (strat predictions)
If the world economy were a game where the objective is to gain resources & power, then I would say that we're collectively newbs. We're presently protected from some of the extremes of capitalism (etc) by the fact that the board (etc) does not actually MAXIMIZE shareholder value.
As tech advances, the more ambitious players will be able to take more resources more easily. The sleepier corps will lose almost all non-govt-protected markets and die. The world will be a more pushy, grabby, competitive, profitable place.[1] People will not foolishly leave so much money on the table so often.
A few examples of recentish advances in strategy to show what I mean:
* Mobile game companies realized they can be casinos where you never actually win money. In 2022, the mobile game market was ~$180B vs ~$40B for console or ~$30B for PC.
* All the apps realized subscriptions are more profitable than sales.
* Anyone who can possibly offer a credit card (Apple, Lowe's, airlines, Amazon...) does now. For airlines, the cards are ~1/3ish of profit.
* Landlords in some cities are effectively fixing prices by ubiquitously using a "recommended market-rate price" from an app.
* Google's aquire-and-kill approach seems to be operating successfully at unprecedented scale (1 company per weekday in 2022)
Which new strategies will be viable depends much on regulation & the order of tech advances, but here's my bets for how the state of the game will change over the next ten years.
* better pre-IPO PR
* ^ partly enabled by social network AI influencer swarms available for hire
* more price "recommendation" apps across industries
* more & better phishing & scams ofc
* misleading loans everywhere, especially construction & renovation[2]
* more we-don't-tell-you-the-price like hospitals are somehow permitted to do
* more acquisitions across industries — why shouldn't Microsoft buy an oil company?
* more sea steading and secret private militaries (something being a common joke do |
74e3eec2-127a-43f3-af4a-c78bce0899a9 | trentmkelly/LessWrong-43k | LessWrong | Combat vs Nurture & Meta-Contrarianism
My initial reaction to Combat vs Nurture was to think "I already wrote about that!" and "but there should be three clusters, not two". However, looking at my old posts, I see that my thinking has shifted since I wrote them, and I don't describe the three clusters in quite the way I currently would. So, here is how I think about it:
* "Face Culture" / "Playing Team": When people offer ideas, their ego/reputation is on the line. It is therefore important to "recognize the value of every contribution" -- accepting or rejecting an idea has a strong undercurrent of accepting or rejecting the person offering the idea. This sometimes makes rational decision impossible; the value of team members is greater than the value of the specific decision, so incorporating input from a team member can be a higher priority than making the best decision. Much of the time, bad ideas can be discarded, but it involves a dance of "due consideration": an idea may be entertained for longer than necessary in order to signal strongly that the team valued the input, and downsides may be downplayed to avoid making anyone look stupid. Eventually you want someone on the team to point out a decisive downside in order to reject a bad idea, but ideally you coax this out of the person who originated the idea. (I called this "face culture", but I have heard people call it "playing team".)
* Intellectual Debate: It is assumed that engaging with an idea involves arguing against it. Approving an idea doesn't necessarily signal anything bad, but unlike face culture, arguing against an idea doesn't signal anything bad either. Arguing against someone (genuinely or in a devil's-advocate position) shows that you find the idea interesting and worth engaging with. Concepts like burden of proof are often applied; one tends to operate as if there were an objective standard of truth which both sides of the debate are accountable to. This warps the epistemic standards in a variety of ways, for example, making it |
694d2896-3fd4-4a89-a5ae-4d144a6af775 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | The Epistemic Authority of Deep Learning Pioneers
Introduction
------------
In response to recent advances in machine learning and the subsequent public outcry over the safety of AI systems, deep learning pioneers Yann LeCun, Yoshua Bengio, and Geoffrey Hinton (henceforth “the pioneers”) have spoken up about what they perceive to be the likely dangers as AI progress continues. Hinton and Bengio have advocated for AI safety in the same vein as the alignment community, while LeCun staunchly remains unconvinced that AI could pose an existential risk to humanity. The actions of Hinton and Bengio have lent credence to the alignment community as a whole, as one of the main criticisms of the movement had been that few modern machine learning experts endorsed it. Anecdotally, a friend of mine who had previously been unconvinced by the alignment community’s arguments was swayed by Hinton’s words this past May.
This context raises the central question of this post: how much credence should one lend to the opinions of these deep learning pioneers on AI risk over other classes of experts? Was it epistemically sound for my aforementioned friend to shift his views so drastically due to the input of a single expert? And more specifically, should we expect the intuitions and knowledge of experimentally successful AI pioneers to generalize to predicting outcomes of artificial superintelligence?
Background
----------
**Geoffrey Hinton** received a PhD in artificial intelligence from the University of Edinburgh in 1978. The history of the backpropagation algorithm (the central optimization tool in deep learning) is sort of murky but it’s generally agreed that Hinton, along with David Rumelhart, helped to popularize it in papers during the late 1980s. Hinton has primarily worked at the University of Toronto, where he was most notably a co-author on the 2012 AlexNet paper that kicked off the current boom in deep learning. He also co-authored papers on dropout and layer normalization, common regularization techniques in deep learning (both are used in the GPT family of models).
After completing a PhD in computer science from McGill University, **Yoshua Bengio** worked as a postdoc at Bell Labs where he assisted Yann LeCun on using backpropagation with convolutional neural networks to implement a handwritten check processor, one of the first big breakthroughs for neural networks. As a professor at the University of Montreal, he was the PI on the original GAN paper and introduced the attention mechanism to language translation tasks, paving the way for the first transformer paper.
**Yann LeCun** received his PhD from Sorbonne University in 1987 and joined Hinton’s research group as a postdoc afterwards. He then joined Bell Labs to lead the aforementioned handwriting recognition project and pioneered an early sparsity effort called optimal brain damage. In 2003, LeCun left Bell Labs to become a professor at NYU, where he continued to work on computer vision, specifically within robotics. In 2013 LeCun was appointed as head of research at Facebook AI.
Evaluation
----------
One reason to lend credence to these pioneers is their knowledge base within deep learning, having worked in the field for 30-40 years each. However, this only lends them as much credence as any other deep learning *expert*, and for the details of modern state-of-the-art models they fall behind the researchers actually building such models. The [SimCLR paper](http://proceedings.mlr.press/v119/chen20j/chen20j.pdf) with Hinton as PI is the only big paper that any of the pioneers have authored in the past five years. This provides us a baseline–the pioneers should *at least* be provided credence as general deep learning experts, though not credence as cutting-edge experimentalists.[[1]](#fn94ifh88hjgd)
It’s also important to note that among the broader class of AI researchers, the pioneers are some of the only researchers that actually got anywhere experimentally. When the pioneers began their careers, symbolic AI was still the dominant paradigm, and their intuitions in this domain were what guided them towards neural networks. From a [2023 interview](https://www.technologyreview.com/2023/05/02/1072528/geoffrey-hinton-google-why-scared-ai/), “‘My father was a biologist, so I was thinking in biological terms,’ says Hinton. ‘And symbolic reasoning is clearly not at the core of biological intelligence.’” Furthermore, they stuck with deep learning through a dry period from the late 1990s to the mid 2000s where there were very few developments in the field. On this side of the matter, the pioneers should be provided credence as “AI researchers whose predictions were accurate.”[[2]](#fnugsnn3qr13)
However, the pioneers’ intuitions might still be misguided, as it seems their initial inclination to work with neural networks was motivated for the wrong reasons: the efficacy of neural networks (probably) comes not from their nominal similarity to biological brains but rather the richness of high-dimensional representations. This wasn't well-understood when the pioneers entered the field; at the time, neural network approaches were rooted in cognitive science and neuroscience. While some concepts in deep learning have biological analogues, like convolutional layers, other tools like backpropagation bear little similarity to how our brains learn. Therefore, one should be wary not to award the pioneers too much credit for the later successes of neural networks.
And like pretty much everyone else, the pioneers did not predict the acceleration and course of AI progress within the last 3-5 years. While this does diminish their epistemic authority, it simultaneously reduces the credence one should have in any source trying to predict what the course of AI progress will look like. There are no real experts in this domain, at least not to the same level as the physicists predicting the capabilities of the atomic bomb, or even climatologists forecasting the rate at which the Earth will warm.
Thus, when looking to update one’s credence on AI safety by deferring to expert judgment, **one should weigh the input of the three deep learning pioneers more than most other sources, but in general the weight placed on any individual expert should be less for AI x-risk forecasting than most other problems**. The friend mentioned at the beginning should not have updated so steeply, but it’s understandable why he (and the general public) would, because analogous experts in other domains carry much more epistemic authority.
1. **[^](#fnref94ifh88hjgd)**Relatedly, one should have very low prior credence in schemes where some lauded AI researcher comes up with some grand unified scheme behind neural networks/intelligence as a whole, like LeCun’s I-JEPA. The older generation of deep learning researchers, like the pioneers, Jurgen Schmidhuber, etc., haven’t made any monumental discoveries on the cutting edge for quite some time so it’s unclear why their approaches should be lent substantial credence given the magnitude of the claims about these systems.
2. **[^](#fnrefugsnn3qr13)**Using similar logic, one should *a priori* discount the claims of non-DL AI researchers like Judea Pearl ([pro-risk](https://twitter.com/yudapearl/status/1695509566015082894)) and Melanie Mitchell ([anti-risk](https://www.youtube.com/watch?v=qLVTc1IUHPE)) relative to the pioneers, since Pearl and Mitchell’s approaches to AI (Bayesian networks and genetic algorithms, respectively) haven’t gotten anywhere near the capabilities of deep learning. It’s also doubtful that these approaches are compute-limited in the same way that neural networks were for ~15 years, so it’s unlikely that we see large capabilities developments in the same way that we saw a jump for deep learning. |
bec72a0e-3c18-4b96-b339-443a9f5cefb1 | trentmkelly/LessWrong-43k | LessWrong | Search, and ye shall find
Hey, reader, how did you get here? Are you subscribed to Putanumonit? Clicked on a tweet or a status or a post? Followed a trail from SlateStarCodex or LessWrong? Either way, you probably know what to expect here. And, you deserve what’s coming to you.
But some people took a more interesting path to this blog. According to WordPress Stats, back in January someone googled ‘nextpart and hack’ and the 7th page of the results took them to my post on water charities. Others arrived on Putanumonit after googling ‘gern’, ‘sci-put’, ‘utrgv “wewillfail”‘ and ‘ya fat njie’. I have no clue what any of these people were looking for , but I doubt they found it on Putanumonit. And if there’s anything I really hate is leaving a client unsatisfied.
I decided to go through the list of search terms that brought people to the blog, and help them out with their unanswered queries. At the end, I’ll share my absolute favorite link to Putanumonit, one that by itself would have made the hundreds of hours I spend on this site worthwhile.
----------------------------------------
There but for the grace of Google go I
Original search terms in blockquotes.
> messi before growth hormones
> messi b4 when messi was a dwarf
He was actually a hobbit.
Credit: tiobolasdoro, via cheeksoftheweek
> lotto king
That’s what my LinkedIn profile says.
> “intelligent life economist”
That’s what my Tinder profile says.
> second option to keep 3galloping horse in tge living room of narth facing house
I don’t know what the first option was, but I would go with that one.
> tribalism in the old testament
How about Samuel 15:3 – “Now go, attack the Amalekites and totally destroy all that belongs to them. Do not spare them; put to death men and women, children and infants, cattle and sheep, camels and donkeys.”
> March 2016: is an ugly guy doomed as a pua
> June 2016: dry spell pua
I guess you got your answer. Maybe you should try a more considerate and respectful approach to dating?
> August 20 |
0ca4233c-3ff4-4848-999c-534167700230 | trentmkelly/LessWrong-43k | LessWrong | GroupThink, Theism ... and the Wiki
In response to the The uniquely awful example of theism, I presented myself as a datapoint of someone in the group who disagrees that theism is uncontroversially irrational.
With a loss of considerable time, several karma points and two bad posts, I now retract my position.
Because I have deconverted? (Sorry, but no.)
I had a working assumption (inferred from here) that rationality meant believing that all beliefs must be rigorously consistent with empirical observation. I now think of this as a weak form of rationalism (see full definition below). A stronger form of rationalism held by (many, most?) rationalists is that there is no other valid source of knowledge. If we define a belief system as religious if and only if it claims knowledge that is independent of empirical experience (i.e., metaphysical) then it is trivially true that all religions are irrational -- using the stronger definition of rational.
A disagreement of definitions is not really a disagreement. Someone suggested on the April open thread that we define "rationality". My idea of a definition would look something like this:
Rationality assumes that:
(1) The only source of knowledge is empirical experience.
(2) The only things that are known are deduced from empirical experience by valid logical reasoning and mathematics.
Weak Rationality assumes that:
(1) The first source of knowledge is empirical experience.
(2) The only things that are known with certainty are deduced from empirical experience by valid logical reasoning and mathematics.
(3) Define a belief system as all knowledge deduced from empirical observation with all metaphysical beliefs, if any. Then the belief system is rational (nearly rational or weakly rational) if the belief system is internally consistent.
Probably these definitions have been outlined somewhere better than they are here. Perhaps I have misplaced emphasis and certainly there are important nuances and variations. Whether this definition works or |
aa2b5b4f-06a9-4316-85f0-d0e421b9e9f1 | trentmkelly/LessWrong-43k | LessWrong | Tessercube — OpenPGP Made Mobile
I cited several words from my colleague Neruthes. He's not at LessWrong but he also joined the most recent Shanghai LessWrong meetup: https://www.lesswrong.com/events/zR4atrRmiaqGLvjYj/shanghai-lesswrong-meetup#NZHcXphJXcsTFwpmf
----------------------------------------
Recently we have been working on this project in the sense that I feel there is no good OpenPGP utility on Mobile (especially iOS). By good, I mean the UX should be good, and the license should be AGPL or GPL.
In the process, we got the idea of App Penetration — by making OpenPGP into keyboards (input methods), we can literally be end-to-end encrypted when using any channel of communication, as long as the other side can decrypt — on Facebook Messenger, on Telegram, on iMessage, whatever.
For now, we have been releasing Android beta test versions on Google Play. The iOS version on App Store. It might be a bit early to announce because there are plenty of bugs and a big shortage of tutorials, but I believe hardcore users can go through it.
There can be a lot of bugs and UX flaws. If you find any bug, just go to GitHub and open an issue. And I will appreciate!
----------------------------------------
In larger perspective, building Tessercube is just a humble beginning. We would like to give general public proper tools of encryption and make them possible to protect their privacy and enable the people to really own their data. That's why we also made Maskbook , an encryption and programmable layer on top of all existing giants, such as Facebook, Twitter, etc. I will write a separate post for our story and our approach.
([I:b]) |
bf1acd6f-dde8-4095-a648-b1f9b3bdb10a | StampyAI/alignment-research-dataset/blogs | Blogs | Conversation with Tom Griffiths
### Participants
* [**Professor Tom Griffiths**](http://cocosci.berkeley.edu/tom/index.php), Director of the Computational Cognitive Science Lab and the Institute of Cognitive and Brain Sciences at the University of California, Berkeley.
* **Finan Adamson**, AI Impacts.
**Note**: These notes were compiled by AI impacts and give an overview of the major points made by Professor Tom Griffiths.
They are available as a pdf [here](http://aiimpacts.org/wp-content/uploads/2016/09/AConversationwithTomGriffithsFinal.pdf).
### Summary
Professor Tom Griffiths answered questions about the intersection between cognitive science and AI. Topics include how studying human brains has helped with the development of AI and how it might help in the future.
How has cognitive science helped with the development of AI in the past?
------------------------------------------------------------------------
AI and cognitive science were actually siblings, born at around the same time with the same parents. Arguably the first AI system, the Logic Theorist, was developed by Herb Simon and Allen Newell and was a result of thinking about the cognitive processes that human mathematicians use when developing proofs. Simon and Newell presented that work at a meeting at MIT in 1956 that many regard as the birth of cognitive science it was a powerful demonstration of how thinking in computational terms could make theories of cognition precise enough that they could be tested rigorously. But it was also a demonstration of how trying to understand the ways that people solve complex problems can inspire the development of AI systems.
How is cognitive science helping with the development of AI presently?
----------------------------------------------------------------------
When I think about this relationship, I imagine a positive feedback loop where cognitive science helps support AI and AI helps support cognitive science. Human beings remain the best examples of systems that can solve many of the problems that we want our AI systems to solve. As a consequence, insights that we get from studying human cognition can inform strategies that we take in developing AI systems. At the same time, progress in AI gives us new tools that we can use to formalize aspects of human cognition that we previously didn’t understand. As a consequence, we can rigorously study a wider range of questions about the mind.
How can cognitive science help with the development of AI in the future?
------------------------------------------------------------------------
### Deep Learning Systems
Deep learning systems are mastering a variety of basic perceptual and learning tasks, and the challenges that these systems now face look a lot like the first important stages of cognitive development in human children: identifying objects, formulating goals, and generating highlevel conceptual representations. I think understanding how children do these things is potentially very relevant to making progress.
### Efficient Strategies
One of the things that people have to be good at, given the limited computational capacity of our minds, is developing efficient strategies for solving problems given limited resources. That’s exactly the kind of thing that AI systems need to be able to do to operate in the real world.
What are the challenges to progress in studying brains as they relate to AI?
----------------------------------------------------------------------------
### Birds and Planes
One important thing to keep in mind is that there are different levels at which we might see a correspondence between human minds/brains and AI systems. Critics of the idea that AI researchers can learn something from human cognition sometimes point out that the way jet airplanes work has little relationship to how birds fly, and in fact trying to mimic birds held back the development of planes. However, this analogy misses the fact that there is something important that both jets and birds share: they both have to grapple with aerodynamics. Ultimately, we can see them both as solutions to the same underlying physical problem, constrained by the same mathematical principles.It isn’t clear which aspects of human brains have the best insights that could cross over to AI. Examples of places to look include the power of neurons as computational units, the efficiency of particular cognitive strategies, or the structure of the computational problem that is being solved. This last possibility — looking at abstract computational problems and their ideal solutions — is the place where I think we’re likely to find the equivalent of aerodynamics for intelligent systems.
What blind spots does the field of AI have that could be addressed by studying cognitive science?
-------------------------------------------------------------------------------------------------
I don’t think they’re blind spots, they are problems that everybody is aware are hard things like forming highlevel actions for reinforcement learning, formulating goals, reasoning about the intentions of others, developing highlevel conceptual representations, learning language from linguistic input alone, learning from very small amounts of data, discovering causal relationships through observation and experimentation, forming effective cognitive strategies, and managing your cognitive resources are all cases where we can potentially learn a lot from studying human cognition.
How does cognitive science relate to AI value alignment?
--------------------------------------------------------
### Theory of Mind
Inferring the preferences or goals of another person from their behavior is something that human children begin to do in infancy and gradually develop in greater sophistication over the first few years of life. This is part of a broader piece of cognitive machinery that developmental psychologists have studied extensively.
What risks might be mitigated by greater collaboration between those who study human brains and those building AI?
------------------------------------------------------------------------------------------------------------------
We’re already surrounded by autonomous agents that have the capacity to destroy all human life, but most of the time operate completely safely. Those autonomous agents are of course human beings. So that raises an interesting question: how is it that we’re able to create humancompatible humans? Answering this question might give us some insights that are relevant to building humancompatible AI systems. It’s certainly not going to give us all the answers many of the issues in AI safety arise because of concerns about superhuman intelligence and a capacity for selfmodification that goes beyond the human norm but I think it’s an interesting avenue to pursue. |
917b7bea-de43-4493-8225-4a8424f494c7 | trentmkelly/LessWrong-43k | LessWrong | How worried are the relevant experts about a magnetic pole reversal?
This article and many others like it describe a phenomenon where the Earth's magnetic poles occasionally switch directions. The key ideas seem to be:
1. Earth's magnetic poles occasionally switch places.
2. This probably happens on a semi-regular basis, and according to our best guess at the schedule, we're long overdue for a switch right now.
3. The magnetic north pole has been moving South much faster than normal recently (25 mi/year).
4. We don't know exactly what happens during a switch, but it's probably not great.
The main concern seems to be that the Earth's magnetic field is what's responsible for deflecting/weakening a lot of incoming radiation, but the magnetic field gets a lot weaker during a switching event. Then again, the article linked above suggests that the period is about 300,000 years, which is far shorter than the length of our evolutionary history. That puts a bound on how bad it can be.
Nonetheless, this all somewhat concerning to me. Who are the relevant authority figures on the subject? How worried are they that we might be on the brink of some kind of extinction event? How worried are they about the interaction of such an event with everyday electronics? Do they make any recommendations for how individual people can prepare themselves for a potential switch? How about for society? |
67d3d693-40ed-4681-b24f-2dda4ca52556 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Capsules for Object Segmentation
1 Introduction
---------------
Object segmentation in the medical imaging and computer vision communities has remained an interesting and challenging problem over the past several decades. Early attempts in automated object segmentation were analogous to the if-then-else expert systems of that period, where the compound and sequential application of low-level pixel processing and mathematical models were used to build-up complex rule-based systems of analysis. Over time, the community came to favor supervised techniques, where algorithms were developed using training data to teach systems the optimal decision boundaries in a constructed high-dimensional feature space. In computer vision fields, superpixels and various sets of feature extractors such as scale-invariant feature transform (SIFT) (Lowe ([1999](#bib.bib13))) or histogram of oriented gradients (HOG) (Dalal and Triggs ([2005](#bib.bib4))) were used to construct these spaces. Specifically in medical imaging, methods such as level sets (Vese and Chan ([2002](#bib.bib23))), fuzzy connectedness (Udupa and Samarasekera ([1996](#bib.bib21))), graph-based (Felzenszwalb and Huttenlocher ([2004](#bib.bib5))), random walk (Grady ([2006](#bib.bib6))), and atlas-based algorithms (Pham et al. ([2000](#bib.bib18))) have been utilized in different application settings.
In the last few years, deep learning methods, in particular convolutional neural networks (CNNs), have become the state-of-the-art for various image analysis tasks. Specifically related to the object segmentation problem, U-Net (Ronneberger et al. ([2015](#bib.bib19))), Fully Convolutional Networks (FCN) (Long et al. ([2015](#bib.bib12))), and other encoder-decoder style CNNs (*e.g.* Mortazi et al. ([2017a](#bib.bib15))) have become the desired models for various medical image segmentation tasks. Most recent attempts in the computer vision and medical imaging literature utilize the extension of these methods to address the segmentation problem. Since the success of deep learning depends on finding an architecture to fit the task, currently several researchers work on designing new and more complex deep networks to improve the expected outcome. This naturally brings high number of hyperparameters to be configured, makes the overall network too complex to be optimized.
### Drawbacks of CNNs and how capsules solves them
The CNNs, despite showing remarkable flexibility and performance in a wide range of computer vision tasks, do come with their own set of flaws. Due to the scalar and additive nature of neurons in CNNs, neurons at any given layer of a network are ambivalent to the spatial relationships of neurons within their kernel of the previous layer, and thus within their effective receptive field of the given input. Recently Sabour et al. ([2017](#bib.bib20)) introduced the idea of capsule networks, where information at the neuron level is stored as vectors, rather than scalars. These vectors contain information about:
1. spatial orientation,
2. magnitude/prevalence, and
3. other attributes of the extracted feature
represented by each capsule type of that layer. These sets of neurons, henceforth referred to as capsule types, are then “routed” to capsules in the next layer via a dynamic routing algorithm which takes into account the agreement between these capsule vectors, thus forming meaningful part-to-whole relationships not found in standard CNNs.
The overall goal of this study is to extend these capsule networks and the dynamic routing algorithm to accomplish the task of object segmentation. We hypothesize that capsules can be used effectively for object segmentation with high accuracy and heightened efficiency compared to the state of the art segmentation methods. To show the efficacy of the capsules for object segmentation, we choose a challenging application of pathological lung segmentation from computed tomography (CT) scans.
2 Background and Related Works
-------------------------------
Problem definition: The task of segmenting objects from images can be formulated as a joint object recognition and delineation problem. The goal in object recognition is to locate an object’s presence in an image, whereas delineation attempts to draw the object’s spatial extent and composition (Bagci et al. ([2012](#bib.bib2))). Solving these tasks jointly (or sequentially) results in partitions of non-overlapping, connected regions, homogeneous with respect to some signal characteristics. Object segmentation is an inherently difficult task; apart from recognizing the object, we also have to label that object at the pixel level, which is an ill-posed problem.
State-of-the-art methods: Object segmentation literature is vast, both before and in the deep learning era. Herein, we only summarize the most popular deep learning-based segmentation algorithms. Based on FCN (Long et al. ([2015](#bib.bib12))) for semantic segmentation, Ronneberger et al. ([2015](#bib.bib19)) introduced an alternative CNN-based pixel label prediction algorithm, called U-Net, which forms the backbone of many deep learning-based segmentation methods in medical imaging today. Following this, many subsequent works follow this encoder-decoder structure, experimenting with dense connections, skip connections, residual blocks, and other types of architectural additions to improve segmentation accuracies for particular medical imaging applications. For instance, a recent example by Jégou et al. ([2017](#bib.bib8)) combines a U-Net-like structure with the very successful DenseNet (Huang et al. ([2017](#bib.bib7))) architecture, creating a densely connected U-Net structure, called Tiramisu. As another example, Mortazi et al. ([2017b](#bib.bib16)) proposed a multi-view CNN, following this encoder-decoder structure and adding a novel loss function, for segmenting the left atrium and proximal pulmonary veins from MRI. Other successful frameworks for segmentation are SegNet (Badrinarayanan et al. ([2017](#bib.bib1))), RefineNet (Lin et al. ([2017](#bib.bib11))), PSPNet (Zhao et al. ([2017](#bib.bib24))), Large Kernel Matters (Peng et al. ([2017](#bib.bib17))), ClusterNet (LaLonde et al. ([2018](#bib.bib10))), and DeepLab (Chen et al. ([2018](#bib.bib3))).
Pathological lung segmentation: Anatomy and pathology segmentation have been central to the most medical imaging applications. Recently, deep learning algorithms have been shown to be generally successful for image segmentation problems. Specific to radiology scans, accurately segmenting anatomical structures and/or pathologies is a continuing concern in clinical practice because even small segmentation errors can cause major problems in disease diagnosis, severity estimation, prognosis, and other clinical evaluations. Despite its importance, accurate segmentation of pathological lungs from CT scans remains extremely challenging due to a wide spectrum of lung abnormalities such as consolidations, ground glass opacities, fibrosis, honeycombing, tree-in-buds, and nodules (Mansoor et al. ([2014](#bib.bib14))). In this study, we test the efficacy of the proposed SegCaps algorithm for pathological lung segmentation due to precise segmentation’s importance as a precursor to the deployment of nearly any computer aided diagnosis (CAD) tool for pulmonary image analysis.
3 Building Blocks of Capsules for Image Segmentation
-----------------------------------------------------
Figure 1: A simple three-layer capsule segmentation network closely mimicking the work by Sabour et al. ([2017](#bib.bib20)). This network uses our proposed locally-constrained dynamic routing algorithm as well as the masked reconstruction of the positive input class.
A simple three-layer capsule network showed remarkable initial results in Sabour et al. ([2017](#bib.bib20)), producing state-of-the-art classification results on the MNIST dataset and relatively good classification results on the CIFAR10 dataset. Since then, researchers have begun extending the idea of capsule networks to other applications; nonetheless, no work yet exists in literature for a method of capsule-based object segmentation.
Performing object segmentation with a capsule-based network is difficult for a number of reasons. The original capsule network architecture and dynamic routing algorithm is extremely computationally expensive, both in terms of memory and run-time. Additional intermediate representations are needed to store the output of “child” capsules in a given layer while the dynamic routing algorithm determines the coefficients by which these children are routed to the “parent” capsules in the next layer. This dynamic routing takes place between every parent and every possible child. One can think of the additional memory space required as a multiplicative increase of the batch size at a given layer by the number of capsule types at that layer. The number of parameters required quickly swells beyond control as well, even for trivially small inputs such as MNIST and CIFAR10. For example, given a set of 32 capsule types with 6×6, 8D-capsules per type, being routed to 10×1, 16D-capsules, the number of parameters for this layer alone is 10×(6×6×32)×16×8=1,474,560 parameters. This one layer contains, coincidentally, roughly the same number of parameters as our entire proposed deep convolutional-deconvolutional capsule network with locally-constrained dynamic routing which itself operates on 512×512 pixel inputs.
We solve this memory burden and parameter explosion by extending the idea of convolutional capsules (primary capsules in Sabour et al. ([2017](#bib.bib20)) are technically convolutional capsules without any routing) and rewriting the dynamic routing algorithm in two key ways. First, children are only routed to parents within a defined spatially-local kernel. Second, transformation matrices are shared for each member of the grid within a capsule type but are not shared across capsule types. To compensate for the loss of global connectivity with the locally-constrained routing, we extend capsule networks by proposing “deconvolutional” capsules which operates using transposed convolutions, routed by the proposed locally-constrained routing. These innovations allow us to still learn a diverse set of different capsule types. Also, with the proposed deep convolutional-deconvolutional architecture, we retain near-global contextual information, while dramatically reducing the number of parameters in the network, addressing the memory burden, and producing state-of-the-art results for our given application. Our proposed SegCaps architecture is illustrated in Figure [2](#S3.F2 "Figure 2 ‣ 3 Building Blocks of Capsules for Image Segmentation ‣ Capsules for Object Segmentation"). As a comparative baseline, we also implement a simple three-layer capsule structure, more closely following that of the original capsule implementation, shown in Figure [1](#S3.F1 "Figure 1 ‣ 3 Building Blocks of Capsules for Image Segmentation ‣ Capsules for Object Segmentation").
Figure 2: The proposed SegCaps architecture for object segmentation.
###
3.1 Summary of Our Contributions
The novelty of this paper can be summarized as follows:
1. Our proposed SegCaps is the first use of a capsule network architecture for object segmentation in literature.
2. We propose two modifications to the original dynamic routing algorithm where (i) children are only routed to parents within a defined spatially-local window and (ii) transformation matrices are shared for each member of the grid within a capsule type.
3. These modifications, combined with convolutional capsules, allow us to operate on large images sizes (512×512 pixels) for the first time in literature, where previous capsule architectures do not exceed inputs of 32×32 pixels in size.
4. We introduce the concept of "deconvolutional" capsules and create a novel deep convolutional-deconvolutional capsule architecture, far deeper than the original three-layer capsule network, implement a three-layer convolutional capsule network baseline using our locally-constrained routing to provide a comparison with our SegCaps architecture, investigate two different routing iteration schemes for our SegCaps, and extend the masked reconstruction of the target class as a method for regularization to the problem of segmentation as described in Section [4](#S4 "4 SegCaps: Capsules for Object Segmentation ‣ Capsules for Object Segmentation").
5. SegCaps produces slightly improved results for lung segmentation on the LUNA16 subset of the LIDC-IDRI database, in terms of dice coefficient, when compared with state-of-the-art methods U-Net (Ronneberger et al. ([2015](#bib.bib19))) and Tiramisu (Jégou et al. ([2017](#bib.bib8))), while dramatically reducing the number of parameters needed to achieve this performance. The proposed SegCaps architecture contains 95.4% fewer parameters than U-Net and 38.4% fewer than Tiramisu.
4 SegCaps: Capsules for Object Segmentation
--------------------------------------------
As illustrated in Figure 2, the input to our SegCaps network is a 512×512 pixel image, in this case, a slice of a CT Scan. This image is passed through a 2D convolutional layer which produces 16 feature maps of the same spatial dimensions. This output forms our first set of capsules, where we have a single capsule type with a grid of 512×512 capsules, each of which is a 16 dimensional vector. This is then followed by our first convolutional capsule layer. We will now generalize this process to any given layer ℓ in the network.
At layer ℓ, ∃ a set of capsule types Tℓ={tℓ1,tℓ2,...,tℓn∣n∈N}. For every tℓi∈Tℓ,∃ an hℓ×wℓ grid of zℓ-dimensional child capsules, C={c11,...,c1wℓ,...,chℓ1,...,chℓwℓ}, where hℓ×wℓ is the spatial dimensions of the output of layer ℓ−1. At the next layer of the network, ℓ+1, ∃ a set of capsule types Tℓ+1={tℓ+11,tℓ+12,...,tℓ+1m∣m∈N}. And for every tℓ+1j∈Tℓ+1,∃ an hℓ+1×wℓ+1 grid of zℓ+1-dimensional parent capsules, P={p11,...,p1wℓ+1,...,phℓ+11,...,phℓ+1wℓ+1}, where hℓ+1×wℓ+1 is the spatial dimensions of the output of layer ℓ.
In convolutional capsules, every parent capsule pxy∈P receives a set of “prediction vectors”, {^uxy∣tℓ1,^uxy∣tℓ2,...,^uxy∣tℓn}, one for each capsule type in Tℓ. Hence, this set is defined as the matrix multiplication between a learned transformation matrix, Mtℓi, and the sub-grid of child capsules outputs, Uxy∣tℓi, within a user-defined kernel centered at position (x,y) in layer ℓ; hence ^uxy∣tℓi=Mtℓi×Uxy∣tℓi,∀ tℓi∈Tℓ. Therefore, we can see each Uxy∣tℓi has shape kh×kw×zℓ, where kh×kw are the dimensions of the user-defined kernel. Each Mtℓi has shape kh×kw×zℓ×∣Tℓ+1∣×zℓ+1 for all capsule types Tℓ, where ∣Tℓ+1∣ is the number of parent capsule types in layer ℓ+1. Note that each Mtℓi does not depend on the spatial location (x,y), as the same transformation matrix is shared across all spatial locations within a given capsule type (similar to how convolutional kernels scan an input feature map), and this is one way our method can exploit parameter sharing to dramatically cut down on the total number of parameters to be learned. The values of these transformation matrices for each capsule type in a layer are learned via the backpropagation algorithm with a supervised loss function.
To determine the final input to each parent capsule pxy∈P, we compute the weighted sum over these “prediction vectors”, pxy=∑nrtℓi∣xy^uxy∣tℓi, where rtℓi∣xy are the routing coefficients determined by the dynamic routing algorithm. These routing coefficients are computed by a “routing softmax”,
| | | | |
| --- | --- | --- | --- |
| | rtℓi∣xy=exp(btℓi∣xy)∑kexp(btℓik), | | (1) |
whose initial logits, btℓi∣xy are the log prior probabilities that prediction vector ^uxy∣tℓi should be routed to parent capsule pxy.
Our method differs from the dynamic routing implemented by Sabour et al. ([2017](#bib.bib20)) in two ways. First, we locally constrain the creation of the prediction vectors. Second, we only route the child capsules within the user-defined kernel to the parent, rather than routing every single child capsule to every single parent. The output capsule is then computed using a non-linear squashing function
| | | | |
| --- | --- | --- | --- |
| | vxy=||pxy||21+||pxy||2pxy||pxy||, | | (2) |
where vxy is the vector output of the capsule at spatial location (x,y) and pxy is its final input. Lastly, the agreement is measured as the scalar product atℓi∣xy=vxy⋅^uxy∣tℓi. A final segmentation mask is created by computing the length of the capsule vectors in the final layer and assigning the positive class to those whose magnitude is above a threshold, and the negative class otherwise. The pseudocode for this locally constrained dynamic routing is summarized in Algorithm [1](#alg1 "Algorithm 1 ‣ 4 SegCaps: Capsules for Object Segmentation ‣ Capsules for Object Segmentation").
1:procedure Routing(^uxy∣tℓi, d, ℓ, kh, kw)
2: for all capsule types tℓi within a kh×kw kernel centered at position (x,y) in layer ℓ and capsule xy centered at position (x,y) in layer (ℓ+1): btℓi∣xy←0.
3: for d iterations do
4: for all capsule types tℓi in layer ℓ: ctℓi←softmax(btℓi) ▹ softmax computes Eq. [1](#S4.E1 "(1) ‣ 4 SegCaps: Capsules for Object Segmentation ‣ Capsules for Object Segmentation")
5: for all capsules xy in layer (ℓ+1): pxy←∑nrtℓi∣xy^uxy∣tℓi
6: for all capsules xy in layer (ℓ+1): vxy←squash(pxy) ▹ squash computes Eq. [2](#S4.E2 "(2) ‣ 4 SegCaps: Capsules for Object Segmentation ‣ Capsules for Object Segmentation")
7: for all capsule types tℓi in layer ℓ and capsules xy in layer (ℓ+1): btℓi∣xy←btℓi∣xy+^uxy∣tℓi.vxy
return vxy
Algorithm 1 Locally-Constrained Dynamic Routing.
As a method of regularization, we extend the idea of reconstructing the input to promote a better embedding of our input space. This forces the network to not only retain all necessary information about a given input, but also encourages the network to better represent the full distribution of the input space, rather than focusing only on its most prominent modes. Since we only wish to model the distribution of the positive input class and treat all other pixels as background, we mask out segmentation capsules which do not belong to the positive class and reconstruct a similarly masked version of the input image. We perform this reconstruction via a three layer 1×1 convolutional network, then compute a weighted mean-squared error loss between only the positive input pixels and this reconstruction.
5 Experiments and Results
--------------------------
Experiments were conducted on the LUNA16 subset of the LIDC-IDRI database, randomly split into four training/testing folds for performing k-fold cross-validation. The LUNA16 subset contains a range of lung CT scans from severe to no pathologies present. Ground-truth annotations were provided in the form of segmentation masks created by an automated algorithm (van Rikxoort et al. ([2009](#bib.bib22))). Manual inspection led to the removal of 10 of the 888 CT scans due to exceedingly poor annotations. Because of the lack of expert human-annotations, we observed that the proposed methods and baselines usually outperformed these ground-truth segmentation masks for particularly difficult scans. This, in turn, lead to higher dice scores for worse performance in those cases, as they typically failed in a similar way. To compensate for such outliers, all numeric results are reported in terms of median rather than mean averages.
U-Net, Tiramisu, our three-layer baseline capsule segmentation network, and SegCaps are all implemented using Keras with TensorFlow. For the baseline capsule network, we modify the margin loss from Sabour et al. ([2017](#bib.bib20)) to the weighted binary version. All other methods are trained using the weighted BCE loss for the segmentation output. We note that in small-scale experiments, the weighted margin loss seemed to perform comparable to the weighted BCE loss for SegCaps. Yet, more thorough experiments are needed to draw conclusions from this. The reconstruction output loss is computed via the masked MSE as described in Section. [4](#S4 "4 SegCaps: Capsules for Object Segmentation ‣ Capsules for Object Segmentation"). All possible experimental factors are controlled between different networks. All networks are trained from scratch, using the same data augmentation methods (scale, flip, shift, rotate, elastic deformations, and random noise) and Adam optimization (Kingma and Ba ([2014](#bib.bib9))) with an initial learning rate of 0.00001. A batch size of 1 is chosen for all experiments to match the original U-Net implementation. The learning rate is decayed by a factor of 0.05 upon validation loss stagnation for 50,000 iterations and early stopping is performed with a patience of 250,000 iterations based on validation dice scores.
The final quantitative results of these experiments are shown in Table [1](#S5.T1 "Table 1 ‣ 5 Experiments and Results ‣ Capsules for Object Segmentation"). SegCaps slightly outperforms all other compared approaches with an average dice score of 98.479%, while requiring far fewer parameters, a reduction in parameters of over 95% from U-Net and over 38% compared with Tiramisu. For qualitative evaluations, we have shown two different slices from two different CT scan and highlighted the segmentation leakages that U-NET caused in Fig. [3](#S5.F3 "Figure 3 ‣ 5 Experiments and Results ‣ Capsules for Object Segmentation").
Further, we investigate how different capsule vectors in the final segmentation capsule layer are representing different visual attributes. Figure [4](#S5.F4 "Figure 4 ‣ 5 Experiments and Results ‣ Capsules for Object Segmentation") shows the selected 5 visual attributes (each row) out of 16 (dimension of final capsule segmentation vector) across different values of the vectors (each column). We observe that regions with different textural properties (i.e., small and large homogeneous) are progressively captured by the capsule segmentation vectors.
| Method | Parameters | Split-0 (%) | Split-1 (%) | Split-2 (%) | Split-3 (%) | Average (%) |
| --- | --- | --- | --- | --- | --- | --- |
| U-Net | 31.0 M | 98.353 | 98.432 | 98.476 | 98.510 | 98.449 |
| Tiramisu | 2.3 M | 98.394 | 98.358 | 98.543 | 98.339 | 98.410 |
| Baseline Caps | 1.7 M | 82.287 | 79.939 | 95.121 | 83.608 | 83.424 |
| SegCaps (R1) | 1.4 M | 98.471 | 98.444 | 98.401 | 98.362 | 98.419 |
| SegCaps | 1.4 M | 98.499 | 98.523 | 98.455 | 98.474 | 98.479 |
Table 1: Dice Coefficient results on a 4-fold cross-validation split of the LUNA16 dataset. For SegCaps (R1), dynamic routing is only performed on layers which change spatial dimensions. All other layers are routed with equal-weight coupling coefficients.
Figure 3: Qualitative results for U-Net and the proposed SegCaps on two different slices from CT scans. The left column is the results produced by U-Net. The right column is the results produced by the proposed SegCaps algorithm on these same slices. The green arrows highlight areas where U-Net made errors in segmentation.
Figure 4: Each row includes 5 selected visual attributes from 16 dimensional capsule segmentation vectors. Each column, on the other hand, indicate different visual attributes of these vectors for a given interval obtained from the capsule segmentation vector.
6 Conclusion
-------------
We propose a novel deep learning algorithm, called SegCaps, for object segmentation, and showed its efficacy in a challenging problem of pathological lung segmentation from CT scans. The proposed framework is the first use of the recently introduced capsule network architecture and expands it in several significant ways. First, we modify the original dynamic routing algorithm to act locally when routing children capsules to parent capsules and to share transformation matrices across capsules within the same capsule type. These changes dramatically reduce the memory and parameter burden of the original capsule implementation and allows for operating on large image sizes, whereas previous capsule networks were restricted to very small inputs. To compensate for the loss of global information, we introduce the concept of a deep convolutional-deconvolutional capsule architecture for pixel level predictions of object labels. Finally, we extend the masked reconstruction of the target class as a regularization strategy for the segmentation problem. Experimentally, SegCaps produces slightly improved accuracies for lung segmentation on the LUNA16 subset of the LIDC-IDRI database, in terms of dice coefficient, when compared with state-of-the-art networks U-Net (Ronneberger et al. ([2015](#bib.bib19))) and Tiramisu (Jégou et al. ([2017](#bib.bib8))). More importantly, the proposed SegCaps architecture contains 95.4% fewer parameters than U-Net and 38.4% fewer than Tiramisu. The proposed algorithm fundamentally improves the current state-of-the-art object segmentation approaches, and provides strong evidence that capsules can successfully model the spatial relationships of the objects better than traditional CNNs. |
a1f8c706-fd88-4677-b6d2-22814efe1890 | trentmkelly/LessWrong-43k | LessWrong | [AN #141]: The case for practicing alignment work on GPT-3 and other large models
Alignment Newsletter is a weekly publication with recent content relevant to AI alignment around the world. Find all Alignment Newsletter resources here. In particular, you can look through this spreadsheet of all summaries that have ever been in the newsletter.
Audio version here (may not be up yet).
Please note that while I work at DeepMind, this newsletter represents my personal views and not those of my employer.
HIGHLIGHTS
The case for aligning narrowly superhuman models (Ajeya Cotra) (summarized by Rohin): One argument against work on AI safety is that it is hard to do good work without feedback loops. So how could we get feedback loops? The most obvious approach is to actually try to align strong models right now, in order to get practice with aligning models in the future. This post fleshes out what such an approach might look like. Note that I will not be covering all of the points mentioned in the post; if you find yourself skeptical, you may want to read the full post as your question might be answered there.
The author specifically suggests that we work on aligning narrowly superhuman models to make them more useful. Aligning a model roughly means harnessing the full capabilities of the model and orienting these full capabilities towards helping humans. For example, GPT-3 presumably “knows” a lot about medicine and health. How can we get GPT-3 to apply this knowledge as best as possible to be maximally useful in answering user questions about health?
Narrowly superhuman means that the model has more knowledge or “latent capability” than either its overseers or its users. In the example above, GPT-3 almost certainly has more medical knowledge than laypeople, so it is at least narrowly superhuman at “giving medical advice” relative to laypeople. (It might even be so relative to doctors, given how broad its knowledge is.)
Learning to Summarize with Human Feedback (AN #116) is a good example of what this could look like: that paper attempted to “bring |
81520554-1496-453c-bddf-8dd0cf1b12ea | awestover/filtering-for-misalignment | Redwood Research: Alek's Filtering Results | id: post913
One of the biggest reasons alignment might be hard is what I’ll call threat obfuscation: various dynamics that might make it hard to measure/notice cases where an AI system has problematic misalignment (even when the AI system is in a controlled environment and one is looking for signs of misalignment). I write more about this general fear in AI Safety Seems Hard to Measure . I think it’s very important to understand threat obfuscation dynamics; to try to design ways of assessing danger that are designed with threat obfuscation dynamics in mind; and to avoid interpreting reassuring-seeming evidence too optimistically, by keeping in mind the various threat obfuscation dynamics that could affect it. I’ve seen three noticeably different threat obfuscation dynamics discussed, in ways that often leave me confused about exactly what dynamic someone is talking about, and when. This post aims to clear that up, based on a conversation with Paul Christiano. The three dynamics are: Simple “training game” dynamics: an AI system appears to behave helpfully, rather than attempting deception/manipulation/disempowerment of humans, because (a) it is good at predicting whether such attempts will be caught; (b) it is aiming for good in-episode performance (in terms of loss or reward); (c) such attempts would result in poor in-episode performance if caught. But if an AI system had a chance to do deception/manipulation/disempowerment that would not be detected, and would result in “good performance” within an episode, it would do so (this could include large-scale world takeover). These dynamics are dangerous because we might observe dangerous behavior, train AIs to behave less dangerously, then observe less dangerous behavior, and conclude that we’ve become more safe (when in fact we’ve trained the AIs to be more effective at deception). But we can still detect dangerous behavior in cases where it would not result in poor in-episode performance. For example, we might decide that when an AI system in the wild hacks its way to good in-episode performance, and is caught after the fact, we will at that point act as if the performance was actually good/intended (for example, no gradient descent against the behavior). Or we might deliberately conduct tests in which AI systems have opportunities to take over digital environments, obtaining high reward (no catch, not a trap, no attempt to "train out" the behavior when it happens). 1 In these cases, situationally aware AI systems might “show their true colors” with fairly straightforward attempts to draw them out. Deceptive alignment : an AI system forms some aim (via inner misalignment) other than good in-episode performance. (This could include something like “an AI system that wants to maximize digital representations of reward across the galaxy, and is willing to get poor in-episode performance as part of a longer-term mission for that to happen.”) It then gets good in-episode performance as part of a longer-term scheme to accomplish its other aim. In this case, the above techniques for getting an AI to “show its true colors” might 2 fail. However, we could still get an AI to engage in specific behaviors that might be helpful to us and harmful to its goals, by specifically training the AI to behave in these ways. For example, we might want an AI to find security vulnerabilities that we can then patch. Even if doing so would undermine the AI’s longer-term aim, we can use gradient descent to put a lot of pressure on it to find security vulnerabilities anyway. (Related: the “no sandbagging on checkable tasks” hypothesis ) Gradient hacking : this also requires that an AI system has some aim other than good in-episode performance. Gradient hacking is a way for an AI system to specifically “defend itself” against being trained to do things counter to its goals. It might look something like: “An AI system checks its own policy against some reference policy that is good for its goals; the greater the divergence, the more it sabotages its own performance, with the result that gradient descent has trouble getting its policy to diverge from the reference policy.” Each of these seems more devastating to alignment research than the previous one - and less likely to emerge in early transformative AI systems. I consider the “training game” very likely, and quite devastating; but it leaves us many possible interventions that the less likely “deceptive alignment” and “gradient hacking” don’t. That was most of what I wanted to cover, but the rest of this post will give some quick notes on (a) why the more severe forms of threat obfuscation seem relatively less likely; (b) how bad things look if we only have to deal with the first one (“training game”). Training game seems very likely by default, deceptive alignment much less so, gradient hacking still less so This is pretty superficial and mostly based on a couple of conversations with Paul Christiano. I’m hoping he’ll write a better version of this at some point. The “training game” mode of threat obfuscation seems like it is very likely by default, because a normal AI training process will probably directly incentivize it. Errors in labeling will directly favor the training game over intended generalization; even with perfect labeling, there’s no particular reason to think that the training game is disfavored compared to the intended generalization (if anything it vaguely/intuitively seems simpler - “do that which gets treated as intended behavior” as opposed to “do that which was intended.”) By contrast, deceptive alignment seems to face a bigger hurdle. For example, if you train an AI with positive reinforcement for getting cookies: A deceptively aligned AI has to, every time it’s deciding how to get cookies, go through a “thought process” like: “I am aiming for [thing other than cookies]. But humans want me to get cookies. And humans are in control right now [if not I need to behave differently]. Therefore I should get cookies.” Contrast this with an AI that just responds to rewards for cookies by going through this thought process: “I’m trying to get cookies.” The former “thought process” could be noticeably more expensive (compute wise or whatever), in which case heavy optimization would push against it. (I think this is plausible, though I’m less than convinced; the former thought process doesn’t seem like it is necessarily much more expensive conditional on already having a situationally aware agent that thinks about the big picture a lot.) An additional issue is that deceptive alignment only happens if you get inner misalignment resulting in an AI with some nonindexical “aim” other than in-episode reward. This could happen but it’s another conjunct. Paul thinks it’s <50% likely (clarification in footnote 3 ) that deceptive alignment is going to be a major issue by default (at least if people make reasonable efforts to avoid the sorts of dynamics discussed in the next section ), and other folks whose judgment I’d put a lot of weight on (due to seeming fluent with the concern about deceptive alignment, while also experienced and/or accomplished with ML) seem to generally be in the same camp. Gradient hacking intuitively seems harder still. The preconditions for it seem to be something like “the preconditions for deceptive alignment, plus the AI figuring out some effective maneuver to execute with the design of its own brain.” How bad is the situation if straightforward training leads to the training game, but not to deceptive alignment or gradient hacking? At first glance, this situation doesn’t seem too bad in terms of threat obfuscation. It basically means that measuring an AI’s dangerousness is as simple as “Put it in an environment where doing malicious things is good in-episode performance. A possible example: if it does malicious things to gain control of its environment, simply let it control its environment for a while and don’t do any of the things one would normally do for unintended behavior (such as trying to use gradient descent to correct against it).” This could result in a dangerous, situationally aware AI (e.g., one that has learned a pattern "Do malicious things iff they won't be caught and trained against") going ahead and doing malicious things while being tested, with no particular reason to conceal them. However, it’s pretty easy for this to go off the rails if people aren’t careful about it. I’d be concerned about a dynamic like this: AIs frequently have opportunities to behave maliciously and gain resources within an episode. These opportunities look good on their reward models. They take these opportunities. Humans respond by trying to train out this sort of behavior - e.g., finding incidents like this via audits and news stories, and going back and doing major gradient updates against the corresponding behaviors/reward models, long after the fact. In so doing, they could effectively be training AIs to build longer-term, more careful models of human reactions to their behavior. For a situationally aware AI, training like this could lead to behavior like “Try to accumulate resources and do things associated with high reward, unless humans might catch it later via an audit; hence, look for opportunities to avoid or prevent or manipulate eventual audits.” This sort of behavior could end up looking a lot like deceptive alignment: optimizing heavily around factors and events well outside the “episode” as traditionally defined (you could also think of this as effectively extending the “episode” to encompass a very long time period during which humans might do audits and respond to what they see). To avoid this, we need to be careful with how we handle unexpected incidents of AIs’ behaving maliciously. It won’t always be the right move to straightforwardly “train against” such incidents; it may sometimes be important to do something more like “starting over,” e.g. insisting on a more principled way 4 of making AIs less prone to malicious behavior. This situation seems scary: We won’t necessarily be able to come up with a principled (and competitive) way of driving out malicious behavior. At least some AI developers (and maybe many) will be doing non-principled things (e.g., the problematic “train out the malicious behavior” thing above); saying “Behold, our systems are safe!”; and wanting to deploy them. The line between principled and non-principled things doesn’t feel super clear. But this situation still seems much better than having to deal with deceptive alignment: we do have a tangible, non-interpretability-dependent way of detecting danger if we can be disciplined about it! In particular, a successful “evals” project looks very tractable in this world - you’re basically running the “make rewards available for bad behavior” test, and then auditing training for signs that it’s been explicitly trained against. I'm not saying this situation is the only one we should plan around. But it is an example of how relatively good luck could put us in a position where some theoretically very difficult aspects of alignment end up being less difficult in reality. Notes The “catch” would be that we would want to avoid deploying similar AI systems in other contexts. But this is not a concern for the “simple training game” AI that is focused on obtaining good in-episode performance. It is a concern for an AI that has a goal other than good in-episode performance. ↩ Though some versions might work - see "Providing unusual incentives ..." here . ↩ “Without thinking carefully I’d guess there is a 25% chance that deceptive alignment is a major challenge that arises early enough that we need to deal with it ourselves. Then another 25% that it’s a challenge that someone needs to deal with some day, but is too far in the future to be worth thinking about now e.g. because AI systems will have obsoleted all of our thinking before they are deceptively aligned. And another 25% chance that deceptive alignment just never arises in the real world if you are careful and take the kind of simple precautions that we already understand. And a final 25% where the entire concern was misguided and it never happens in practice with modern ML no matter how careless you are.” ↩ Such as via internals-based training . ↩ |
e4ee341f-3bff-4f76-b216-ab0c7994ae3b | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Views on when AGI comes and on strategy to reduce existential risk
Summary: AGI isn't super likely to come super soon. People should be working on stuff that saves humanity in worlds where AGI comes in 20 or 50 years, in addition to stuff that saves humanity in worlds where AGI comes in the next 10 years.
Thanks to Alexander Gietelink Oldenziel, Abram Demski, Daniel Kokotajlo, Cleo Nardo, Alex Zhu, and Sam Eisenstat for related conversations.
My views on when AGI comes
==========================
AGI
---
By "AGI" I mean the thing that has very large effects on the world (e.g., it kills everyone) via the same sort of route that humanity has large effects on the world. The route is where you figure out how to figure stuff out, and you figure a lot of stuff out using your figure-outers, and then the stuff you figured out says how to make powerful artifacts that move many atoms into very specific arrangements.
This isn't the only thing to worry about. There could be transformative AI that isn't AGI in this sense. E.g. a fairly-narrow AI that just searches configurations of atoms and finds ways to do atomically precise manufacturing would also be an existential threat and a possibility for an existential win.
Conceptual capabilities progress
--------------------------------
The "conceptual AGI" view:
>
> The first way humanity makes AGI is by combining some set of significant ideas about intelligence. Significant ideas are things like (the ideas of) gradient descent, recombination, probability distributions, universal computation, search, world-optimization. Significant ideas are to a significant extent bottlenecked on great natural philosophers doing great natural philosophy about intelligence, with sequential bottlenecks between many insights.
>
>
>
The conceptual AGI doesn't claim that humanity doesn't already have enough ideas to make AGI. I claim that——though not super strongly.
Timelines
---------
Giving probabilities here doesn't feel great. For one thing, it seems to contribute to information cascades and to shallow coalition-forming. For another, it hides the useful models. For yet another thing: A probability bundles together a bunch of stuff I have models about, with a bunch of stuff I don't have models about. For example, how many people will be doing original AGI-relevant research in 15 years? I have no idea, and it seems like largely a social question. The answer to that question does affect when AGI comes, though, so a probability about when AGI comes would have to depend on that answer.
But ok. Here's some butt-numbers:
* 3%-10% probability of AGI in the next 10-15ish years. This would be lower, but I'm putting a bit of model uncertainty here.
* 40%-45% probability of AGI in the subsequent 45ish years. This is denser than the above because, eyeballing the current state of the art, it seems like we currently lack some ideas we'd need——but I don't know how many insights would be needed, so the remainder could be only a couple decades around the corner. It also seems like people are distracted now.
* Median 2075ish. IDK. This would be further out if an AI winter seemed more likely, but LLMs seem like they should already be able to make a lot of money.
* A long tail. It's long because of stuff like civilizational collapse, and because AGI might be really really hard to make. There's also a sliver of a possibility of coordinating for a long time to not make AGI.
If I were trying to make a model with parts, I might try starting with a mixture of Erlang distributions of different shapes, and then stretching that according to some distribution about the number of people doing original AI research over time.
Again, this is all butt-numbers. I have almost no idea about how much more understanding is needed to make AGI, except that it doesn't seem like we're there yet.
Responses to some arguments for AGI soon
========================================
The "inputs" argument
---------------------
At about [1:15 in this interview](https://www.youtube.com/watch?v=_kRg-ZP1vQc&t=4450s), Carl Shulman argues (quoting from the [transcript](https://www.dwarkeshpatel.com/p/carl-shulman)):
>
> We've been scaling [compute expended on ML] up four times as fast as was the case for most of the history of AI. We're running through the orders of magnitude of possible resource inputs you could need for AI much much more quickly than we were for most of the history of AI. That's why this is a period with a very elevated chance of AI per year because we're moving through so much of the space of inputs per year [...].
>
>
>
This isn't the complete argument Shulman gives, but on its own it's interesting. On its own, it's valid, but only if we're actually scaling up all the needed inputs.
On the conceptual AGI view, this isn't the case, because we aren't very greatly increasing the amount of great natural philosophers doing great natural philosophy about intelligence. That's a necessary input, and it's only being somewhat scaled up. For one thing, many new AI researchers are correlated with each other, and many are focused on scaling up, applying, and varying existing ideas. For another thing, sequential progress can barely be sped up with more bodies.
The "big evolution" argument
----------------------------
Carl goes on to argue that eventually, when we have enough compute, we'll be able to run a really big evolutionary process that finds AGIs (if we haven't already made AGI). This idea also appears in [Ajeya Cotra's report](https://www.lesswrong.com/posts/KrJfoZzpSDpnrv9va/draft-report-on-ai-timelines) on the compute needed to create AGI.
I broadly agree with this. But I have two reasons that this argument doesn't make AGI seem very likely very soon.
The first reason is that running a big evolution actually seems kind of hard; it seems to take significant conceptual progress and massive engineering effort to make the big evolution work. What I'd expect to see when this is tried, is basically nothing; life doesn't get started, nothing interesting happens, the entities don't get far (beyond whatever primitives were built in). You can get around this by invoking more compute, e.g. by simulating physics more accurately at a more detailed level, or by doing hyperparameter search to find worlds that lead to cool stuff. But then you're invoking more compute. (I'd also expect a lot of the hacks that supposedly make our version of evolution much more efficient than real evolution, to actually result in our version being circumscribed, i.e. it peters out because the shortcut that saved compute also cut off some important dimensions of search.)
The second reason is that evolution seems to take a lot of serial time. There's probably lots of clever things one can do to shortcut this, but these would be significant conceptual progress.
"I see how to do it"
--------------------
My (limited / filtered) experience with these ideas leads me to think that [ideas knowably sufficient to make an AGI in practice] aren't widespread or obvious. (Obviously it is *somehow* feasible to make an AGI, because evolution did it.)
The "no blockers" intuition
---------------------------
An intuition that I often encounter is something like this:
>
> Previously, there were blockers to current systems being developed into AGI. But now those blockers have been solved, so AGI could happen any time now.
>
>
>
This sounds to my ears like: "I saw how to make AGI, but my design required X. Then someone made X, so now I have a design for an AGI that will work.". But I don't think that's what they think. I think they don't think they have to have a design for an AGI in order to make an AGI.
I kind of agree with some version of this——there's a lot of stuff you don't have to understand, in order to make something that can do some task. We observe this in modern ML. But current systems, though they impressively saturate some lower-dimensional submanifold of capability-space, don't permeate a full-dimensional submanifold. Intelligence is a positive thing. Most computer code doesn't put itself on an unbounded trajectory of gaining capabilities. To make it work you have to do engineering and science, at some level. Bridges don't hold weight just because there's nothing blocking them from holding weight.
Daniel Kokotajlo points out that for things that grow, it's kind of true that they'll succeed as long as there aren't blockers——and for example animal husbandry kind of just works, without the breeders understanding much of anything about the internals of why their selection pressures are met with adequate options to select. This is true, but it doesn't seem very relevant to AGI because we're not selecting from an existing pool of highly optimized "genomic" (that is, mental) content. If instead of tinkering with de novo gradient-searched circuits, we were tinkering with remixing and mutating whole-brain emulations, then I would think AGI comes substantially sooner.
Another regime where "things just work" is many mental contexts where a task is familiar enough in some way that you can expect to succeed at the task by default. For example, if you're designing a wadget, and you've previously designed similar wadgets to similar specifications, then it makes sense to treat a design idea as though it's going to work out——as though it can be fully fleshed out into a satisfactory, functioning design——*unless* you see something clearly wrong with it, a clear blocker like a demand for a metal with unphysical properties. Again, like the case of animal husbandry, the "things just work" comes from the (perhaps out of sight) preexisting store of optimized content that's competent to succeed at the task given a bit of selection and arrangement. In the case of AGI, no one's ever built anything like that, so the store of knowledge that would automatically flesh out blockerless AGI ideas is just not there.
Yet another such regime is markets, where the crowd of many agents can be expected to figure out how to do something as long as it's feasible. So, a version of this intuition goes:
>
> There are a lot of people trying to make AGI. So either there's some strong blocker that makes it so that no one can make AGI, or else someone will make AGI.
>
>
>
This is kind of true, but it just goes back to the question of how much conceptual progress will people make towards AGI. It's not an argument that we already have the understanding needed to make AGI. If it's used as an argument that we already have the understanding, then it's an accounting mistake: it says "We already have the understanding. The reason we don't need more understanding, is that if there were more understanding needed, someone else will figure it out, and then we'll have it. Therefore no one needs to figure anything else out.".
Finally: I also see a fair number of specific "blockers", as well as some indications that existing things don't have properties that would scare me.
"We just need X" intuitions
---------------------------
Another intuition that I often encounter is something like this:
>
> We just need X to get AGI. Once we have X, in combination with Y it will go all the way.
>
>
>
Some examples of Xs: memory, self-play, continual learning, curricula, AIs doing AI research, learning to learn, neural nets modifying their own weights, sparsity, learning with long time horizons.
For example: "Today's algorithms can learn anything given enough data. So far, data is limited, and we're using up what's available. But self-play generates infinite data, so our systems will be able to learn unboundedly. So we'll get AGI soon.".
This intuition is similar to the "no blockers" intuition, and my main response is the same: the reason bridges stand isn't that you don't see a blocker to them standing. See above.
A "we just need X" intuition can become a "no blockers" intuition if someone puts out an AI research paper that works out some version of X. That leads to another response: just because an idea is, at a high level, some kind of X, doesn't mean the idea is anything like the fully-fledged, generally applicable version of X that one imagines when describing X.
For example, suppose that X is "self-play". One important thing about self-play is that it's an infinite source of data, provided in a sort of curriculum of increasing difficulty and complexity. Since we have the idea of self-play, and we have some examples of self-play that are successful (e.g. AlphaZero), aren't we most of the way to having the full power of self-play? And isn't the full power of self-play quite powerful, since it's how evolution made AGI? I would say "doubtful". The self-play that evolution uses (and the self-play that human children use) is much richer, containing more structural ideas, than the idea of having an agent play a game against a copy of itself.
Most instances of a category are not the most powerful, most general instances of that category. So just because we have, or will soon have, some useful instances of a category, doesn't strongly imply that we can or will soon be able to harness most of the power of stuff in that category. I'm reminded of [the politician's syllogism](https://en.wikipedia.org/wiki/Politician%27s_syllogism): "We must do something. This is something. Therefore, we must do this.".
The bitter lesson and the success of scaling
--------------------------------------------
[Sutton's bitter lesson](http://www.incompleteideas.net/IncIdeas/BitterLesson.html), paraphrased:
>
> AI researchers used to focus on coming up with complicated ideas for AI algorithms. They weren't very successful. Then we learned that what's successful is to leverage computation via general methods, as in deep learning and massive tree search.
>
>
>
Some add on:
>
> And therefore what matters in AI is computing power, not clever algorithms.
>
>
>
This conclusion doesn't follow. Sutton's bitter lesson is that figuring out *how to leverage* computation using *general methods* that *scale with more computation* beats trying to perform a task by encoding human-learned specific knowledge about the task domain. You still have to come up with the general methods. It's a different sort of problem——trying to aim computing power at a task, rather than trying to work with limited computing power or trying to "do the task yourself"——but it's still a problem. To modify a famous quote: "In some ways we feel we are as bottlenecked on algorithmic ideas as ever, but we believe we are bottlenecked on a higher level and about more important things."
Large language models
---------------------
Some say:
>
> LLMs are already near-human and in many ways super-human general intelligences. There's very little left that they can't do, and they'll keep getting better. So AGI is near.
>
>
>
This is a hairy topic, and my conversations about it have often seemed not very productive. I'll just try to sketch my view:
* The existence of today's LLMs is scary and should somewhat shorten people's expectations about when AGI comes.
* LLMs have fixed, partial concepts with fixed, partial understanding. An LLM's concepts are like human concepts in that they can be combined in new ways and used to make new deductions, in some scope. They are unlike human concepts in that they won't grow or be reforged to fit new contexts. So for example there will be some boundary beyond which a trained LLM will not recognize or be able to use a new analogy; and this boundary is well within what humans can do.
* An LLM's concepts are mostly "in the data". This is pretty vague, but I still think it. A number of people who think that LLMs are basically already AGI have seemed to agree with some version of this, in that when I describe something LLMs can't do, they say "well, it wasn't in the data". Though maybe I misunderstand them.
* When an LLM is trained more, it gains more partial concepts.
* However, it gains more partial concepts with poor sample efficiency; it mostly only gains what's in the data.
* In particular, even if the LLM were being continually trained (in a way that's similar to how LLMs are already trained, with similar architecture), it still wouldn't do the thing humans do with quickly picking up new analogies, quickly creating new concepts, and generally reforging concepts.
* LLMs don't have generators that are nearly as powerful as the generators of human understanding. The stuff in LLMs that seems like it comes in a way that's similar to how stuff in humans comes, actually comes from a lot more data. So LLMs aren't that much of indication that we've figured out how to make things that are on an unbounded trajectory of improvement.
* LLMs have a weird, non-human shaped set of capabilities. They go much further than humans on some submanifold, and they barely touch some of the full manifold of capabilities. (They're "unbalanced" in Cotra's terminology.)
* There is a *broken inference*. When talking to a human, if the human emits certain sentences about (say) category theory, that strongly implies that they have "intuitive physics" about the underlying mathematical objects. They can recognize the presence of the mathematical structure in new contexts, they can modify the idea of the object by adding or subtracting properties and have some sense of what facts hold of the new object, and so on. This inference——emitting certain sentences implies intuitive physics——doesn't work for LLMs.
* The broken inference is broken because these systems are optimized for being able to perform all the tasks that don't take a long time, are clearly scorable, and have lots of data showing performance. There's a bunch of stuff that's really important——and is a key indicator of having underlying generators of understanding——but takes a long time, isn't clearly scorable, and doesn't have a lot of demonstration data. But that stuff is harder to talk about and isn't as intuitively salient as the short, clear, demonstrated stuff.
* Vaguely speaking, I think stable diffusion image generation is comparably impressive to LLMs, but LLMs seem even more impressive to some people because LLMs break the performance -> generator inference more. We're used to the world (and computers) creating intricate images, but not creating intricate texts.
* There is a *missing update*. We see impressive behavior by LLMs. We rightly update that we've invented a surprisingly generally intelligent thing. But we should also update that this behavior surprisingly turns out to not require as much general intelligence as we thought.
Other comments on AGI soon
==========================
* There's a seemingly wide variety of reasons that people I talk to think AGI comes soon. This seems like evidence for each of these hypotheses: that AGI comes soon is overdetermined; that there's one underlying crux (e.g.: algorithmic progress isn't needed to make AGI) that I haven't understood yet; that I talked to a heavily selected group of people (true); that people have some other reason for saying that AGI comes soon, and then rationalize that proposition.
* I'm somewhat concerned that people are being somewhat taken in by hype (experiments systematically misinterpreted by some; the truth takes too long to put on its pants, and the shared narrative is already altered).
* I'm kind of baffled that people are so willing to say that LLMs understand X, for various X. LLMs do not behave with respect to X like a person who understands X, for many X.
* I'm pretty concerned that many people are fairly strongly deferring to others, in a general sense that includes updating off of other people's actions and vibes. Widespread deference has many dangers, which I list in "[Dangers of deference](https://tsvibt.blogspot.com/2022/09/dangers-of-deferrence.html)".
* I'm worried that there's a bucket error where "I think AGI comes soon." isn't separated from "We're going to be motivated to work together to prevent existential risk from AGI.".
My views on strategy
====================
* Alignment is really hard. No one has good reason to think any current ideas would work to make an aligned / corrigible AGI. If AGI comes, everyone dies.
* If AGI comes in five years, everyone dies. We won't solve alignment well enough by then. This of course doesn't imply that AGI coming soon is less likely. However, it does mean that some people should focus on somewhat different things. Most people trying to make the world safe by solving AGI alignment should be open to trains of thought that likely will only be helpful in twenty years. There will be a lot of people who can't help the world if AGI comes in five years; if those people are going to stress out about how they can't help, instead they should work on stuff that helps in twenty or fifty years.
* A consensus belief is often inaccurate, e.g. because of deference and information cascades. In that case, the consensus portfolio of strategies will be incorrect.
* Not only that, but furthermore: Suppose there is a consensus believe, and suppose that it's *totally correct*. If funders, and more generally anyone who can make stuff happen (e.g. builders and thinkers), use this *totally correct* consensus belief to make local decisions about where to allocate resources, *and they don't check the global margin*, then they will in aggregrate follow a portfolio of strategies that is incorrect. The make-stuff-happeners will each make happen the top few things on their list, and leave the rest undone. The top few things will be what the consensus says is most important——in our case, projects that help if AGI comes within 10 years. If a project helps in 30 years, but not 10 years, then it doesn't get any funding at all. This is not the right global portfolio; it oversaturates fast interventions and leaves slow interventions undone.
* Because the shared narrative says AGI comes soon, there's less shared will for projects that take a long time to help. People don't come up with such projects, because they don't expect to get funding; and funders go on not funding such projects, because they don't see good ones, and they don't particularly mind because they think AGI comes soon.
Things that might actually work
-------------------------------
Besides the standard stuff (AGI alignment research, moratoria on capabilities research, explaining why AGI is an existential risk), here are two key interventions:
* Human intelligence enhancement. Important, tractable, and neglected. Note that if alignment is hard enough that we can't solve it in time, but enhanced humans could solve it, then making enhanced humans one year sooner is almost as valuable as making AGI come one year later.
* Confrontation-worthy empathy. Important, probably tractable, and neglected.
+ I suspect there's a type of deep, thorough, precise understanding that one person (the intervener) can have of another person (the intervened), which makes it so that the intervener can confront the intervened with something like "If you and people you know succeed at what you're trying to do, everyone will die.", and the intervened can hear this.
+ This is an extremely high bar. It may go beyond what's normally called empathy, understanding, gentleness, wisdom, trustworthiness, neutrality, justness, relatedness, and so on. It may have to incorporate a lot of different, almost contradictory properties; for example, the intervener might have to at the same time be present and active in the most oppositional way (e.g., saying: I'm here, and when all is said and done you're threatening the lives of everyone I love, and they have a right to exist) while also being almost totally diaphanous (e.g., in fact not interfering with the intervened's own reflective processes). It may involve irreversible changes, e.g. risking innoculation effects and unilateralist commons-burning. It may require incorporating very distinct skills; e.g. being able to make clear, correct, compelling technical arguments, and also being able to hold emotional space in difficult reflections, and also being interesting and socially competent enough to get the appropriate audiences in the first place. It probably requires seeing the intervened's animal, and the intervened's animal's situation, so that the intervener can avoid being a threat to the intervened's animal, and can help the intervened reflect on other threats to their animal. Developing this ability probably requires recursing on developing difficult subskills. It probably requires to some extent thinking like a cultural-rationalist and to some extent thinking very much not like a cultural-rationalist. It is likely to have discontinuous difficulty——easy for some sorts of people, and then very difficult in new ways for other sorts of people.
+ Some people are working on related abilities. E.g. Circlers, authentic relaters, therapists. As far as I know (at least having some substantial experience with Circlers), these groups aren't challenging themselves enough. Mathematicians constantly challenge themselves: when they answer one sort of question, that sort of question becomes less interesting, and they move on to thinking about more difficult questions. In that way, they encounter each fundamental difficulty eventually, and thus have likely already grappled with the mathematical aspect of a fundamental difficulty that another science encounters.
+ Critch talks about empathy [here](https://www.lesswrong.com/posts/gZkYvA6suQJthvj4E/my-may-2023-priorities-for-ai-x-safety-more-empathy-more), though maybe with a different emphasis. |
04bd0c8f-9466-41af-aacc-8f0ae3496f86 | trentmkelly/LessWrong-43k | LessWrong | post
Previously proposed LessWrong projects that have not materialized and could form the basis for such a group:
1. The Simple Math of Everything
2. Social.lesswrong.com/CUA (mea culpa)
3. Threaded sub-reddits
4. Lesswrong University
|
1e50b782-dc0e-4f2f-896a-97ac18a47cf6 | trentmkelly/LessWrong-43k | LessWrong | 2023 Survey Results
The Data
0. Population
There were 558 responses over 32 days. The spacing and timing of the responses had hills and valleys because of an experiment I was performing where I'd get the survey advertised in a different place, then watch how many new responses happened in the day or two after that.
Previous surveys have been run over the last decade or so.
2009: 166
2011: 1090
2012: 1195
2013: 1636
2014: 1503
2016: 3083
2017: "About 300"
2020: 61
2022: 186
2023: 558
Last year when I got a hundred and eighty six responses, I said that the cheerfully optimistic interpretation was "cool! I got about as many as Scott did on his first try!" This time I got around half of what Scott did on his second try. A thousand responses feels pretty firmly achievable.
This is also the tenth such survey that’s been run. We missed a proper ten year anniversary in 2019, and in 2022 I was mostly focused on making the survey happen at all. Still, this is a cool milestone, and in celebration I’m going to be dipping into the datasets from previous years a lot. Unfortunately that doesn’t mean I have ten surveys worth of data; bit rot and the rotation of census runners means I only have access to about half of these.
I’ll talk about other surveys more later on. For the moment, let's talk about the basic breakdowns.
There’s two main formats I’m going to present information in.
One of them is where I'm not treating the answers as numbers. Here, I present
* The option selected
* How many people selected that option
* The percentage of people who selected that option
Let's use Relationship Status as an example.
Relationship Status:
Single: 263, 48.3%
Relationship: 170, 31.3%
Married: 111, 20.4%
263 people said they were single. That's 48.3% of the answers.
The other format is where I have the mean and standard deviation. If you see a sequence of numbers like "30.1 + 8.9 (24, 28, 34) [n=186]" those numbers are "Mean + standard deviation (1st quartile, 2nd quartile, 3rd quarti |
516be071-926f-4840-9b3d-744669111500 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Less Wrong NH Meetup
Discussion article for the meetup : Less Wrong NH Meetup
WHEN: 23 February 2016 07:00:00PM (-0500)
WHERE: 269 Pearl St, Manchester, NH 03104
The thirteenth NH meet-up is Tuesday, 2/23, in Manchester, NH at 7 pm at a private residence. Light refreshments will be provided.
Have you read Rationality: from AI to Zombies, or any of the Sequences on Less Wrong? Maybe you're just a fan of Harry Potter and the Methods of Rationality. Come hang out with us and discuss optimization of whatever it is you want to optimize.
Agenda: Instrumental techniques
You may want to bring a notebook.
https://www.facebook.com/events/501964119975678/
https://www.facebook.com/groups/695201067251306/
Discussion article for the meetup : Less Wrong NH Meetup |
75004c3d-dfd1-4e74-90ef-b8a812af3243 | trentmkelly/LessWrong-43k | LessWrong | Four Phases of AGI
AGI is not discrete, and different phases lead to different opportunities, risks, and strategies
> “How long until AGI? I’d say -1 years from now.”
> “Considering the worst-case outcomes of AGI, I’m most concerned about the effect on jobs.”
> “AGI will be like corporations, so we will control it like corporations.”
> “AGI will be millions of times smarter than us and take over control with methods we can’t comprehend.”
These kinds of statements are all over AI discussions. While these statements are all about “artificial general intelligence” (AGI)—meaning AI systems that can perform as well or better than humans on a wide range of cognitive tasks—they have wildly different assumptions and implications. You could decide that some of these takes are ridiculous--that one particular framing of AGI is “correct” and the others “wrong.” People often decide this, leading to unproductive fights.
But what may actually be happening is that different people have valid reasons for believing these different takes. They are simply referring to different concepts—AGI is ambiguous. To further complicate matters, AGI is broad: “as well or better than humans” is a wide range with no upper bound, so the term “AGI” is used for AI systems spanning from advanced chatbots to incomprehensibly superhuman intelligence.
This post lays out a particular framework for conceptualizing different levels of AGI. I aim for it to be particularly useful to non-technical decision-makers who may need to deal with policy and governance issues with respect to all the above perspectives.
Four Phases
One way we can think about the progression of AGI is to split it into four separate phases—Below-human AGI (BAGI), Human-level AGI (HAGI), Moderately-superhuman AGI (MAGI), and Superintelligent AGI (SAGI). Figure 1 conceptualizes these phase transitions as general capabilities (y-axis) advances with increasing time and AI investment (x-axis):
Figure 1: Conceptual illustration of the Four Phases of AGI us |
2ea382ba-ee3a-4f74-9a89-914ccf3008f3 | StampyAI/alignment-research-dataset/special_docs | Other | Attention in value-based choice as optimal sequential sampling.
Fixation patterns in simple choice re
ect
optimal information sampling
Frederick Callaway1*, Antonio Rangel2, and Thomas L. Griths1,3
1Department of Psychology, Princeton University, Princeton, NJ 08544
2Departments of Humanities and Social Sciences and Computation and Neural Systems,
California Institute of Technology, Pasadena, CA 91125
3Department of Computer Science, Princeton University, Princeton, NJ 08544
*fredcallaway@princeton.edu
December 30, 2020
Abstract
Simple choices (e.g., eating an apple vs. an orange) are made by integrating noisy
evidence that is sampled over time and in
uenced by visual attention; as a result,
uctuations in visual attention can a ect choices. But what determines what is xated
and when? To address this question, we model the decision process for simple choice as
an information sampling problem, and approximate the optimal sampling policy. We
nd that it is optimal to sample from options whose value estimates are both high and
uncertain. Furthermore, the optimal policy provides a reasonable account of xations
and choices in binary and trinary simple choice, as well as the di erences between the
two cases. Overall, the results show that the xation process during simple choice is
in
uenced dynamically by the value estimates computed during the decision process,
in a manner consistent with optimal information sampling.
Author summary
Any supermarket shopper is familiar with the problem of choosing between a small number
of items. Even these \simple choices" can be challenging because we have to think about
the options to determine which one we like most, and we can't think about all of them at
once. This raises a question: what should we think about|and for how long should we
think|before making a decision? We formalize this question as an information sampling
problem, and identify an optimal solution. Observing what people look at while making
choices, we nd that many of the key patterns in their eye xations are consistent with
optimal information sampling.
Introduction
Consider the problems faced by a diner at a bu et table or a shopper at a supermarket shelf.
They are presented with a number of options and must evaluate them until they identify
the most desirable one. A central question in psychology and neuroscience is to understand
the algorithms, or computational processes, behind these canonical simple choices.
Previous work has established two important features of the processes underlying simple
value-based choices. First, choices and reaction times are well explained by information
sampling models like the di usion decision model (DDM) [1{3] and the leaky competing
accumulator model [4, 5]. In these models, individuals are initially uncertain about the
desirability of each option, but they receive noisy signals about the options' values that
1
they integrate over time to form more accurate estimates. A central insight of these models
is that sampling information about unknown subjective values is a central feature of simple
choice. Second, visual attention a ects the decision-making process. In particular, items
that are xated longer are more likely to be chosen [6{13], unless they are aversive, in which
case they are chosen lessfrequently [7, 14]. These ndings have been explained by the
Attentional Drift Di usion Model (aDDM), in which the value samples of the xated item
are over-weighted relative to those of un xated ones [9, 10, 12, 13]. See refs 15 and 16 for
reviews.
These insights raise an important question: What determines what is xated and when
during the decision process? Previous work has focused on two broad classes of theories.
One class suggests that decisions and xations are driven by separate processes, so that
xations a ect how information about values is sampled and integrated, but not the other
way around. In this view, although xations can be modulated by features like visual
saliency or spatial location, they are assumed to be independent of the state of the decision
process. This is the framework behind the aDDM [9, 10, 12] and related models [17, 18].
Another class of theories explores the idea that the decision process a ects xations,
especially after some information about the options' values has been accumulated. Examples
of this class include the Gaze Cascade Model [6], an extension of the aDDM in which options
with more accumulated evidence in their favor are more likely to be xated [19], and a
Bayesian sampling model in which options with less certain estimates are more likely to be
xated [20]. However, these models have not considered how uncertainty and value might
interact, nor have they considered the optimality of the posited xation process.
Research on eye movements in the perceptual domain suggests a third possibility: that
xations are deployed to sample information optimally in order to make the best possible
choice. Previous work in vision has shown that xations are guided to locations that pro-
vide useful information for performing a task, and often in ways that are consistent with
optimal sampling [21]. For example, in visual search (e.g., nding an `M' in a eld of `Ns')
people xate on areas most likely to contain the target [22, 23]; in perceptual discrimina-
tion problems, people adapt their relative xation time to the targets' noise levels [24, 25];
and in naturalistic task-free viewing, xations are drawn to areas that have high \Bayesian
surprise", i.e., areas where meaningful information is most likely to be found [26]. The
properties of xations in these types of tasks are captured by optimal sampling models that
maximize expected information gain [21, 27]. However, these models have not been applied
in the context of value-based decision making, and thus the extent to which xations pat-
terns during simple choices are consistent with optimal information sampling is an open
question.
In this paper, we draw these threads together by de ning a model of optimal information
sampling in canonical simple choice tasks and investigating the extent to which it accounts
for xation patterns and their relation to choices. In a value-based choice, optimal informa-
tion sampling requires maximizing the di erence between the value of the chosen item and
the cost of acquiring the information needed to make the choice. Our model thus falls into
a broad class of models that extend classical rational models of economic choice [28, 29] to
additionally account for constraints imposed by limited cognitive resources [30{35]. How-
ever, as is common in this approach, we stop short of specifying a full algorithmic model of
simple choice. Instead, we ask to what extent people's xations are consistent with optimal
information sampling, without specifying how the brain actually implements an optimal
sampling policy.
Exploring an optimal information sampling model of xations in simple choice is use-
ful for several reasons. First, since xations can a ect choices, understanding what drives
the xation process can provide critical insight into the sources of mistakes and biases in
decision-making. In particular, the extent to which behaviors can be characterized as mis-
takes depends on the extent to which xations sample information sub-optimally. Second,
simple choice algorithms like the DDM have been shown to implement optimal Bayesian
information processing when the decision-maker receives the same amount of information
2
about all options at the same rate [36{42], and this is often viewed as an explanation for why
the brain uses these algorithms in the rst place. In contrast, the optimal algorithm when
the decision-maker must sample information sequentially and selectively is unknown. Third,
given the body of evidence showing that xations are deployed optimally in perceptual de-
cision making, it is interesting to ask if the same holds for value-based decisions. Given
that such problems are characterized by both a di erent objective function (maximizing a
scalar value rather than accuracy) and a di erent source of information (e.g. sampling from
memory [43{45] rather than from a noisy visual stimulus), it is far from clear that optimal
information sampling models will still provide a good account of xations in this setting.
Building on the previous literature, our model assumes that the decision maker estimates
the value of each item in the choice set based on a sequence of noisy samples of the items' true
values. We additionally assume that these samples can only be obtained from the attended
item, and that it is costly to take samples and to switch xation locations. This sets up
a sequential decision problem: at each moment the decision maker must decide whether
to keep sampling, and if so, which item to sample from. Since the model does not have a
tractable analytical solution, in order to solve it and take it to the data, we approximate
the optimal solution using tools from metareasoning in arti cial intelligence [46{49].
We compare the optimal xation policy to human xation patterns in two in
uential
binary and trinary choice datasets [9, 10]. We nd that the model captures many previously
identi ed patterns in the xation data, including the e ects of previous xation time [20]
and item value [17, 19]. In addition, the model makes several novel predictions about
the di erences in xations between binary and trinary choices and about xation durations,
which are consistent with the data. Finally, we identify a critical role of the prior distribution
in producing the classic e ects of attention on choice [7, 9, 10, 14]. Overall, the results show
that the xation process during simple choice is in
uenced by the value estimates computed
during the decision process, in a manner consistent with optimal information sampling.
Model
Sequential sampling model
We consider simple choice problems in which a decision maker (DM) is presented with a
set of items (e.g. snacks) and must choose one. Each item iis associated with some true
but unknown value, u(i), the utility that the DM would gain by choosing it. Following
previous work [1, 9, 36, 38{42, 50], we assume that the DM informs her choice by collecting
noisy samples of the items' true values, each providing a small amount of information,
but incurring a small cost. The DM integrates the samples into posterior beliefs about
each item's value, choosing the item with maximal posterior mean when she terminates the
sampling process.
As illustrated in Fig. 1, we model attention by assuming that the DM can only sample
from one item at each time point, the item she is xating on. This sets up a fundamental
problem: How should she allocate xations in order to make good decisions without incurring
too much cost? Speci cally, at each time point, the DM must decide whether to select an
option or continue sampling, and in the latter case, she must also decide which item to
sample from. Importantly, she cannot simply allocate her attention to the item with the
highest true value because she does not know the true values. Rather, she must decide
which item to attend to based on her current value estimates and their uncertainty.
The DM's belief about the item values at time tis described by a set of Gaussians, one
for each item, with means (i)
tand precisions (i)
t(the precision is the inverse of the vari-
ance). These estimated value distributions are initialized to the DM's prior belief about the
distribution of values in the environment. That is, she assumes that u(i)Gaussian(;2)
and consequently sets (i)
0= and(i)
0= 2for alli. We further discuss the important
role of the prior below.
3
Figure 1: Sampling and belief updating in the binary choice task. The top row shows the experimental display, with the
xated item denoted by the eye symbol. The bottom two rows depict the rst few steps of the sampling and belief updating
process. The decision maker's beliefs about the value of each item are denoted by the Gaussian probability density curves.
The true values of each item (dashed lines) are sampled from standard normal distributions; this is captured in the decision
maker's initial belief state ( rst column). Every time step, t, the decision maker xates one of the items and receives a noisy
sample about the true value of that item ( xtmarks). She then updates her belief about the value of the xated item using
Bayesian updating (shift from light to dark curve). The beliefs for the un xated item are not updated. The process repeats
each time step until the decision maker terminates sampling, at which point she chooses the item with maximal posterior
mean.
We model the control of attention as the selection of cognitive operations, ct, that specify
either an item to sample, or the termination of sampling. If the DM wishes to sample from
itemcat time-step t, she selects ct=cand receives a signal
xtGaussian(u(c);2
x); (1)
whereu(c)is the unknown true value of the item being sampled, and 2
xis a free parameter
specifying the amount of noise in each signal. The belief state is then updated in accordance
with Bayesian inference:
(c)
t+1=(c)
t+ 2
x
(c)
t+1= 2
xxt+(c)
t(c)
t
(c)
t+1
(i)
t+1=(i)
tand(i)
t+1=(i)
tfori6=c:(2)
The cognitive cost of each step of sampling and updating is given by a free parameter,
sample . We additionally impose a switching cost,
switch , that the DM incurs whenever she
samples from an item other than the one sampled on the last timestep (i.e., makes a saccade
to a di erent item). Thus, the cost of sampling is
cost(ct) =
sample +1(ct6=ct 1)
switch: (3)
Note that the model includes the special case in which there are no switching costs (
switch =
0).
4
In addition to choosing an item to sample, the DM can also decide to stop sampling and
choose the item with the highest expected value. In this case, she selects ct=?. It follows
that if the choice is made at time step T(i.e.,cT=?) the chosen item is i= arg max i(i)
T.
The DM's total payo on a single decision is given by:
payo = u(i)
|{z}
utility of
chosen item T 1X
t=1cost(ct):
|{z}
cognitive cost(4)
Optimal policy
We assume that the decisions about where to xate and when to stop sampling are made
optimally, subject to the informational constraints described in the previous section. For-
mally, we assume that the ctare selected by an optimal policy . A policy selects the next
cognitive operation to execute, ct, given the current belief state, ( t;t); it is optimal if it
selectsctin a way that maximizes the expectation of Equation 4. How can we identify such
a policy? Problems of this kind have been explored in the arti cial intelligence literature
on rational metareasoning [46, 47]. Thus, we cast the model described above as a metalevel
Markov decision process [48], and identify a near-optimal policy using a recently developed
method that has been been shown to achieve strong performance on a related problem [49].
In accordance with past work modeling people's choices [51] and xations [19, 20], we assume
that people follow a softmax policy in selecting each cognitive operation by sampling from
a Boltzmann distribution based on their estimated values. Thus, their choices of cognitive
operations are guided by the optimal policy, but subject to some noise. See Methods for
details.
What does optimal attention allocation look like? In order to provide an intuitive un-
derstanding, we focus on two key properties of belief states: 1) uncertainty about the true
values and (2) di erences in the value estimates. Fig. 2A shows the probability of the opti-
mal policy (for a model with parameters t to human data) sampling an item as a function
of these two dimensions (marginalizing over the other dimensions according to their prob-
ability of occurring in simulated trials). We see that the optimal policy tends to xate on
items that are uncertain and have estimated values similar to the other items. In the case
of trinary|but not binary|choice, we additionally see a stark asymmetry in the e ect of
relative estimated value. While the policy is likely to sample from an item whose value is
substantially higher than the competitors, it is unlikely to sample from an item with value
well below. In particular, the policy has a strong preference to sample from the items with
best or second-best value estimates.
To see why this is optimal, note that sampling is only valuable insofar as it a ects choice,
and that the chosen item is the one with maximal estimated value when sampling stops.
Thus, the optimal policy generally xates on the item for which gathering more evidence is
most likely to change which item has maximal expected value. There are two ways for this
to happen: either the value of the current best item is reduced below the second-best item,
or the value of some alternative item is increased above the best item. The former can only
happen by sampling the best item, and the latter is ceteris paribus most likely to occur by
sampling the second-best item because it is closer to the top position than the third-best
item (Fig. 2B bottom). However, if uncertainty is much greater for the third-best item, this
can outweigh the larger di erence in estimated value (Fig. 2B top).
The prior distribution
Recall that the initial belief about each item's value is set to the DM's prior belief about the
distribution of values in the environment; that is (i)
0= and(i)
0= 2. This corresponds
to the DM assuming that each item's value is drawn from a prior distribution of true values
5
A
B
Figure 2: Optimal xation policy. (A) Probability of xating on item 1 as a function
of the precision of its value estimate, (1), and the mean of its relative value estimate,
(1) mean((2);(3)). The heat map denotes the probability of xating item 1 as opposed
to xating one of the other items or terminating the sampling process. (B) Illustration of
the value of sampling. Each panel shows a belief state for trinary choice. The curves depict
the estimated beliefs for each item's value, and the shaded regions show the probability
that the item's true value is higher than the current best value estimate. This probability
correlates strongly with the value of sampling the item because sampling is only valuable
if it changes the choice (the full value of sampling additionally depends on the size of the
potential gain in value, as well as the cost of future samples and the possibility of sampling
other items). In each case, it is more valuable to sample the orange item than the purple
item because either (top) its value is more uncertain, or (bottom) its value is closer to the
leading value.
given byu(i)Gaussian(;2). This assumption is plausible if this is the actual distribution
of items that the DM encounters, and she is a Bayesian learner with sucient experience
in the context under study. However, given that these models are typically used to study
choices made in the context of an experiment (as we do here), the DM might not have
learned the exact prior distribution at work. As a result, we must consider the possibility
that she has a biased prior .
In order to investigate the role of the prior on the model predictions, we assume that
6
it takes the form of a Gaussian distribution with a mean and standard deviation related to
the actual empirical distribution as follows:
= mean(ratings)
= std(ratings) :(5)
Here, mean(ratings) denotes the mean value ratings of all items, which provide independent
and unbiased measures of the true value of the items (computed across trials in both ex-
periments), and is a free parameter that speci es the amount of bias in the prior ( = 0
corresponds to a strong bias and = 1 corresponds to no bias). As a result, the DM has
correct beliefs about the prior variance, but is allowed to have a biased belief about the
prior mean. This case could arise, for example, if the average true value of the items used
in the experiment di ers from the average item that the DM encounters in her daily life.
Model tting
We apply the model to two in
uential simple choice datasets: a binary food choice task [9]
and a trinary food choice task [10]. In each study, participants rst provided liking-ratings
for 70 snack items on a -10 to 10 scale, which are used as an independent measure of the
items' true values. They then made 100 choices among items that they had rated positively,
while the location of their xations was monitored at a rate of 50 Hz. See the Supplementary
Information for more details on the experiments.
The model has ve free parameters: the standard deviation of the sampling distribution
x, the cost per sample
sample , the cost of switching attention
switch , the prior bias , and
the inverse temperature of the softmax policy used to select cognitive operations, . This
last parameter controls the amount of noise in the xation decisions. In order to t the
model, we need to make an assumption about the time that it takes to acquire each sample,
which we take to be 100 ms. Note, however, that this choice is not important: changing
the assumed duration leads to a change in the tted parameters, but not in the qualitative
model predictions.
We use an approximate maximum likelihood method to t these parameters to choice
and xation data, which is described in the Methods section. Importantly, since the same
model can be applied to N-item choices, we t a common set of parameters jointly to the
pooled data in both datasets. Thus, any di erences in model predictions between binary
and trinary choices are a priori predictions resulting from the structure of the model, and
not di erences in the parameters used to explain the two types of choices. We estimate the
parameters using only the even trials, and then simulate the model in odd trials in order
to compare the model predictions with the observed patterns out-of-sample. Because the
policy optimization and likelihood estimation methods that we use are stochastic, we display
simulations using the 30 top performing parameter con gurations to give a sense of the
uncertainty in the predictions. The parameter estimates were (mean std)x= 2:60:216,
= 0:5810:118,
switch = 0:009950:001,
sample = 0:003730:001, and = 364:081:2
As explained in the Methods, the units of these parameter estimates are standard deviations
of value (i.e., ).
In order to explore the role of the prior, we also t versions of the model in which the
prior bias term was xed to = 0 or = 1. The former corresponds to a strongly biased
prior and the latter corresponds to a completely unbiased prior. For = 0, the tted
parameters were x= 3:160:409,
switch = 0:008750:002,
sample = 0:003190:001,
and = 326:081:2. For = 1, they were x= 2:660:272,
switch = 0:01180:002,
sample = 0:005060:001, and = 330:097:9.
All of the gures below are based on model ts estimated at the group level on the
pooled data. However, for completeness we also t the model separately for each individual,
and report these ts in the Supplementary Information. We also carry out a validation of
our model tting approach in the SI (see Figure S1 for results).
7
A
5
0 5
Left rating - right rating0.000.250.500.751.00P(left chosen)
5
0 5
Left rating - mean other rating0.000.250.500.751.00
B
0200040006000
Total fixation time [ms]Density
0200040006000
Total fixation time [ms]
C
02468
Best rating - worst rating100015002000250030003500Total fixation time [ms]
02468
Best rating - mean other rating100015002000250030003500
D
02468
Mean item rating100015002000250030003500Total fixation time [ms]
02468
Mean item rating100015002000250030003500
Figure 3: Basic psychometrics. Each panel compares human data (black) and model predictions for binary choice (left,
two dots) and trinary choice (right, three dots). The main model predictions are shown in purple. The restricted model
predictions for the case of a highly biased prior mean ( = 0) is shown in blue, and for the case of a highly unbiased prior
mean ( = 1) are shown in pink. These colors were chosen to illustrate that the main model falls between these two extremes.
The aDDM predictions are shown in dashed green. Error bars (human) and shaded regions (model) indicate 95% con dence
intervals computed by 10,000 bootstrap samples (the model con dence intervals are often too small to be visible). Note that
the method used to compute and estimate the model parameters is noisy. To provide a sense of the e ect of this noise on the
main model predictions, we depict the predictions of the thirty best- tting parameter con gurations. Each light purple line
depicts the predictions for one of those parameters, whereas the darker purple line shows the mean prediction. In order to
keep the plot legible, only the mean predictions of the biased priors models are shown). (A) Choice probability as a function
of relative rating. (B) Kernel density estimation for the distribution of total xation time. Quartiles (25%, 50%, and 75%
quantiles) for the data, aDDM and main model predictions are shown at the bottom. (C) Total xation time as a function
of the relative rating of the highest rated item. (D) Total xation time as a function of the mean of all the item ratings
(overall value).
Results
We now investigate the extent to which the predictions of the model, tted on the even
trials, are able to account for observed choice, reaction time and xation patterns in the
out-of-sample odd trials.
Basic psychometrics
We begin by looking at basic psychometric patterns. Fig. 3A compares the choice curves
predicted by the model with the actual observed choices, separately for the case of binary
and trinary choice. It shows that the model captures well the in
uence of the items' true
values (as measured by liking ratings) on choice.
Fig. 3B plots the distribution of total xation times. This measure is similar to reaction
time except that it excludes time not spent xating on one of the items. We use total
xation time instead of reaction time because the model does not account for the initial
xation latency nor the time spent saccading between items (although it does account
for the opportunity cost of that time, through the
sample parameter). As shown in the
gure, the model provides a reasonable qualitative account of the distributions, although it
8
A
1 5 10
Number of fixations0.00.10.20.3Proportion of trials
1 5 10
Number of fixations0.00.10.20.3
B
02468
Best rating - worst rating3456Number of fixations
02468
Best rating - mean other rating3456
C
12345>5final
Fixation number30060090012001500Fixation duration [ms]
12345>5final
Fixation number30060090012001500
Figure 4: Basic xation patterns. (A) Histogram of number of xations in a trial. (B) Number of xations as a function of
decision diculty, as measured by the relative rating of the best item. (C) Duration of xation by xation number. Final
xations are excluded from all but the last bin. See Fig. 3 for more details.
underpredicts the mode in the case of two items and the skew in both cases.
Fig. 3C shows the relationship between total xation time and trial diculty, as mea-
sured by the relative liking rating of the best item. We nd that the model provides a
reasonable account of how total xation time changes with diculty. This prediction fol-
lows from that fact that fewer samples are necessary to detect a large di erence than to
either detect a small di erence or determine that the di erence is small enough to be unim-
portant. However, the model exhibits considerable variation in the predicted intercept and
substantially overpredicts total xation time in dicult trinary choices.
Finally, Fig. 3D shows the relationship between total xation time and the average rating
of all the items in the choice set. This \overall value e ect" has been emphasized in recent
research [13, 16] because it is consistent with multiplicative attention weighting (as in the
aDDM) but not an additive boosting model (e.g., ref. 11). Bayesian updating results in a
form of multiplicative weighting (speci cally, a hyperbolic function, c.f. 14), and thus our
model also predicts this pattern. Surprisingly, we do not see strong evidence for the overall
value e ect in the datasets we consider, but we note that the e ect has been found robustly
in several other datasets [13, 52{55]. Note that, in the binary case, the predicted overall
value e ect is symmetric around the prior mean; that is, choices between two very bad items
will also be made quickly. Indeed, with an unbiased-prior, the model predicts an inverted-U
relationship around the prior mean.
Several additional patterns in Fig. 3 are worth highlighting. First, all of the models
make similar and reasonable predictions of the psychometric choice curve and xation time
distributions. Second, the models with some prior bias provide a better account of the
xation time curves in binary choice than the unbiased model, and qualitatively similar
predictions to the aDDM. Finally, despite using a common set of parameters, all of the
models capture well the di erences between binary and trinary choice.
9
A
1000
500
05001000
Fixation advantage
of fixated item [ms]Density
1000
500
05001000
Fixation advantage
of fixated item [ms]
B
1000
500
05001000
Alternative fixation advantage
of fixated item [ms]0.00000.00050.00100.0015Density
C
200400600800100012001400
Alternative fixation advantage
of more-fixated item [ms]0.20.30.40.50.60.70.8P(fixate more-fixated item)
Figure 5: Uncertainty-directed attention. (A) Distribution of xation advantage of the xated item, computed at the
beginning of each new xation. Fixation advantage is de ned as the cumulative xation time to the item minus the mean
cumulative xation time to the other item(s). First xations are excluded in this plot. (B) Similar to A, except that we
compare the xation advantage between the xated item and the other item that could have been xated but was not. First
and second xations are excluded in this plot. (C) The probability that the item with greater alternative xation advantage
is xated, as a function of that advantage. See Fig. 3 for more details.
Basic xation properties
We next compare the predicted and observed xation patterns. An observed \ xation"
refers to a contiguous span of time during which a participant looks at the same item. A
predicted model xation refers to a continuous sequence of samples taken from one item.
Fig. 4A shows the distribution of the number of xations across trials. The model-
predicted distribution is reasonably similar to the observed data. However, in the two-item
case, the model is more likely to make only one xation, suggesting that people have a
tendency to xate both items at least once that the model does not capture.
Fig. 4B shows the relationship between the total number of xations and decision di-
culty. We nd that the model captures the relationship between diculty and the number
of xations reasonably well, with the same caveats as for Fig. 3b.
The original binary and trinary choice papers [9, 10] observed a systematic change in
xation durations over the course of the trial, as shown in Fig. 4C. Although the model tends
to underpredict the duration of the rst two xations in the three-item case, it captures well
three key patterns: (a) the nal xation is shorter, (b) later (but non- nal) xations are
longer and (c) xations are substantially longer in the two-item case. The nal prediction
is especially striking given that the model uses the same set of tted parameters for both
datasets. The model predicts shorter nal xations because they are cut o when a choice
is made [9, 10]. The model predicts the other patterns because more evidence is needed to
alter beliefs when their precision is already high; this occurs late in the trial, especially in
the two-item case where samples are split between fewer items.
Fig. 4 also shows that the main model provides a more accurate account than the aDDM
of how the number of xations changes with trial diculty, and of how xation duration
evolves over the course of a trial. One diculty in making this comparison is that the
aDDM assumes that non- nal xation durations are sampled from the observed empirical
distribution, conditional on a number of observable variables, and thus the accuracy of its
predictions regarding xation duration and xation number depends on the details of this
sampling. To maximize comparability with the existing literature, here we use the same
methods as in the original implementations [9, 10].
Uncertainty-directed attention
As we have seen, one of the key drivers of xations in the optimal policy is uncertainty about
the items' values. Speci cally, because the precision of the posteriors increases linearly with
the number of samples, the model predicts that, other things being equal, xations should
10
go to items that have received less cumulative xation time. However, the di erence in
precision must be large enough to justify paying the switching cost. In this section we
explore some of the xation patterns associated with this mechanism.
Fig. 5A depicts the distribution of relative cumulative xation time at the beginning of
a new xation, starting with the second xation. That is, at the onset of each xation, we
ask how much time has already been spent xating the newly xated item, compared to the
other items. In both cases, the actual and predicted distributions are centered below zero,
so that items tend to be xated when they have received less xation time than the other
items. Additionally, the model correctly predicts the lower mode and fatter left tail in the
two-item case.
Note, however, that a purely mechanical e ect can account for this basic pattern: the
item that is currently xated will on average have received the most xation time, but it
cannot be the target of a new xation, which drives down the xation advantage of newly
xated items. For this reason, it is useful to look further at the three-item case, which
a ords a stronger test of uncertainty-directed attention. In this case, the target of each new
xation (excluding the rst) must be one of the two items that are not currently xated.
Thus, comparing the cumulative xation times for these items avoids the previous confound.
Fig. 5B thus plots the distribution of xation time for the xated item minus that of the
item which could have been xated but was not. We see a similar pattern to Fig. 5A (right)
in both the data and model predictions. This suggests that uncertainty is not simply driving
the decision to make a saccade, but is also in
uencing the location of that saccade.
Fig. 5C explores this further by looking at the location of new xations in the three-item
case, as a function of the di erence in cumulative xation time between the two possible
xation targets. Although the more-previously- xated item is always less likely to be xated,
the probability of such a xation actually increases as its xation advantage grows. This
counterintuitive model prediction results from the competing e ects of value and uncertainty
on attention. Since items with high estimated value are xated more, an item that has been
xated much less than the others is likely to have a lower estimated value, and is therefore
less likely to receive more xations. However, we see that the predicted e ect is much
stronger than the observed e ect, and that the aDDM model provides a better account
of this pattern than our main model. However, note that the accuracy of this t follows
from the fact that the aDDM samples xation locations and durations from the empirical
distribution, conditioned on the previous three xation locations and the item ratings.
Value-directed attention
A second key driver of attention in the optimal policy is estimated value, which directs
xations to the twoitems with highest posterior means. As illustrated in Fig. 2A, this
implies that xation locations should be sensitive to relative estimated values in the trinary
but not in the binary case.
Although we cannot directly measure the participants' evolving value estimates, we can
use the liking ratings as a proxy for them because higher-rated items will tend to result
in higher value estimates. Using this idea, Fig. 6A shows the proportion of xation time
devoted to the left item as a function of its relative rating. Focusing rst on the three-item
case, both the model and data show a strong tendency to spend more time xating on higher
rated items (which are therefore likely to have higher estimated values). In the two item
case, the model simulations show a smaller but positive e ect. This is counterintuitive since
the model predicts that in the two-item case xation locations are insensitive the sign of
the relative value estimates (Fig. 2A). However, the patter likely arises due to the tendency
to xate last on the chosen item (see Fig. 7A below).
Fig. 6B provides an alternative test that avoids confounds associated with the nal
xation. It shows the duration of the rst xation, which is rarely nal, as a function of the
rating of the rst xated item. In the three-item case, both the model and data show longer
initial xations to high-rated items, although the model systematically underpredicts the
11
A
5
0 5
Left rating - right rating0.20.30.40.50.6Proportion fixate left
5
0 5
Left rating - mean other rating0.20.30.40.50.6
B
0 5 10
First fixated item rating200300400500600700First fixation duration [ms]
0 5 10
First fixated item rating200300400500600700
C
0500100015002000
Cumulative fixation time [ms]0.250.300.350.400.450.50P(fixate worst)
0500100015002000
Cumulative fixation time [ms]0.250.300.350.400.450.50
D
6
4
2
0246
Rating of first minus
second fixated item0.20.40.60.8P(4th fixation is refixation
to first fixated item)
E
246810
Rating of first fixated item0.10.20.30.40.5P(third fixation is refixation
to first fixated item)
Figure 6: Value-directed attention. (A) Proportion of time xating the left item as a function of its relative rating. (B) First
xation duration as a function of the rating of the rst- xated item. (C) Probability of xating the lowest rated item as a
function of the cumulative xation time to any of the items. (D) Probability that the fourth xation is to the rst- xated
item as a function of the di erence in rating between that item and the second- xated item. (E) Probability that the third
xation is to the rst xated item as a function of its rating. See Fig. 3 for more details.
mean rst xation duration. This prediction follows from the fact that, under the optimal
policy, xations are terminated when the xated item's estimated value falls out of the top
two (below zero for the rst xation); the higher the true value of the item, the less likely
this is to happen. In the two-item case, however, the model predicts that rst xation
duration should be largely insensitive to estimated value; highly valuable items actually
receive slightly shorter xations because these items are more likely to generate extremely
positive samples that result in terminating the rst xation and immediately choosing the
xated item. Consistent with this prediction, humans show little evidence for longer rst
xations to high-rated items in the binary case.
Previous work has suggested that attention may be directly in
uenced by the true value
of the items [17, 18, 56]. In our model, however, attention is driven only by the internal
value estimates generated during the decision making process. To distinguish between these
two accounts, we need a way to dissociate estimated value from true value. One way to
do this is by looking at the time course of attention. Early in the decision making process,
estimated values will be only weakly related to true value. However, with time the value
estimates become increasingly accurate and thus more closely correlate with true value.
Thus, if the decision maker always attends to the items with high estimated value, she
should be increasingly likely to attend to items with high true value as the trial progresses.
Fig. 6C shows the probability of xating on the worst item as a function of the cumulative
xation time to any of the items. In both the two- and three-item cases, the probability
begins near chance. In the three-item case, however, the probability quickly falls. This is
consistent with a model in which attention is driven by estimated value rather than value
itself.
The model makes even starker predictions in the three-item case. First, take all trials
in which the decision-maker samples from di erent items during the rst three xations.
Consider the choice of where to deploy the fourth xation. The model predicts that this
xation should be to the rst- xated item if its posterior mean is larger than that of the
12
A
5.0
2.5
0.02.55.0
Last fixated rating - other rating0.20.40.60.81.0P(last fixated item chosen)
5.0
2.5
0.02.55.0
Last fixated rating - mean other0.20.40.60.81.0
B
1500
1000
500
05001000
Final time advantage left [ms]0.00.20.40.60.8P(left chosen)
1500
1000
500
05001000
Final time advantage left [ms]0.00.20.40.60.8
C
2505007501000
First fixation duration [ms]0.00.20.40.60.8P(first fixated chosen)
2505007501000
First fixation duration [ms]0.00.20.40.60.8
Figure 7: Choice biases. (A) Probability that the last xated item is chosen as a function of its relative rating. (B)
Probability that the left item is chosen as a function of its nal xation advantage, given by total xation time to the left
item minus the mean total xation time to the other item(s). (C) Probability of choosing the rst-seen item as a function
of the rst- xation duration. See Fig. 3 for more details.
second- xated item, and vice versa. As a result, the probability that the fourth xation is
a re xation to the rst- xated item should increase with the di erence in ratings between
the rst- and second- xated items. As shown in Fig. 6D, the observed pattern follows the
model prediction.
Finally, the model makes a striking prediction regarding the location of the third xation
in the three-item case. Consider the choice of where to xate after the rst two xations.
The decision maker can choose to xate on the item that she has not seen yet, or to re xate
the rst- xated item. The model predicts a re xation to the rst-seen item if both that
item and the second-seen item already have high value estimates (leaving the un xated
item with the lowest value estimate). Consistent with this prediction, Fig. 6E shows that
the probability of the third xation being a re xation to the rst-seen item increases with
that item's rating. Note that the model with xed to zero (corresponding to a strong
prior bias), dramatically overpredicts the intercept. This is because this model greatly
underestimates the value of the not-yet- xated item.
Fig. 6 shows that our main model provides a better prediction of some of the xation
patterns, whereas the aDDM provides a better t of others. However, it it important to keep
in mind that whereas our model provides predictions for these xation patterns based on rst
principles, the predictions of the aDDM for these patterns are largely mechanistic since that
model samples xation locations and durations from the observed empirical distribution. As
a result, it is not surprising that Fig. 6B shows a better match between the aDDM and the
data since the predicted durations are, literally, sampled from the observed data conditional
on the rst item rating.
Choice Biases
Previous work has found a systematic positive correlation between relative xation time
and choice for appetitive (i.e., positively valenced) items [6, 7, 9, 10, 14, 19]. In particular,
models like the aDDM propose that an exogenous or random increase in xations towards
13
an appetive item increase the probability that it will be chosen, which leads to attention
driven choice biases. Here we investigate whether the optimal model can account for these
types of e ects.
Importantly, in the type of optimal xation model proposed here, there are two potential
mechanisms through which such correlations can emerge in the optimal model. The rst
is driven by the prior. If the prior mean is negatively biased, then sampling from an item
will on average increase its estimated value. This follows from the fact that sampling will
generally move the estimated value towards the item's true value, and a negatively biased
prior implies that the initial value estimate is generally less than the true value. The second
mechanism, which is only present in trinary choice, is the result of value-directed attention.
Here, the causal direction is
ipped, with value estimates driving xations rather than
xations driving value estimates. In particular, items with higher estimated value are both
more likely to be xated, and more likely to be chosen. Thus, xations and choice are
correlated through a common cause structure. Importantly, the two mechanisms are not
mutually exclusive; in fact, our model predicts that both will be in e ect for choice between
more than two items.
Fig. 7A shows that there is a sizable choice bias towards the last-seen item in both
datasets, as evidenced by the greater-than-chance probability of choosing an item whose
value is equal to the mean of the other items. Our model provides a strong quantitative
account of the pattern in trinary choice, but substantially underpredicts the e ect in binary
choice. Interestingly, it predicts a weaker e ect than the aDDM in the binary case, but a
stronger e ect in the trinary case.
To understand this result, it is important to think about the prior beliefs implicit in the
aDDM and related models [9, 10, 19]. Since these are not Bayesian models, they do not
posit an explicit prior that is then modi ed by evidence. However, the aDDM can be viewed
as an approximation to a Bayesian model with a prior centered on zero, as re
ected by the
initial point of the accumulator (zero) and the multiplicative discounting (the evidence for
the non-attended item is discounted towards zero). The latter roughly corresponds to the
Bayesian regularization e ect, wherein the posterior mean falls closer to the prior mean
when the likelihood is weak (low precision). Given this, our model predicts a weaker e ect
in the binary case because it has a weaker prior bias ( = 0:58) than the one implicit in
the aDDM ( = 0). Our model predicts a stronger e ect in the trinary case due to the
value-directed attention mechanism. Critically, although the aDDM accounts for the e ect
oftrue value on xations (by sampling from the empirical xation distribution), only the
optimal model accounts for the e ects of estimated value. Thus, conditioning on true value
(as we do in Fig. 7A) breaks the value-based attention mechanism in the aDDM but not
in the optimal model. Finally, note that the optimal model with = 0 provides a good
account of the bias in the binary case, but dramatically overpredicts it in the trinary case.
Fig. 7B shows that the average probability of choosing the left item increases substan-
tially with its overall relative xation time. As before, in comparison with the aDDM, the
optimal model provides better captures the full strength of the bias in the trinary case, but
underpredicts the e ect in the binary case. The optimal model with xed to zero performs
best in both cases. Note that the t of the aDDM is not as close as for similar gures in the
original papers because we simulate all models with the observed ratings (rather than all
possible combination of item ratings) and we consider a larger range of nal time advantage.
We replicate the original aDDM gures in the SI.
Finally, Fig. 7C shows that the probability of choosing the rst xated item increases
with the duration of the rst xation. Importantly, this gure shows that the attention-
choice correlation cannot be explained solely by the tendency to choose the last- xated item.
Again, all four models qualitatively capture the e ect, with varying degrees of quantitative
t.
14
Discussion
We have built a model of optimal information sampling during simple choice in order to
investigate the extent to which it can provide a quantitative account of xation patterns,
and their relationship with choices, during binary and trinary decisions. The model is based
on previous work showing that simple choices are based on the sequential accumulation of
noisy value samples [1, 40, 57{60] and that the process is modulated by visual attention
[7, 9, 10, 17, 19, 20, 61]. However, instead of proposing a speci c algorithmic model of
the xation and choice process, as is common in the literature, our focus has been on
characterizing the optimal xation policy and its implications. We build on previous work
on optimal economic decision-making in which samples are acquired for all options at the
same rate [36, 40{42], and extend it to the case of endogenous attention, where the decision
maker can control the rate of information acquired about each option. We formalized the
selection of xations as a problem of dynamically allocating a costly cognitive resource in
order to gain information about the values of the available options. Leveraging tools from
metareasoning in arti cial intelligence [46{49], we approximated the optimal solution to this
problem, which takes the form of a policy that selects which item to xate at each moment
and when to terminate the decision-making process.
We found that, despite its simplicity, the optimal model accounts for many key xation
and choice patterns in two in
uential binary and trinary choice datasets [9, 10]. The model
was also able to account for striking di erences between the two- and three-item cases using
a common set of parameters tted out of sample. More importantly, the results provide
evidence in favor of the hypothesis that the xation process is in
uenced by the evolving
value estimates, at least to some extent. Consider, for example, the increase in xation
duration over the course of the trial shown in Fig. 4C, the tendency to equate xation time
across items (Fig. 5B), and the relationship between the rating of the rst xated item and
the probability of re- xating it (Figs. 6D-E). These e ects are explained by our model, but
are hard to explain with exogenous xations, or with xations that are correlated with the
true value of the items, but not with the evolving value estimates (e.g., as in [17, 18, 62]).
Optimal information sampling models may appear inappropriate for value-based decision-
making problems, in which perceptual uncertainty about the identity of the di erent choice
items (often highly familiar junk foods) is likely resolved long before a choice is made. Two
features of the model ameliorate this concern. First,the samples underlying value-based de-
cisions are not taken from the external display (as in perceptual decisions), but are instead
generated internally, perhaps by some combination of mental simulation and memory recall
[43{45]. Second, the model makes the eye-mind assumption [15, 63]: what a person is look-
ing at is a good indicator of what they are thinking about. Importantly, these assumptions
implicitly underlie all sequential sampling models of value-based decision-making.
Our model is not the rst to propose that the xation and value-estimation processes
might interact reciprocally. However, no previous models fully capture the key character-
istics of optimal attention allocation, which appear to be at least approximated in human
xation behavior. For example, the Gaze Cascade Model [6] proposes that late in a trial
subjects lock-in xations on the favored option until a choice is made, [19] propose an aDDM
in which the probability of xating an item is given by a softmax over the estimated val-
ues, and [20] propose a Bayesian model of binary choice in which xations are driven by
relative uncertainty. In contrast to these models, the optimal model predicts that xations
are driven by a combination of the estimated uncertainty and relative values throughout
the trial, and that attention is devoted speci cally to the items with the top two value esti-
mates. Although the data strongly support the rst prediction, further data are necessary
to distinguish between the top-two rule and the softmax rule of [19].
Our results shed further light on the mechanisms underlying the classic attention-choice
correlation that has motivated previous models of attention-modulated simple choice. First,
our results highlight an important role of prior beliefs in sequential sampling models of simple
choice (c.f. ref 64). All previous models have assumed a prior mean of zero, either explicitly
15
[20, 64] or implicitly [9, 10, 19]. Such a prior is negatively biased when all or most items
have positive value, as is often the case in experimental settings. This bias is critical in
explaining the classic attention-choice correlation e ects because it creates a net-positive
e ect of attention on choice: if one begins with an underestimate, attending to an item will
on average increase its estimated value. However, we found that the best characterization
of the full behavior was achieved with a moderately biased prior, both in terms of our
approximate likelihood and in the full set of behavioral patterns in the plots.
Our results also suggest another (not mutually exclusive) mechanism by which the
attention-choice correlation can emerge: value-directed attention. We found that the opti-
mal model with no prior bias ( = 1) predicts an attention-choice correlation in the trinary
choice case. This is because, controlling for true values, an increase in estimated value (e.g.,
due to sample noise) makes the model more likely to both xate and choose an item. This
could potentially help to resolve the debate over additive vs. multiplicative e ects of atten-
tion on choice [11, 13]. While the prior-bias mechanism predicts a multiplicative e ect, the
value-directed attention mechanism predicts that xation time and choice will be directly
related (as predicted by the additive model). Although we did not see strong evidence
for value-directed attention in the binary dataset, such a bias has been shown in explicit
information gathering settings [65] and could be at work in other binary choice settings.
Our work most closely relates to two recent lines of work on optimal information sam-
pling for simple choice. First, H ebert and Woodford [66, 67] consider sequential sampling
models based on rational inattention. They derive optimal sampling strategies under highly
general information-theoretic constraints, and establish several interesting properties of op-
timal sampling, such as the conditions under which the evidence accumulation will resemble
a jump or a di usion process. In their framework, the decision maker chooses, at each time
point, an arbitrary information structure , the probability of producing each possible signal
under di erent true states of the world. In contrast, we specify a very small set of informa-
tion structures, each of which corresponds to sampling a noisy estimate of one item's value
(Equation 1). This naturally associates each information structure with xating on one of
the items, allowing us to compare model predictions to human xation patterns. Whether
human attention more closely resembles
exible construction of dynamic information struc-
tures, or selection from a small set of xed information structures is an interesting question
for future research.
In a second line of work, concurrent to our own, Jang, Sharma, and Drugowitsch [64]
develop a model of optimal information sampling for binary choice with the same Bayesian
structure as our model and compare their predictions to human behavior in the same binary
choice dataset that we use [9]. There are three important di erences between the studies.
First, they consider the possibility that samples can also be drawn in parallel for the unat-
tended item, but with higher variance. However, they nd that a model in which almost
no information is acquired for the unattended item ts the data best, consistent with the
assumptions of our model. Second, they use dynamic programming to identify the optimal
attention policy almost exactly. This allows them to more accurately characterize truly
optimal attention allocation. However, dynamic programming is intractable for more than
two items, due to the curse of dimensionality. Thus they could not consider trinary choice,
which is of special interest because only this case makes value-directed attention optimal,
and forces the decision-maker to decide which of the unattended items to xate next, rather
than simply when to switch to the other item. Third, they assumed (following previous
work) that the prior mean is zero. In contrast, by varying the prior, we show that although
a biased prior is needed to account for the attention-choice correlation in binary choice, the
data is best explained by a model with only a moderately biased prior mean, about halfway
between zero and the empirical mean.
We can also draw insights from the empirical patterns that the model fails to capture.
These mismatches suggest that the model, which was designed to be as simple as possible,
is missing critical components that should be explored in future work. For example, the
underprediction of xation durations early in the trial could be addressed by more realistic
16
constraints on the xation process such as inhibition of return, and the overprediction of the
proportion of single- xation trials in the two-item case could be explained with uncertainty
aversion. Although not illustrated here, the model's accuracy could be further improved by
including bottom-up in
uences on xations (e.g., spatial or saliency biases [18, 68]).
While we have focused on attention in simple choice, other studies have explored the role
of attention in more complicated multi-attribute choices [5, 69{78]. None of these studies
have carried out a full characterization of the optimal sampling process or how it compares
to observed xation patterns, although see [79, 80] for some related results. Extending the
methods in this paper to that important case is a priority for future work.
Methods
The model was implemented in the Julia programming language [81]. The data and code are
available at https://github.com/fredcallaway/optimal-fixations-simple-choice .
Attention allocation as a metalevel Markov decision process
To characterize optimal attention allocation in our model, we cast the model as a metalevel
Markov decision process (MDP) [48]. Like a standard MDP, a metalevel MDP is de ned
by a set of states, a set of actions, a transition function giving the probability of moving
to each state by executing a given action in a given state, and a reward function giving
the immediate utility gained by executing a given action in a given state. In a metalevel
MDP, the states, B, correspond to beliefs (mental states), and the actions, C, correspond to
computations (cognitive operations). However, formally, it is identical to an MDP, and can
be interpreted as such.
In our model, a belief state, b2B, corresponds to a set of posterior distributions over
each item's value. Because the distributions are Gaussian, the belief can be represented by
two vectors, and, that specify the mean and precision of each distribution. That is
p(u(i)jb) = Gaussian( u(i);(i);1=(i))
To model the switching cost, the belief state must also encode the currently attended item,
i.e., the item sampled last (taking a null value, , in the initial belief). Thus, a belief is a
tuplebt= (t;t;lastt). The dimensionality of the belief space is 2 N+ 1 whereNis the
number of items.
A computation, c2C, corresponds to sampling an item's value and updating the corre-
sponding estimated value distribution. There are Nsuch computations, one for each item.
Additionally, all metalevel MDPs have a special computation, ?, that terminates the com-
putation process (in our case, sampling) and selects an optimal external action given the
current belief state (in our case, choosing the item with maximal posterior mean).
The metalevel transition function describes how computations update beliefs. In our
model, this corresponds to the sampling and Bayesian belief updating procedure speci ed in
Equation 2, which we reproduce here for the reader's convenience. Note that we additionally
make explicit the variable that tracks the previously sampled item. Given the current belief,
bt= (t;t;lastt), and computation, c, the next belief state, bt+1= (t+1;t+1;lastt+1), is
sampled from the following generative process:
xtGaussian(u(c);2
x)
(c)
t+1=(c)
t+ 2
x
(c)
t+1= 2
xxt+(c)
t(c)
t
(c)
t+1
(i)
t+1=(i)
tand(i)
t+1=(i)
tfori6=c:
lastt+1=c(6)
17
Finally, the metalevel reward function incorporates both the cost of computation and
the utility of the chosen action. The metalevel reward for sampling is de ned
R(bt;ct) = cost(bt;ct) = (
sample +1(lastt6= ^ct6= lastt)
switch ):
That is, the cost of sampling includes a xed cost,
sample , as well as an additional switching
cost,
switch , that is paid when sampling from a di erent item than that sampled on the
last time step. We assume that this cost is not paid for the rst xation; however, this
assumption has no e ect on the optimal policy for reasonable parameter values. The action
utility is the true value of the chosen item, i.e., u(i
T)wherei
T= argmax
i(i)
T. The metalevel
reward for the termination computation, ?, is the expectation of this value. Because we
assume accurate priors and Bayesian belief updating, this expectation can be taken with
respect to the agent's own beliefs [48], resulting in
R(bt;?) = Eh
u(i
T) bti
= max
i(i)
t:
Optimal metalevel policy
The solution to a metalevel MDP takes the form of a Markov policy, , that stochastically
selects which computation to take next given the current belief state. Formally, ct(bt).
The optimal metalevel policy, , is the one that maximizes expected total metalevel reward,
= argmax
E"TX
tR(bt;ct) ct(bt)#
:
ReplacingRwith its de nition, we see that this requires striking a balance between the
expected value of the chosen item and the computational cost of the samples that informed
the choice,
= argmax
E"
max
i(i)
T T 1X
tcost(bt;ct) ct(bt)#
:
That is, one wishes to acquire accurate beliefs that support selecting a high-value item,
while at the same time minimizing the cost of the samples necessary to attain those beliefs.
This suggests a strategy for selecting computations optimally. For each item, estimate how
much one's decision would improve if one sampled from it (and then continued sampling
optimally). Subtract from this number the cost of taking the sample (and also the estimated
cost of the future samples). Now identify the item for which this value is maximal. If it is
positive, it is optimal to take another sample for this item; otherwise, it is optimal to stop
sampling and make a decision.
This basic logic is formalized in rational metareasoning as the value of computation
(VOC) [47]. Formally, VOC( b;c) is de ned as the expected increase in total metalevel
reward if one executes a single computation, c, and continues optimally rather than making
a choice immediately (i.e., executing ?):
VOC(bt;c) =R(bt;c) + E"TX
t0=t+1R(bt0;ct0) ct0(bt0)#
R(b;?):
In our model, this can be rewritten
VOC(bt;c) = cost(bt;c) + E"
max
i(i)
T T 1X
t0=t+1cost(bt0;ct0) ct0(bt0)#
max
i(i)
t:
That is, the VOC for sampling a given item in some belief state is the expected improvement
in the value of the chosen item (rather than making a choice based on the current belief)
minus the cost of sampling that item and the expected cost of all future samples.
18
We can then de ne the optimal policy as selecting computations with maximal VOC:
(b)Uniform(argmax
cVOC(b;c)):
For those familiar with reinforcement learning, this recursive joint de nition of and VOC
is exactly analogous to the joint de nition of the optimal policy with the state-action value
function,Q[82]. Indeed, VOC( b;c) =Q(b;c) R(b;?).
Finally, by de nition, VOC( b;?) = 0 for all b. Thus, the optimal policy terminates
sampling when no computation has a positive VOC.
Approximating the optimal policy
For small discrete belief spaces, the optimal metalevel policy can be computed exactly using
standard dynamic programming methods such as value iteration or backwards induction
[83]. These methods can also be applied to low-dimensional, continuous belief spaces by
rst discretizing the space on a grid [41], and this approach has recently been used to
characterize the optimal xation policy in binary choice [64]. Unfortunately, these methods
are infeasible in the trinary choice case, since the belief space has six continuous dimensions.
Instead, we approximate the optimal policy by extending the method proposed in [49]. This
method is based on an approximation of the VOC as a linear combination of features,
[VOC(b;c;w) =w1VOI myopic (b;c) +w2VOI item(b;c) +w3VOI full(b) (cost(c) +w4);
(7)
for all for all c6=?, with[VOC(b;?;w) = VOC(b;?) = 0.
We brie
y de ne the features here, and provide full derivations in the SI appendix.
The VOI terms quantify the value of information [84] that might be gained by di erent
additional computations. Note that the VOI is di erent from the VOC because the latter
includes the costs of computation as well as its bene ts. In general, the VOI is de ned as the
expected improvement in the utility of the action selected based on additional information
rather than the current belief state: E~bjb[R(~b;?) R(b;?)], where ~bis a hypothetical future
belief in which the information has been gained, the distribution of which depends on the
current belief.
VOI myopic (b;c) denotes the expected improvement in choice utility from drawing one
additional sample from item cbefore making a choice, as opposed to making a choice
immediately based on the current belief, b. VOI item(b;c) denotes the expected improvement
from learning the true value of item c, and then choosing the best item based on that
information. Finally, VOI full(b) denotes the improvement from learning the true value of
every item and then making an optimal choice based on that complete information.
Together, these three features approximate the expected value of information that could
be gained by the (unknown) sequence of future samples. Importantly, this true value of
information always lies between the lower bound of VOI myopic and the upper bound of
VOI full(see Fig. S4), implying that the true VOI is a convex combination of these two
terms. Note, however, that the weights on this combination are not constant across beliefs,
as assumed in our approximation. Thus, including the VOI itemterm, improves the accuracy
of the approximation, by providing an intermediate value between the two extremes. Finally,
the last two terms in Equation 7 approximate the cost of computation: cost( c) is the cost of
carrying out computation candw4approximates the expected future costs incurred under
the optimal policy.
Although maximizing [VOC(b;c;w) identi es the policy with best performance, it is un-
likely that humans make attentional decisions using such perfect and noiseless maximization.
Thus, we assume that computations are chosen using a Boltzmann (softmax) distribution
[51] given by
(cjb;w; )/expn
[VOC(b;c;w)o
;
19
where the inverse temperature, , is a free parameter that controls the degree of noise.
Note that computation selection is fully random when = 0 and becomes deterministic as
!1 .
To identify the weights used in the approximation, we rst assume that wi0 and
w1+w2+w3= 1, since w1:3features form a convex combination and w4captures the
non-negative future cost. Previous work [49] used Bayesian optimization to identify the
weights within this space that maximize total expected metalevel reward. However, we
found that often a large area of weight space resulted in extremely similar performance,
despite inducing behaviorally distinct policies. Practically, this makes identifying a unique
optimal policy challenging, and theoretically we would not expect all participants to follow
a single unique policy when there is a wide plateau of high-performing policies. To address
this, we instead identify a set of near-optimal policies and assume that human behavior will
conform to the aggregate behavior of this set.
To identify this set of near-optimal policies, we apply a method based on Upper Con -
dence Bound (UCB) bandit algorithms [85]. We begin by sampling 8000 weight vectors to
roughly uniformly tile the space of possible weights. Concretely, we divide a three dimen-
sional hypercube into 800 = 203equal-size boxes and sample a point uniformly from each
box. The rst two dimensions are bounded in (0 ;1) and are used to produce w1:3using
the following trick: Let x1andx2be the lower and higher of the two sampled values. We
then de ne w1:3= [x1;x2 x1;1 x2]. Ifx1andx2are uniformly sampled from (0 ;1),
and indeed they are, then this produces w1:3uniformly sampled from the 3-simplex. The
third dimension produces the future cost weight; we set w4=x3maxcost where maxcost
is the lowest cost for which no computation has positive [VOC in the initial belief state. We
then simulate 100 decision trials for each of the resulting policies, providing a baseline level
of performance. Using these simulations, we compute an upper con dence bound of each
policy's performance equal to ^ i+3^i, where ^iand ^iare the empirical mean and standard
deviations of the metalevel returns sampled for policy i. A standard UCB algorithm would
then simulate from the policy maximizing this value. However, because we are interested in
identifying a set of policies, we instead select the top 80 (i.e. 1% of) policies and simulate 10
additional trials for each, updating ^ iand ^ifor each one. We iterate this step 5000 times.
Finally, we select the 80 policies with highest expected performance as our characterization
of optimal behavior in the metalevel MDP. To eliminate the possibility of tting noise in the
optimization procedure, we use one set of policies to compute the likelihood on the training
data and re-optimize a new set of policies to generate plots and compute the likelihood of
the test data. Note that we use the box sampling method described in the previous para-
graph rather than a deterministic low discrepancy sampling strategy [86] so that the set of
policies considered are not exactly the same in the tting and evaluation stages.
How good is the approximation method? Previous work found that this approach gener-
ates near-optimal policies on a related problem, with Bernoulli-distributed samples and no
switching costs [49]. Note that in the case of Bernoulli samples, the belief space is discrete
and thus the optimal policy can be computed exactly if an upper bound is placed on the
number of computations that can be performed before making a decision. Although intro-
ducing switching costs makes the metareasoning problem more challenging to solve, in the
Bernoulli case we have found that they only induce a modest reduction in the performance
of the approximation method relative to the full optimal policy, achieving 92% of optimal
reward in the worst case (see SI for details). This suggests that this method is likely to pro-
vide a reasonable approximation to the optimal policy in the model with Gaussian samples
used here, but a full veri cation of this fact is beyond the scope of the current study.
Implementation of the prior
In the main text, we speci ed the prior as a property of the initial belief state. However,
for technical reasons (in particular, to reuse the same set of optimized policies for multiple
values of ), it is preferable to perform policy optimization and simulation in a standardized
20
space, in which the initial belief state has 0=0and0=1. We then capture the prior
over the ratings of items in the experiment by transforming the ratings into this standardized
space such that the transformed values are in units de ned by the prior. Concretely, given
an item rating r(i), we set the true value to
u(i)=r(i)
; (8)
where and denote the prior mean and standard deviation. Modulo the resultant change
in units (all parameter values are divided by ), this produces the exact same behavior as
the na ve implementation, in which the initial belief itself varies.
There is one non-trivial consequence of using this approach when jointly tting multiple
datasets: The jointly t parameters are estimated in the standardized space, rather than
the space de ned by the raw rating scale. As a result, if we transform the parameters back
into the raw rating space, the parameters will be slightly di erent for the two datasets (even
though they are identical in the transformed space). This was done intentionally because we
expect that the parameters will be consistent in the context-independent units (i.e., standard
deviations of an internal utility scale). However, this decision turns out to have negligible
impact in our case because the empirical rating distributions are very similar. Speci cally,
the empirical rating distributions are (mean std) 3:4922:631 for the binary dataset
and 4:2952:524 for the trinary dataset. Due to the di erence in standard deviations, all
parameters (except , which is not a ected) are 2 :631=2:524 = 1:042 times larger in the
raw rating space for the binary dataset compared to the trinary dataset. The di erence
in empirical means a ects , which is 3 :492=4:295 = 0:813 times as large in the binary
compared to trinary dataset. However, given our interpretation of as a degree of updating
towards the empirical mean, this di erence is as intended.
Model simulation procedure
Given a metalevel MDP and policy, , simulating a choice trial amounts to running a single
episode of the policy on the metalevel MDP. To run an episode, we rst initialize the belief
state,b0= (0=0;0=1;last0= ). Note that last 0= indicates that no item is
xated at the onset of a trial.
The agent then selects an initial computation c0(b0) and the belief is updated ac-
cording to the transition dynamics (Equation 6). Note that (cjb0) assigns equal sampling
probability to all of the items, since the subject starts with symmetrical beliefs. This pro-
cess repeats until some time step, T, when the agent selects the termination action, ?. The
predicted choice is the item with maximal posterior value, i
T= argmax
i(i)
T. In the event
of a tie, the choice is sampled uniformly from the set of items with maximal expected value
in the nal belief state; in practice, this never happens with well- tting parameter values.
To translate the sequence of computations into a xation sequence, we assume that each
sample takes 100 ms and concatenate multiple contiguous samples from the same item into
one xation. The temporal duration of a sample is arbitrary; a lower value would result in
ner temporal predictions, but longer runtime when simulating the model. In this way, it is
very similar to the dtparameter used in simulating di usion decision models. Importantly
the qualitative predictions of the model are insensitive to this parameter because xand
sample can be adjusted to result in the same amount of information and cost per ms.
We simulate the model for two di erent purposes: (1) identifying the optimal policy
and (2) comparing model predictions to human behavior. In the former case, we randomly
sample the true utilities on each \trial" i.i.d. from Gaussian(0 ;1). This corresponds to the
assumption that the xation policy is optimized for an environment in which the DM's prior
is accurate. When simulating a speci c trial for comparison to human behavior, the true
value of each item is instead determined by the liking ratings for the items presented on
that trial, as speci ed in Equation 8.
21
Model parameter estimation
The model has ve free parameters: the standard deviation of the sampling distribution,
x, the cost per sample,
sample , the cost of switching attention,
switch , the degree of prior
updating, , and the inverse temperature of the Boltzmann policy, . We estimate a single
set of parameters at the group level using approximate maximum likelihood estimation in
the combined two- and three-item datasets, using only the even trials.
To brie
y summarize the estimation procedure: given a candidate set of parameter val-
ues, we construct the corresponding metalevel MDP and identify a set of 80 near-optimal
policies for that MDP. We then approximate the likelihood of the human xation and choice
data using simulations from the optimized policies. Finally, we perform this full procedure
for 70,000 quasi-randomly sampled parameter con gurations and report the top thirty con-
gurations (those with the highest likelihood) to give a rough sense of the uncertainty in
the model predictions. A parameter recovery exercise (reported in the SI) suggests that
this method, though approximate, is sucient to identify the parameters of the model with
fairly high accuracy. Below, we explain in detail how we estimate and then maximize the
approximate likelihood.
The primary challenge in tting the model is in estimating the likelihood function. In
principle, we could seek to maximize the joint likelihood of the observed xation sequences
and choices. However, like most sequential sampling models, our model does not have
an analytic likelihood function. Additionally, the high dimensionality of the xation data
makes standard methods for approximating the likelihood [87, 88] infeasible. Thus, taking
inspiration from Approximate Bayesian Computation methods [89, 90], we approximate the
likelihood by collapsing the high dimensional xation data into four summary statistics:
the identity of the chosen item, the number of xations, the total xation time, and the
proportion of xation time on each item. As described below, we estimate the joint likelihood
of these summary statistics as a smoothed histogram of the statistics in simulated trials,
and then approximate the likelihood of a trial by the likelihood of its summary statistics.
We emphasize, however, that we do not use this approximate likelihood to evaluate the
performance of the model. Instead, we intend it to be a maximally principled (and minimally
researcher-speci ed) approach to choosing model parameters, given that computing a true
likelihood is computationally infeasible.
Given a set of near-optimal policies, we estimate the likelihood of the summary statistics
for each trial using a smoothed histogram of the summary statistics in simulated trials.
Critically, this likelihood is conditional on the ratings for the item in that trial. However,
it depends only on the (unordered) set of these ratings; thus, we estimate the conditional
likelihood once for each such set. Given a set of ratings, we simulate the model 625 times
for each of the 80 policies, using the resulting 50,000 simulations to construct a histogram of
the trial summary statistics. The continuous statistics (total and proportion xation times)
are binned into quintiles (i.e., ve bins containing equal amounts of the data) de ned by
the distribution in the experimental data. For the xation proportions, the quintiles are
de ned on the rating rank of the item rather than the spatial location because we expect
the distributions to depend on relative rating in the three-item case. Values outside the
experimental range are placed into the corresponding tail bin. Similarly trials with ve
or more xations are all grouped into one bin (including e.g. six and seven xations) and
cases in which the model predicts zero xations are grouped into the one- xation bin. This
latter case corresponds to choosing an item immediately without ever sampling, and occurs
rarely in well- tting instantiations of the model, but happens frequently when
sample is
set too high. For each simulation, we compute the binned summary statistics, identify the
corresponding cell in the histogram, and increase its count by one. Finally, we normalize this
histogram, resulting in a likelihood over the summary statistics. To compute the likelihood
of a trial,L(dj), we compute the binned summary statistics for the trial and look up the
corresponding value in the normalized histogram for that trial's rating set.
To account for trials that are not well explained by our model, we use add- nsmoothing,
22
wherenwas chosen independently for each so as to maximize its likelihood. This is equiv-
alent to assuming a mixture between the empirical distribution and a uniform distribution
with mixing weight . Thus, the full approximate likelihood is
L(Dj) = max
2[0;0:5]Y
d2D
1
C+ (1 )L(dj)
;
whereC=N5N+1is the total number of cells in the histogram. Importantly, this error
model is only used to approximate the likelihood; it is not used for generating the model
predictions in the gures|indeed, it could not be used in this way because the error model
is de ned over the summary statistics, and cannot generate full sequences of xations. Thus,
theparameter should be interpreted in roughly the same way as the bandwidth parameter
of a kernel density estimate [87], rather than as an additional free parameter of the model.
We then use this approximate likelihood function to identify a maximum likelihood
estimate, ^= arg maxL(Dj). Based on manual inspection, we identi ed the promising
region of parameter space to be x2(1;5),
sample2(0:001;0:01),
switch2(0:003;0:03),
and 2(100;500). We then ran an additional quasi-random search of 10,000 points within
this space using Sobol low-discrepancy sequences [86]. This approach has been shown to
be more e ective than both grid search and random search, while still allowing for massive
parallelization [91].
Note that the optimal policy does not depend on because the DM believes her prior to
be unbiased (by de nition) and makes her xation decisions accordingly. The alternative,
optimizing the policy conditional on , would imply that the DM is internally inconsistent,
accounting for the bias in her xations but not in the prior itself. Thus, we optimize
separately from the other parameters. Speci cally, we consider 10,000 possible instantiations
of all the other parameters, nd optimal policies once for each instantiation, and evaluate
the likelihood for seven values of ; these seven values included the special cases of 0 and 1
as well as ve additional randomly-spaced values with a random o set (roughly capturing
the low-discrepancy property of the Sobol sequence).
We found that the stochasticity in the policy optimization and likelihood estimation
coupled with weak identi ability for some parameters resulted in slightly di erent results
when re-running the full procedure; thus, to give a rough sense of the uncertainty in the
estimate, we identify the top thirty parameters, giving us both mean and standard deviation
for each parameter and the total likelihood.
Acknowledgements This research was supported by a grant from Facebook Reality Labs.
Antonio Rangel gratefully acknowledges support from the NOMIS foundation. We thank
Ian Krajbich for his help in simulating the aDDM, and Bas van Opheusden for suggesting
the method for eciently computing VOI full.
References
[1] Ratcli R, McKoon G. The di usion decision model: theory and data for two-choice
decision tasks. Neural computation. 2008;20(4):873{922.
[2] Ratcli R, Smith PL, Brown SD, McKoon G. Di usion Decision Model: Current Issues
and History. Trends in Cognitive Science. 2016;20(4):260{281.
[3] Milosavljevic M, Malmaud J, Huth A, Koch C, Rangel A. The Drift Di usion Model
can account for the accuracy and reaction time of value-based chioces under high and
low time pressure. Judgment and Decision Making. 2010;5(6):437{449.
[4] Usher M, McClelland JL. The Time Course of Perceptual Choice: The Leaky, Com-
peting Accumulator Model. Psychological Review. 2001;108(3):550{592.
23
[5] Usher M, McClelland JL. Loss Aversion and Inhibition in Dynamical Models of Mul-
tialternative Choice. Psychological Review. 2004;111(3):757{769.
[6] Shimojo S, Simion C, Shimojo E, Scheier C. Gaze bias both re
ects and in
uences
preference. Nature Neuroscience. 2003;6(12):1317{1322.
[7] Armel KC, Beaumel A, Rangel A. Biasing Simple Choices by Manipulating Relative
Visual Attention. Judgment and Decision Making. 2008;3(5):396{403.
[8] Glaholt MG, Reingold EM. Stimulus exposure and gaze bias: A further test of the gaze
cascade model. Attention, Perception & Psychophysics. 2009;71(3):445{450.
[9] Krajbich I, Armel C, Rangel A. Visual xations and the computation and comparison
of value in simple choice. Nature Neuroscience. 2010;13(10):1292{1298.
[10] Krajbich I, Rangel A. Multialternative drift-di usion model predicts the relationship
between visual xations and choice in value-based decisions. Proceedings of the National
Academy of Sciences. 2011;108(33):13852{13857.
[11] Cavanagh JF, Wiecki TV, Kochar A, Frank MJ. Eye tracking and pupillometry are
indicators of dissociable latent decision processes. Journal of Experimental Psychology:
General. 2014;143(4):1476{1488.
[12] Tavares G, Perona P, Rangel A. The Attentional Drift Di usion Model of Simple
Perceptual Decision-Making. Frontiers in Neuroscience. 2017;11.
[13] Smith SM, Krajbich I. Gaze Ampli es Value in Decision Making. Psychological Science.
2019;30(1):116{128.
[14] Armel KC, Rangel A. Neuroeconomic models of economic decision making: The impact
of computation time and experience on decision values. American Economic Review.
2008;98(2):163{168.
[15] Orquin JL, Mueller Loose S. Attention and Choice: A Review on Eye Movements in
Decision Making. Acta Psychologica. 2013;144(1):190{206.
[16] Krajbich I. Accounting for Attention in Sequential Sampling Models of Decision Mak-
ing. Current Opinion in Psychology. 2018;29:6{11.
[17] Gluth S, Spektor MS, Rieskamp J. Value-based attentional capture a ects multi-
alternative decision making. Elife. 2018;7:e39659.
[18] Towal RB, Mormann M, Koch C. Simultaneous modeling of visual saliency and value
computation improves predictions of economic choice. Proceedings of the National
Academy of Sciences. 2013;110(40):E3858{E3867.
[19] Gluth S, Kern N, Kortmann M, Vitali CL. Value-Based Attention but Not Divisive Nor-
malization In
uences Decisions with Multiple Alternatives. Nature Human Behaviour.
2020;4(6):634{645.
[20] Song M, Wang X, Zhang H, Li J. Proactive information sampling in value-based
decision-making: Deciding when and where to saccade. Frontiers in Human Neuro-
science. 2019;13(February):1{10.
[21] Gottlieb J, Oudeyer PY. Towards a neuroscience of active sampling and curiosity.
Nature Reviews Neuroscience. 2018;19(12):758{770.
[22] Najemnik J, Geisler WS. Optimal eye movement strategies in visual search. Nature.
2005;434(7031):387{391.
[23] Eckstein MP. Visual search: A retrospective. Journal of Vision. 2011;11(5):1{36.
24
[24] Cassey TC, Evens DR, Bogacz R, Marshall JAR, Ludwig CJH. Adaptive Sampling of
Information in Perceptual Decision-Making. PLOS ONE. 2013;8(11):e78993.
[25] Ludwig CJ, Evens DR. Information foraging for perceptual decisions. Journal of
Experimental Psychology: Human Perception and Performance. 2017;43(2):245{264.
[26] Itti L, Baldi P. Bayesian surprise attracts human attention. Vision Research.
2009;49(10):1295{1306.
[27] Gottlieb J, Oudeyer PY, Lopes M, Baranes A. Information-Seeking, Curiosity, and
Attention: Computational and Neural Mechanisms. Trends in Cognitive Sciences.
2013;17(11):585{593.
[28] Savage LJ. The Foundations of Statistics. The Foundations of Statistics. Oxford,
England: John Wiley & Sons; 1954.
[29] Von Neumann J, Morgenstern O. Theory of Games and Economic Behavior. Princeton,
NJ, US: Princeton University Press; 1944.
[30] Lewis RL, Howes A, Singh S. Computational Rationality: Linking Mechanism
and Behavior through Bounded Utility Maximization. Topics in Cognitive Science.
2014;6(2):279{311.
[31] Griths TL, Lieder F, Goodman ND. Rational Use of Cognitive Resources: Levels of
Analysis between the Computational and the Algorithmic. Topics in Cognitive Science.
2015;7(2):217{229.
[32] Lieder F, Griths TL. Resource-Rational Analysis: Understanding Human Cogni-
tion as the Optimal Use of Limited Computational Resources. Behavioral and Brain
Sciences. 2019;.
[33] Gershman SJ, Horvitz EJ, Tenenbaum JB. Computational Rationality: A Converging
Paradigm for Intelligence in Brains, Minds, and Machines. Science. 2015;349(6245).
[34] Sims CA. Stickiness. Carnegie-Rochester Conference Series on Public Policy.
1998;49:317{356.
[35] Caplin A, Dean M. Behavioral Implications of Rational Inattention with Shannon
Entropy. NBER Working Paper. 2013;(August):1{40.
[36] Bogacz R, Brown E, Moehlis J, Holmes P, Cohen JD. The physics of optimal decision
making: A formal analysis of models of performance in two-alternative forced-choice
tasks. Psychological Review. 2006;113(4):700{765.
[37] Moreno-Bote R. Decision Con dence and Uncertainty in Di usion Models with Par-
tially Correlated Neuronal Integrators. Neural Computation. 2010;22(7):1786{1811.
[38] Drugowitsch J, Moreno-Bote R, Churchland AK, Shadlen MN, Pouget A. The Cost
of Accumulating Evidence in Perceptual Decision Making. Journal of Neuroscience.
2012;32(11):3612{3628.
[39] Bitzer S, Park H, Blankenburg F, Kiebel SJ. Perceptual decision making: drift-
di usion model is equivalent to a Bayesian model. Frontiers in Human Neuroscience.
2014;8(February):1{17.
[40] Tajima S, Drugowitsch J, Pouget A. Optimal policy for value-based decision-making.
Nature Communications. 2016;7:1{12.
[41] Tajima S, Drugowitsch J, Patel N, Pouget A. Optimal policy for multi-alternative
decisions. Nature Neuroscience. 2019;22(9):1503{1511.
25
[42] Fudenberg D, Strack P, Strzalecki T. Speed, accuracy, and the optimal timing of
choices. American Economic Review. 2018;108(12):3651{3684.
[43] Biderman N, Bakkour A, Shohamy D. What Are Memories For? The Hippocam-
pus Bridges Past Experience with Future Decisions. Trends in Cognitive Sciences.
2020;24(7):542{556.
[44] Bakkour A, Palombo DJ, Zylberberg A, Kang YHR, Reid A, Verfaellie M, et al.
The Hippocampus Supports Deliberation during Value-Based Decisions. eLife.
2019;8:unde ned{unde ned.
[45] Wang S, Feng SF, Bornstein A. Mixing memory and desire: How memory reactivation
supports deliberative decision-making. 2020;.
[46] Matheson JE. The Economic Value of Analysis and Computation. IEEE Transactions
on Systems Science and Cybernetics. 1968;4(3):325{332.
[47] Russell S, Wefald E. Principles of metareasoning. Arti cial Intelligence. 1991;49(1-
3):361{395.
[48] Hay N, Russell S, Tolpin D, Shimony SE. Selecting Computations: Theory and Appli-
cations. In: Proceedings of the Twenty-Eighth Conference on Uncertainty in Arti cial
Intelligence. UAI'12. Arlington, Virginia, USA: AUAI Press; 2012. p. 346{355.
[49] Callaway F, Gul S, Krueger P, Griths TL, Lieder F. Learning to select computations.
In: Uncertainty in Arti cial Intelligence: Proceedings of the Thirty-Fourth Conference;
2018.
[50] Gold JI, Shadlen MN. Banburismus and the Brain: Decoding the Relationship between
Sensory Stimuli, Decisions, and Reward. Neuron. 2002;36(2):299{308.
[51] McFadden D. Economic choices. American Economic Review. 2001;91(3):351{378.
[52] Fr omer R, Dean Wolf CK, Shenhav A. Goal Congruency Dominates Reward Value
in Accounting for Behavioral and Neural Correlates of Value-Based Decision-Making.
Nature Communications. 2019;10(1):1{11.
[53] Hunt LT, Kolling N, Soltani A, Woolrich MW, Rushworth MFS, Behrens TEJ. Mecha-
nisms Underlying Cortical Activity during Value-Guided Choice. Nature Neuroscience.
2012;15(3):470{476.
[54] Polan a R, Krajbich I, Grueschow M, Ru CC. Neural Oscillations and Synchronization
Di erentially Support Evidence Accumulation in Perceptual and Value-Based Decision
Making. Neuron. 2014;82(3):709{720.
[55] Pirrone A, Azab H, Hayden BY, Sta ord T, Marshall JAR. Evidence for the
Speed{Value Trade-o : Human and Monkey Decision Making Is Magnitude Sensitive.
Decision. 2018;5(2):129{142.
[56] Anderson BA. The attention habit: How reward learning shapes attentional selection.
Annals of the New York Academy of Sciences. 2016;1369(1):24{39.
[57] Ratcli R. A theory of memory retrieval. Psychological Review. 1978;85(2):59{108.
[58] Teodorescu AR, Usher M. Disentangling Decision Models: From Independence to
Competition. Psychological Review. 2013;120(1):1{38.
[59] Busemeyer JR, Townsend JT. Decision Field Theory: A Dynamic-Cognitive Ap-
proach to Decision Making in an Uncertain Environment. Psychological Review.
1993;100(3):432{459.
26
[60] Holmes WR, Trueblood JS, Heathcote A. A new framework for modeling decisions
about changing information: The Piecewise Linear Ballistic Accumulator model. Cog-
nitive psychology. 2016;85:1{29.
[61] Smith SM, Krajbich I. Attention and Choice across Domains. Journal of Experimental
Psychology General. 2018;147(12):1810{1826.
[62] Stoji c H, Orquin JL, Dayan P, Dolan RJ, Speekenbrink M. Uncertainty in learning,
choice, and visual xation. Proceedings of the National Academy of Sciences of the
United States of America. 2020;117(6):3291{3300.
[63] Just MA, Carpenter PA. Eye Fixations and Cognitive Processes. Cognitive Psychology.
1976;.
[64] Jang A, Sharma R, Drugowitsch J. Optimal policy for attention-modulated decisions
explains human xation behavior. BioRXiv. 2020;.
[65] Hunt LT, Rutledge RB, Malalasekera WMN, Kennerley SW, Dolan RJ. Approach-
induced biases in human information sampling. PLOS Biology. 2016;14(11):e2000638.
doi:10.1371/journal.pbio.2000638.
[66] H ebert B, Woodford M. Rational Inattention with Sequential Information Sampling.
Working Paper. 2017; p. 1{141.
[67] Hebert B, Woodford M. Rational Inattention When Decisions Take Time. Journal of
Chemical Information and Modeling. 2019;53(9):1689{1699.
[68] Itti L, Koch C. A Saliency-Based Search Mechanism for Overt and Covert Shifts of
Visual Attention. Vision Research. 2000;40(10):1489{1506.
[69] Roe RM, Busemeyer JR, Townsend JT. Multialternative decision eld theory: A dy-
namic connectionist model of decision making. Psychological Review. 2001;108(2):370{
392.
[70] Noguchi T, Stewart N. Multialternative decision by sampling: A model of decision
making constrained by process data. Psychological Review. 2018;125(4):512{544.
[71] Russo JE, Dosher BA. Strategies for Multiattribute Binary Choice. Journal of Exper-
imental Psychology Learning, Memory, and Cognition. 1983;9(4):676{696.
[72] Trueblood JS, Brown SD, Heathcote A. The Multiattribute Linear Ballistic Accu-
mulator Model of Context E ects in Multialternative Choice. Psychological Review.
2014;121(2):179{205.
[73] Berkowitsch NAJ, Scheibehenne B, Rieskamp J. Rigorously testing multialternative de-
cision eld theory against random utility models. Journal of Experimental Psychology:
General. 2014;143(3):1331{1348.
[74] Fisher G. An attentional drift di usion model over binary-attribute choice. Cognition.
2017;168:34{45.
[75] Krajbich I, Lu D, Camerer C, Rangel A. The Attentional Drift-Di usion Model Extends
to Simple Purchasing Decisions. Frontiers in Psychology. 2012;3.
[76] Westbrook A, van den Bosch R, M a att a JI, Hofmans L, Papadopetraki D, Cools R, et al.
Dopamine promotes cognitive e ort by biasing the bene ts versus costs of cognitive
work. Science. 2020;367(6484):1362{1366.
[77] Shi SW, Wedel M, Pieters F. Information acquisition during online decision mak-
ing: A model-based exploration using eye-tracking data. Management Science.
2013;59(5):1009{1026.
27
[78] Manohar SG, Husain M. Attention as Foraging for Information and Value. Frontiers
in Human Neuroscience. 2013;7(November):1{16.
[79] Gabaix X, Laibson D, Moloche G, Weinberg S. Costly Information Acquisition: Exper-
imental Analysis of a Boundedly Rational Model. American Economic Review. 2006;96
(4)(4):1043{1068.
[80] Yang L, Toubia O, De Jong MG. A bounded rationality model of information search and
choice in preference measurement. Journal of Marketing Research. 2015;52(2):166{183.
[81] Bezanson J, Edelman A, Karpinski S, Shah VB. Julia: A fresh approach to numerical
computing. SIAM review. 2017;59(1):65{98.
[82] Sutton RS, Barto AG. Reinforcement learning: An introduction. MIT press; 2018.
[83] Callaway F, Lieder F, Das P, Gul S, Krueger PM, Griths TL. A resource-rational
analysis of human planning. In: Proceedings of the 40th Annual Conference of the
Cognitive Science Society. Austin, TX: Cognitive Science Society; 2018.
[84] Howard RA. Information value theory. IEEE Transactions on systems science and
cybernetics. 1966;2(1):22{26.
[85] Auer P, Cesa-Bianchi N, Fischer P. Finite-time analysis of the multiarmed bandit
problem. Machine learning. 2002;47(2-3):235{256.
[86] Sobol' IM. On the distribution of points in a cube and the approximate evaluation of
integrals. Zhurnal Vychislitel'noi Matematiki i Matematicheskoi Fiziki. 1967;7(4):784{
802.
[87] Turner BM, Sederberg PB. A Generalized, Likelihood-Free Method for Posterior Esti-
mation. Psychonomic Bulletin and Review. 2014;21(2):227{250.
[88] van Opheusden B, Acerbi L, Ma WJ. Unbiased and Ecient Log-Likelihood Estimation
with Inverse Binomial Sampling. arXiv:200103985 [cs, q-bio, stat]. 2020;.
[89] Sunn aker M, Busetto AG, Numminen E, Corander J, Foll M, Dessimoz C. Approximate
Bayesian Computation. PLOS Computational Biology. 2013;9(1):e1002803.
[90] Csill ery K, Blum MGB, Gaggiotti OE, Fran cois O. Approximate Bayesian Computation
(ABC) in Practice. Trends in Ecology & Evolution. 2010;25(7):410{418.
[91] Bergstra J, Bengio Y. Random search for hyper-parameter optimization. Journal of
machine learning research. 2012;13(Feb):281{305.
[92] Goodrich B, Gabry J, Ali I, Brilleman S. rstanarm: Bayesian applied regression mod-
eling via Stan.; 2020. Available from: https://mc-stan.org/rstanarm .
[93] Yellott Jr JI. The relationship between Luce's choice axiom, Thurstone's theory of com-
parative judgment, and the double exponential distribution. Journal of Mathematical
Psychology. 1977;15(2):109{144.
28
Supporting Information
Binary choice: Task description
This data was initially reported by Krajbich et al. in 2010. For the convenience of the reader,
we include the task description from the original paper [9]. The experiment consisted of
39 Caltech students. Only subjects who self-reported regularly eating the snack foods (for
example, potato chips and candy bars) used in the experiment and not being on a diet
were allowed to participate. These steps were taken to ensure that the food items we used
would be motivationally relevant. This would not have been the case if the subjects did
not like junk food. Subjects were asked to refrain from eating for 3 h before the start of
the experiment. After the experiment they were required to stay in the room with the
experimenter for 30 min while eating the food item that they chose in a randomly selected
trial (see below). Subjects were not allowed to eat anything else during this time.
In an initial rating phase subjects entered liking ratings for 70 di erent foods using an
on-screen slider bar (how much would you like to eat this at the end of the experiment?,
scale -10 to 10). The initial location of the slider was randomized to reduce anchoring
e ects. This rating screen had a free response time. The food was kept in the room with
the subjects during the experimental session to assure them that all the items were available.
Furthermore, subjects brie
y saw all the items at this point so that they could e ectively
use the rating scale.
In the choice phase, subjects made their choices by pressing the left or right arrow keys
on the keyboard. The choice screen had a free response time. Food items that received a
negative rating in the rating phase of the experiment were excluded from the choice phase.
The items shown in each trial were chosen pseudo-randomly according to the following rules:
(i) no item was used in more than 6 trials; (ii) the di erence in liking ratings between the
two items was constrained to be 5 or less; (iii) if at some point in the experiment (i) and
(ii) could no longer both be satis ed, then the di erence in allowable liking ratings was
expanded to 7, but these trials occurred for only 5 subjects and so were discarded from the
analyses. The spatial location of the items was randomized. After subjects indicated their
choice, a yellow box was drawn around the chosen item (with the other item still on-screen)
and displayed for 1 s, followed by a xation screen before the beginning of the next trial.
Subjects xation patterns were recorded at 50 Hz using a Tobii desktop-mounted eye-
tracker. Before each choice trial, subjects were required to maintain a xation at the center
of the screen for 2 s before the items would appear, ensuring that subjects began every
choice xating on the same location.
Trinary choice: Task description
This data was initially reported by Krajbich and Rangel in 2011. For the convenience of
the reader, we include the task description from the original paper [10].
Thirty Caltech students participated in the experiment. The screening, pre-experimental
instructions, eye-tracking and liking rating phase were identical to those used in the binary
choice task described in the previous section.
In the choice phase, subjects made their choices using the keyboard. The choice screen
had a free response time. The items shown in each trial were randomly chosen. In all trials
the three items were displayed in a triangular formation with the left and right items at
the same vertical position, and the center item at the opposite vertical position. In half of
the trials the center item was on the top half of the screen, and in the other half it was on
the bottom half of the screen. Subjects indicated their choice by pressing the left, down, or
right arrow keys for the left, center, and right items, respectively. After subjects indicated
their choice, a yellow box was drawn around the chosen item (with the other item still on
the screen) and displayed for 1 s, followed by a xation screen, before the beginning of the
next trial.
1
Individual ts
We have focused on group-level ts because we are especially interested in the ability of the
model to predict di erences between binary and trinary decisions. However, it is important
to verify that the qualitative e ects that we emphasize also hold in individual data, and
are not aggregation artifacts. It is also interesting to see to what extent the model can
account for individual variability in xation and choice behavior. To address both of these
concerns, we present versions of each plot shown in the main text with separate panels for
each participant. The model was t to each participant's data following the same tting
procedure as for the group-level t (using the same precomputed likelihood histograms).
Finally, because many of the behavioral patterns are quite noisy with only 50 trials, we
aditionally plot Bayesian linear model ts for both the human and model-simulated data
(using logistic regression for binary dependent variables). These predictions were generated
using the rstanarm package [92]. The plots are included in S2 Appendix.1
In brief, we found that most behavioral patterns shown in the main text gures were
consistently demonstrated by a majority of participants. However, although most e ects
were consistently present and in the correct direction, the strength often varied considerably
across individuals. In many cases, the model showed only a modest ability to capture this
variability. This re
ects the strong a priori assumptions of the model, in particular, the
assumption that attention is allocated optimally.
Parameter recovery
To validate our model tting approach, we conducted a parameter recovery exercise. We
began by sampling 1024 \true" parameter con gurations from the promising region of the
parameter space that we considered when tting human data (see main text Methods ). We
sampled these values using the 5-dimensional Sobol sequence [86] to ensure good coverage
of the space. For each parameter con guration, we computed two sets of 80 near-optimal
policies (one for binary choice and one for trinary choice) using the UCB-based method
described in the main text. Then, for each set, we simulated the even trials of the corre-
sponding dataset. We simulated each trial only once (to match the amount of data when
tting participants), cycling between the 80 near-optimal policies. We then applied the full
approximate maximum likelihood estimation procedure described in the main text for each
dataset.2For each con guration, the maximum likelihood estimate of each parameter was
its mean in the 30 con gurations with highest likelihood (following our reporting approach
for the ts to human data).
The results, shown in Figure S1, suggest that we were able to recover parameters with
fairly high accuracy. For all parameters besides the softmax temperature, the Pearson
correlation was over 0.9. Importantly, we found only slight bias in the estimation procedure,
with the best tting linear regression line falling close to the equality line for all parameters.
The largest bias was for the prior bias parameter, , for which the recovered parameter was
on average .095 less than the true parameter.
To validate our approach when tting individual subjects, we repeated the steps above,
except using only 50 simulated trials (the number of tting trials for each subject). Unsur-
prisingly, we nd that the estimates become less reliable; however the correlations are still
fairly strong. In the trinary case, we see substantial bias for both
sample and . Thus, care
must be taken when interpreting the individual tting results.
Implementation and validation of the aDDM
In order to compare our model to the predictions of the aDDM [9, 10], we reimplemented
it based on code provided from the rst author. We made one change to the simulation
1https://fredcallaway.com/pdfs/callaway-fixation-patterns-individual-fits.pdf
2We reused the likelihood histograms that we computed when tting participant data. Critically, however,
the policies used to generate these histograms were not the same ones used to generate the simulated data.
2
100 500
100500
r=0.762Full combined dataset
1 5
x
15x
r=0.991
0.001 0.01
sample
0.0010.01sample
r=0.978
0.003 0.03
switch
0.0030.03switch
r=0.974
0 1
01
r=0.929
100 500
100500
r=0.590Individual binary dataset
1 5
x
15x
r=0.929
0.001 0.01
sample
0.0010.01sample
r=0.894
0.003 0.03
switch
0.0030.03switch
r=0.906
0 1
01
r=0.821
100 500
100500
r=0.493Individual trinary dataset
1 5
x
15x
r=0.927
0.001 0.01
sample
0.0010.01sample
r=0.805
0.003 0.03
switch
0.0030.03switch
r=0.866
0 1
01
r=0.717Figure S1: Parameter recovery. Each panel plots the estimated parameter value as a function of the true parameter value.
Each black dot corresponds to one simulated dataset. The dotted red line shows equality (i.e., perfect recovery) and the
solid blue line shows the linear trend. The top row shows results when simulating the full joint dataset. The middle row
shows results when simulating 50 trials (the amount of tting data one individual produces) of binary choice. The bottom
row shows the same for trinary choice.
procedure. In the original papers, the model predictions were generated by simulating an
equal number of trials for all possible combinations of item ratings. In contrast, we have
simulated each trial in the dataset a xed number of times. That is, our simulations follow
the empirical distribution of the item ratings. To verify the correctness of our implementa-
tion, we have replicated four key plots from the original binary and trinary papers, shown
in Figures S2 and S2 respectively. Note that for these plots, we use the original approach
of simulating each possible combination a xed number of times.
3
0.0 0.2 0.4 0.6 0.8 1.0
final time advantage left [ms]p(left chosen)
-600 -400 -200 0 200 400 600
p=0.0002
0.0 0.2 0.4 0.6 0.8 1.0
first fixation duration [ms]p(first seen chosen)
0 200 400 600
p=1e-065b
5d4bOriginal Implementation Replication
-4 -2 0 2 40.0 0.2 0.4 0.6 0.8 1.0
last seen item rating - other item ratingp(last fixation to chosen)
4c
-800 -600 -400 -200 0
last fixation duration [ms]time advantage [ms]
0 400 800 1200
0 400 800 1200Figure S2: Replication of Krajbich, Armel, and Rangel (2010). Note that x axis labels in
the orginal plots sometimes re
ected the left tail of the bin; in these cases, we adjusted the
tick locations accordingly.
4
Original Implementation Replication
3a
3b
3c
3dFigure S3: Replication of Krajbich and Rangel (2011). Note that there are slight deviations
in model predictions due to noise in the simulations; the orginal plots are based on 2000
simulated trials.
5
Derivation: Myopic Value of Information
The myopic value of information is the value of the information acquired by a single compu-
tation, that is, the expected increase in decision quality from executing a single computation
and then deciding rather than making a decision immediately. Formally,
VOI myopic (bt;c) = Ebt+1jbt;c[R(bt+1;?)] R(b;?):
In our model, this is equal to the expected value of the item that will be chosen after taking
an additional sample minus the expected value of an item chosen based on the current
beliefs. That is,
VOI myopic (bt;c) = Et+1jt;th
max
i(i)
t+1i
max
i(i)
t:
Becauset+1di ers from tonly for item c, we can rewrite the expectation term as
E(c)
t+1j(c)
t;(c)
t
max
(c)
t+1;max
i6=c(i)
t
: (9)
Thus, the term inside the expectation is the maximum of a constant, max i6=c(i)
t, and a uni-
variate random variable, (c)
t+1j(c)
t;(c)
t. To simplify notation, we suppress the conditioning
variables in the following derivation.
To derive an analytic expression for Equation [9], we rst derive the distribution of
(c)
t+1, that is, the distribution over the posterior mean after taking a sample. Applying the
transition dynamics given in Equation 2, we have
(c)
t+1= 2
xxt+(c)
t(c)
t
(c)
t+ 2x: (10)
Sincextju(c)Gaussian(u(c);2
x) and(c)
t+1is a linear transformation of xt, it follows
that(c)
t+1is a Gaussian random variable. Additionally, because the belief is a distribution
1 25 50 75 100
Number of Computations0.00.20.40.60.81.0Value of Chosen ItemVOIfull
VOImyopic
Figure S4: Illustration of the value of information features. The solid line shows the average
value of the item chosen after di erent numbers of computations selected by a near-optimal
policy assuming no computational costs. The dashed lines show values for two of the VOI
features in the initial belief state: VOI myopic is the value after one computation and VOI full
is the asymptotitc value after in nite computations.
6
over the true utility, we have u(c)j(c)
t;(c)
tGaussian((c)
t;1=(c)
t). Combining these two
statements, we see that xtj(c)
t;(c)
tis a Gaussian whose mean is itself a Gaussian. Applying
the fact that Gaussian( ;2) =+Gaussian(0 ;2), we can then derive that xtj(c)
t;(c)
t
Gaussian((c)
t;1=(c)
t)+Gaussian(0 ;x), which reduces to Gaussian( (c)
t;1=(c)
t+2
x). Finally,
applying the linear transformation of xtgiven by Equation 10, we have
(c)
t+1Gaussian(;2
)
where
= 2
x
(c)
t+1(c)
t+(c)
t(c)
t
(c)
t+1=(c)
t
and
2
=
2
x
(c)
t+1!2
1
(c)
t+2
x!
:
Having derived the distribution of Equation (c)
t+1, we now turn to the expected maximum
in [9]. From basic probability theory we know that for any constant zand random variable
X,
E[maxfX;zg] = Pr[Xz]z+ (1 Pr[Xz])E[XjX >z ]: (11)
Substituting max i6=c(i)
tforzand(c)
t+1forX, we can use this formula to derive an analytical
solution for the myopic value of information. First, we have
Pr
(c)
t+1max
i6=c(i)
t
= ( );
where is the cumulative density function (CDF) of a standard Gaussian , and
=maxi6=c(i)
t :
Next, we apply the standard formula for the expectation of a truncated Gaussian, giving us
E
(c)
t+1j(c)
t+1>max
i6=c(i)
t
=+ ( )
1 ( );
where is the standard normal probability density function. Finally, putting this together
we nd that VOI myopic (b;c) is equal to
( )max
i6=c(i)
t+ (1 ( ))
+ ( )
1 ( )
max
i(i)
t:
Derivation: Value of Perfect Information About One Item
Whereas VOI myopic captures the information value of a single sample, VOI itemcaptures the
information value of an in nite number of samples for one item, that is, the value of knowing
the exact value of one item. Formally,
VOI item(bt;c) = Eu(c)j(c)
t;(c)
t
max
u(c);max
i6=c(i)
t
max
i(i)
t:
The derivation is similar to that of VOI myopic , but instead of taking the expectation over the
posterior mean after one computation, (c)
t+1, we take the expectation over the true utility,
7
u(c)j(c)
t;(c)
tGaussian((c)
t;1=(c)
t). Thus, we apply the same steps beginning with [11],
but replacing (c)
t+1withu(c)j(c)
t;(c)
t. This results in VOI item(b;c) equal to
( 0)max
i6=c(i)
t+ (1 ( 0))
(c)
t+ ( 0)
1 ( 0)p
1=(c)
t
max
i(i)
t
where
0=maxi6=c(i)
t (c)
tp
1=(c)
t:
Derivation: Value of Perfect Information About All Items
VOI fullcaptures the information value of learning the exact value of every item in the choice
set, that is acquiring full information. In this case, the DM will make an exactly optimal
choice, gaining the utility of the item that is in fact best. Formally,
VOI full(b) = Euj(c)
t;(c)
th
max
in
u(i)oi
max
i(i)
t: (12)
For the case of Nitems, the conditional expectation term is given by the integral
Z
:::Z"
max
iu(i)kY
i=1Gaussian(u(i);(i)
t;1=(i)
t)#
du(1):::du(N):
Unfortunately, there is no analytic solution to this integral. However, we can substantially
reduce our computational burden by reducing to a piecewise one-dimensional integral. First,
we can express the expectation of any random variable as a piecewise integral,
E[X] = Z0
1FX(x)dx+Z1
0(1 FX(x))dx; (13)
whereFXis the CDF of X. Next, we can express the CDF of the maximum of a set of
random variables as the product of the CDF for each variable alone,
FmaxX(x) =Y
X2XFX(x); (14)
because the maximum of a set is less than xif and only if each element in the set is less
thanx. In our case, the set contains the belief distributions for each item. Letting M
denote maxX, we can de ne its CDF as
FM(m) =NY
i=1q
(i)
t
m (i)
t
:
Combining Equations [12], [13], and [14], we arrive at the following expression for VOI full(b):
Z0
1FM(x)dx+Z1
0(1 FM(x))dx max
i(i)
t:
We evaluate these two integrals numerically to a minimum precision of 10 5by the
adaptive Gauss-Kronrod quadrature method implemented in the QuadGK Julia package.
Despite the dimensionality reduction, we found that evaluating these integrals was still
the primary computational bottleneck for simulating the model. Thus, in order to reduce
computation time, we only compute VOI fullwhen it is necessary to determine which com-
putation the policy will execute. As detailed below, this is often unnecessary because the
other features already determine which feature has maximal [VOC.
8
Critically, the modi cation that we describe here has no e ect on the behavior of the
policy or the predictions of the models; we have veri ed this assertion through simulation.
This computational trick is based on three insights. First, note that VOI fullhelps to
decide whether or not to take another sample, but not which item to sample from. Thus,
we can determine which computation the policy would take, conditional on taking a sample
at all, based only on the VOI myopic and VOI itemfeatures. Given that these two features have
an analytical solution, as derived above, we can quickly identity the best item to sample
from, which is given by
c= arg max
c6=?
w1VOI myopic (b;c) +w2VOI item(b;c) cost(c) +w4
: (15)
Second, since VOC( b;?) = 0, it follows that if [VOC(b;c)>0, the policy should sample
from itemc, and otherwise it should stop sampling. In general, determining the sign of
[VOC(b;c)requires evaluating VOI full(b). However, in some cases the sign can be deter-
mined without knowing VOI full(b). In particular, we can take advantage of the fact that
VOI item(b;c)VOI full(b) for allb,c. We can thus compute a lower bound on [VOC(b;c)
by replacing VOI full(b) with VOI item(b;c) in Equation [15]. If this lower bound is positive,
then we know the full approximation would also be positive, and thus the optimal choice
is to sample from item c. Otherwise, we compute VOI full(b) and identify the optimal
computation using all of the features.
Third, at rst sight this approach might seem to be incompatible with the soft-maximizing
policy, where computation cis selected with probability proportional to exp [VOC(b;c). In
particular, the standard method for sampling from this distribution requires fully evaluating
[VOC(b;c). However, we can circumvent this issue using the Gumbel-max trick [93], which
provides a way to sample from a Boltzman (softmax) distribution by taking the argmax of
the unexponentiated values corrupted by Gumbel noise. Formally,
Prh
arg max
ifxi+ig=ji
=expxjP
iexpxi:
As a result, we can rewrite the soft-max policy as
(b;w; ) = arg max
cn
\VOC (b;c;w) +co
;
wherecGumbel(0;1). We can then implement steps 1 and 2 of the short-cut, adding cto
the right hand side of Equation 15, and comparing the lower-bound VOC to an independent
Gumbel sample, ?, rather than 0 to capture the noise applied to VOC( b;?).
Quality of the approximation method in Bernoulli model
The approximation method used here has previously been shown to learn policies with
near-optimal performance on a metalevel MDP similar to the one in the present model, but
with Bernoulli-distributed samples and noswitching costs [49]. The logic of the problem is
identical: A DM wants to select the best item and informs her decision by drawing noisy
samples with an expected value equal to the items' true utility. However, in the simpler
Bernoulli case that has been previously studied, true utilities take values between 0 and
1, samples from item care drawn from Bernoulli( u(c)), and the uniform distribution over
all possible utilities, Beta(1 ;1), provides a conjugate prior. Thus, posterior beliefs take the
form Beta(1 + a;1 +b), whereaandbare respectively the number of times 1 and 0 have
been sampled for the given item. Critically, the resulting belief space is discrete because
aandbare integers. This allows the computation of the exact optimal policy by dynamic
programming, if an upper bound on the number of samples that can be taken is assumed.
Callaway et al. [49] take advantage of this fact to show that the policy approximation
method used here provides a highly-accurate approximation of the optimal policy. How-
ever, their model does not have switching costs, which could potentially make the approx-
imation perform much worse. Here, we investigate this issue by adding switching costs to
9
the Bernoulli model, and measuring their impact on the method's performance. Ideally
we would be able to directly assess the performance of our method in the full model with
Gaussian samples, but an optimal solution for this case is not available and deriving one is
beyond the scope of the study.
To aid interpretation, we re-parameterized the switching cost as
switch = (k 1)
sample
such thatkcan be interpreted as a multiplier on the base sample cost. For example, k= 1,
indicated by \1x" in the gure, corresponds to no switching cost. We considered a grid of cost
parameters with
sample2fe 9;e 8;:::;e 3gandk21;2:::;10. We set an upper bound of
75 samples. As shown in Fig. S5, we replicated previous results that the approximated policy
is nearly optimal when there is no switch cost. As the gure shows, relative performance
degrades somewhat when switch costs are added, but the approximation still achieves 92%
of the optimal metalevel reward in the worst case explored.
0.50.60.7Metalevel Return
1x Switch Cost
2x Switch Cost
3x Switch Cost
4x Switch Cost
5x Switch Cost
104
102
Sample Cost0.50.60.7Metalevel Return
6x Switch Cost
104
102
Sample Cost
7x Switch Cost
104
102
Sample Cost
8x Switch Cost
104
102
Sample Cost
9x Switch Cost
104
102
Sample Cost
10x Switch Cost
Figure S5: Performance of the policy approximation (red) compared with the true optimal solution (blue) on the Bernoulli
model with switching costs. The red line shows mean performance from the top 80 policies identi ed by the UCB algorithm.
Additionally each individual policy's performance is plotted as an individual point, but performance is so consistent that
the points are not visually distinct.
10 |
9ac90401-989c-4db3-9135-24db7aefeeb7 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Ontology-based Queries over Cancer Data
1 Introduction
---------------
In the biomedical sciences, the use, exchange and integration of the ever-increasing amount of data has become paramount to accelerate the discovery of new approaches for the detection, diagnosis, treatment and prevention of diseases. In particular, this applies to cancer, for which the US National Cancer Institute (NCI) and the UK National Cancer Research Institute (NCRI) have implemented the caBIG®111caBIG® stands for cancer Biomedical Informatics Grid® programme and the NCRI Informatics Initiative, looking at building and deploying software infrastructure to manage and analyse data generated from heterogenous data sources.
In this paper, we provide an analysis of the caGrid[saltz:Bioinformatics2006] software infrastructure developed within the NCI caBIG® programme and extend it with richer querying capabilities. caGrid supports a collaborative information network for sharing cancer research data, and deals with syntactic and semantic interoperability of the data resources in a service-oriented model-driven architecture. Semantic interoperability is achieved by using a metadata registry, which maintains information models annotated with concepts from a domain ontology: the NCI thesaurus (NCIt)[hartel:JBI2005]. However, the query functionality provided in caGrid does not take into account the semantic annotations, but it only relies on each individual information model.
Our methodology is based on extending the caGrid service-oriented model-driven infrastructure with additional services to support ontology-based queries over the distributed data resources. In this way, the biomedical researchers, as the end-users of our system, will be able to query cancer data by building queries using their domain knowledge (expressed as concepts from the NCIt ontology) rather than having to know the underlying models. This also means that the queries are reusable across resources, which is not the case in the caGrid infrastructure. This functionality will be incorporated into the NCRI ONcology Information eXchange (ONIX222<http://www.ncri-onix.org.uk/>). Our approach involves a customised transformation from annotated information models to an ontological representation using the Web Ontology Language version 2 (OWL333OWL is a recommendation from the World Wide Web Consortium (W3C) and the language overview for its second version can be found at <http://www.w3.org/TR/owl2-overview/>). This representation supports annotations based on a primary concept and a list of qualifiers. Based on these ontological representations of the data resources, we have designed and developed a query rewriting and translation approach that converts concept-based queries into the query language supported by the caGrid infrastructure. This approach is general and could be used to support other target query languages, as the only step dependent on caGrid is the last one. This work presents significant improvements over our previous work[gonzalez-beltran:CBMS2009], as we have significantly modified and improved the OWL representation and the design and implementation of the query rewriting and translation steps. We have developed a caGrid analytical service for the transformation from an annotated information model to OWL. Additionally, we present an analysis of the caGrid query language and information together with an extensive performance evaluation that justifies the applicability of our solution.
This paper is structured as follows. Section [2](#S2 "2 Background ‣ Ontology-based Queries over Cancer Data") introduces background material on the caGrid infrastructure. Section [3](#S3 "3 Analysis of the caGrid Query Language ‣ Ontology-based Queries over Cancer Data") presents an analysis of the caGrid query functionality and the type of queries supported by its query language. Then, we present in section [4.1](#S4.SS1 "4.1 OWL Representation of caGrid Information Models ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data") the OWL representation that is used for query rewriting and translation, which in turn is described in Section [4.2](#S4.SS2 "4.2 Query Rewriting and Translation ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data"). The implementation details and performance evaluation results are given in Sections [4.3](#S4.SS3 "4.3 Implementation ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data") and [4.4](#S4.SS4 "4.4 Performance Evaluation ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data"), respectively. The evaluation includes an analysis of the generated ontologies as well as several performance metrics for OWL generation and query rewriting, which justify the viability of our approach. After comparing our approach with related work in Section [5](#S5 "5 Related Work ‣ Ontology-based Queries over Cancer Data"), we conclude the paper in Section [6](#S6 "6 Conclusions ‣ Ontology-based Queries over Cancer Data"), including considerations for future work.
2 Background
-------------
caBIG®[saltz:Bioinformatics2006] is an NCI programme whose aim is to create a virtual and federated informatics infrastructure for sharing data, tools and connect scientists and organisations in the cancer research community. The computing middleware in caBIG® is called caGrid, which is a Grid[foster:anatomy01] extended to support data modelling and semantics[saltz:Bioinformatics2006]. caGrid has a number of core services and corresponding application programming interfaces (APIs), which we will introduce next, by analogy with the metadata hierarchy[pollock:adaptive04], as per Figure [1](#S2.F1 "Figure 1 ‣ 2 Background ‣ Ontology-based Queries over Cancer Data").
Figure 1: caGrid semantic infrastructure.
The metadata hierarchy represents how the semantics of raw data (instance data) can be augmented by overlaying metadata of increasing descriptiveness [pollock:adaptive04]. The syntactic metadata refers to the language format and data types, and in caGrid is represented by XML schemas managed by the Global Model Exchange (GME)[saltz:Bioinformatics2006] service, which exposes them through the GME API444<http://cagrid.org/display/gme/>. The structural metadata gives form to the units of data. In caGrid, it is implemented as an object-oriented virtualisation of the underlying data resources[saltz:Bioinformatics2006] and it is represented as UML555UML stands for the Unified Modeling Language, a specification of the Object Management Group®(OMG®) models. These UML models can be accessed through the Discovery API666<http://cagrid.org/display/metadata13/Discovery>. The purpose of the referent metadata is to represent the linkages between the different data models. In caGrid, the linkages are provided by a metadata registry, called caDSR777caDSR stands for cancer Data Standards Repository, based on the ISO/IEC 11179 standard888<http://metadata-stds.org/11179/>. caDSR manages common data elements (CDEs) and exposes them through the caDSR API. The domain metadata represents what the data is about. It is implemented by a domain conceptualisation, usually in the form of an ontology[pollock:adaptive04]. In the caGrid case, the NCIt ontology[hartel:JBI2005] is used, accessed via the LexEVS API999<https://cabig.nci.nih.gov/concepts/EVS/>. Finally, the rules constitute an overarching layer that can be applied to all the aforementioned layers. Rules can be used to constrain and extend the semantics of metadata specifications at any of the abstraction levels[pollock:adaptive04]. In the current caGrid semantic infrastructure, there is no equivalent to the rule metadata.
A data resource owner can share the data by developing caGrid data services using common interfaces and metadata, as described above. In this way, a data service encapsulates the data, which is kept in native formats (including, for example, relational data or flat files), exposing an access interface based on the object-oriented (UML) model of the underlying resource. The common interface also exposes a query processor based on the Common Query Language (CQL) defined for caGrid. CQL is an object-oriented query language reflecting the underlying object model of the data resource while abstracting the physical representation of the data[saltz:Bioinformatics2006]. At the time of writing, there exist two versions of CQL and there is a pre-release version of the latest one101010<http://cagrid.org/display/dataservices/CQL+2>. More details on CQL are given in Section [3](#S3 "3 Analysis of the caGrid Query Language ‣ Ontology-based Queries over Cancer Data").
caGrid also supports basic distributed aggregations and joins of queries over multiple data services by means of the caGrid Federated Query Infrastructure111111<http://cagrid.org/display/fqp/Home>, through a distributed extension of CQL called DCQL. Thus, caGrid relies on D/CQL – custom query languages based on the structural characteristics of the resources. In other words, caGrid builds a ’structural layer’, where queries are expressed in terms of objects, attributes and associated objects, without allowing for semantic queries. D/CQL are evolving to provide richer structural queries as new requirements arise from different caBIG® projects. However, these query languages do not allow for data extraction based on semantic information. Thus, a shortcoming of caGrid is that does not currently exploit the referent and domain metadata maintained for its data services. Additionally, as already mentioned, it is not possible to specify rules about the models nor the domain semantics.
As stated in the introduction, this work advocates the extension of the caGrid infrastructure to exploit its rich metadata by building a semantic layer, using semantic web technologies to exploit caGrid’s metadata. Additionally, this extension is capable of:
a) accommodating other resources with different ways of dealing with metadata, and
b) specifying rules at different levels of abstraction.
3 Analysis of the caGrid Query Language
----------------------------------------
A CQL query is defined by an XML document, which must comply to a specified XML schema121212The CQL XML schema is available at: <http://cagrid.org/display/dataservices/CQL+Schemas>. The schema indicates that a CQL query must specify a ⟨Target⟩ element, which is the data type of the query result. Optionally, an ⟨Attribute⟩ element might indicate a predicate over an attribute of the object with ⟨Target⟩ type and an ⟨Association⟩ may specify a link with a related object. In Table [1](#S3.T1 "Table 1 ‣ 3 Analysis of the caGrid Query Language ‣ Ontology-based Queries over Cancer Data") we show how a CQL query is built recursively presenting it as a context-free grammar, where ⟨CQLQuery⟩ is the start symbol, ϵ is the empty string and ⟨xsd:string⟩ is the non-terminal variable representing the *X*SD:string data type.
| |
| --- |
|
⟨CQLQuery⟩ →
⟨Target⟩ |
⟨Target⟩ ⟨QueryModifier⟩
⟨Target⟩ →
⟨Name⟩ ⟨Attribute⟩ |
⟨Name⟩ ⟨Association⟩ |
⟨Name⟩ ⟨Group⟩
⟨Attribute⟩ →⟨Name⟩ ⟨Predicate⟩ ⟨Value⟩
⟨Group⟩ →
⟨LogicalOp⟩ ⟨Attribute⟩ ⟨Group1⟩ |
⟨LogicalOp⟩ ⟨Association⟩ ⟨Group1⟩
⟨Group1⟩ →
⟨Attribute⟩ ⟨Group1⟩ |
⟨Association⟩ ⟨Group1⟩ |
⟨Group⟩ |ϵ
⟨Name⟩ →⟨xsd:string⟩
⟨RoleName⟩ →⟨xsd:string⟩
|
| |
| --- |
|
⟨LogicalOp⟩ →AND |OR
⟨Predicate⟩ →
EQUAL\_TO |NOT\_EQUAL\_TO |
LIKE |IS\_NULL |
IS\_NOT\_NULL |LESS\_THAN |
LESS\_THAN\_EQUAL\_TO |
GREATER\_THAN |
GREATER\_THAN\_EQUAL\_TO
⟨Association⟩ →
⟨RoleName⟩ |
⟨RoleName⟩ ⟨Association⟩ |
⟨RoleName⟩ ⟨Attribute⟩ |
⟨RoleName⟩ ⟨Group⟩
⟨Value⟩ →⟨xsd:string⟩
⟨QueryModifier⟩ →
⟨DistinctAttribute⟩ |
⟨DistinctAttribute⟩ ⟨AttributeNames⟩
|
Table 1: CQL query context-free grammar
So, CQL traverses the UML class diagram graph, where the ⟨Target⟩ is the initial class, the ⟨Association⟩ conditions allow for path navigation by traversing sequences of consecutive classes and ⟨Attribute⟩ conditions apply locally to individual classes. The terminal symbols ⟨Group⟩ and ⟨Group1⟩ represent the combination of two or more constraints over a particular node in the UML class graph.
4 Ontology-based queries over the caGrid infrastructure
--------------------------------------------------------
As shown before, the caGrid queries rely on the structure of the underlying data resources, i.e. their UML models. Thus, a biomedical researcher interested in retrieving data about, for example, a particular gene of interest will need to explore the UML model of each relevant data service and build a query considering the specific attributes and associations of the class maintaining the *Gene* objects. The queries can be built programmatically or also through the caGrid portal131313<http://cagrid-portal.nci.nih.gov>, which allows to explore the UML models and provides a query builder based on these models.
In this work, we propose a system that allows the user to concentrate on the concepts from the domain, as represented by the NCIt ontology on cancer, and build the ontology-based queries which are *high-level*141414By a high-level query, we mean a query that can be written without specific details about the structure of the target resource. and *descriptive*151515By a descriptive query, we refer to queries that provide the criteria for the desired data rather than the procedure to find the data.. Thus, the ontology-based queries can be applicable to any of the underlying data resources.
Apart from the cancer concepts found on NCIt, the queries combine elements from an ontology we built with metadata on UML models161616We will see later, than the queries could also use elements from the list ontology[drummond:putting06]., namely the *UML model* ontology. This ontology contains OWL classes to represent UML classes and attributes (*UMLClass*, *UMLAttribute*), OWL object properties to represent UML associations and the relationship between a UML class and its attributes (*hasAssociation*, *hasAttribute*) and a data property to represent the values of attributes (*hasValue*).
Some simple example queries171717The example queries are given in Manchester OWL Syntax and are just intended to show queries retrieving objects from a UML class, a UML class with a condition over an attribute and two associated UML classes with a restriction over an attribute of one of the classes, respectively. are:
a) *Specimen*to retrieve all the objects that are annotated with the Specimen concept;
b) *Gene and hasAttribute some (Gene\_Symbol and hasValue value ”BRCA%”)*to find all the genes whose symbol starts with the string BRCA;
c) *Single\_Nucleotide\_Polymorphism and hasAssociation some (Gene and hasAttribute some (Gene\_Symbol and hasValue value ”TGFB1”))*to obtain all the *SNPs* associated with the *Gene* *Transforming Growth Factor Beta 1*[gonzalez-beltran:CBMS2009].
In our system, these queries could be submitted to any data service, and they will be converted to the specific CQL query.
We note that the third query specifies SNPs that are associated with genes. This association may be present in different ways in two separate UML models. For example, the two corresponding classes may have a direct UML association, or the association may arise by traversing an association path from the first class to the second one. In order for our system to deal with those paths of associations, without the user requiring to know the specific underlying UML model, we define the *hasAssociation* property as transitive and use reasoning to determine the paths.
Next, we introduce our transformation from caGrid models to an OWL2 representation and the query rewriting/translation approach, which transforms ontology-based queries into CQL queries. The OWL2 ontologies provide an unified view of the UML models and their semantic annotations, which allows us to apply reasoning over them.
###
4.1 OWL Representation of caGrid Information Models
Figure 2: Part of UML class diagram for caBIO 4.2
OWL model of UML class diagrams. First, we present our customised UML-to-OWL transformation. This transformation differs from previous approaches, as explained in Section [5](#S5 "5 Related Work ‣ Ontology-based Queries over Cancer Data"). Next, we describe the transformation and use the portion of the caBIO 4.2 information model in Figure [2](#S4.F2 "Figure 2 ‣ 4.1 OWL Representation of caGrid Information Models ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data") to give examples. Every UML element is related to its counterpart in the *UML model* ontology: all UML classes and attributes are defined as subclasses of *UMLClass* and *UMLAttribute*, respectively (see equations Eq. [1](#S4.E1 "(1) ‣ 4.1 OWL Representation of caGrid Information Models ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data") and Eq. [2](#S4.E2 "(2) ‣ 4.1 OWL Representation of caGrid Information Models ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data") below181818The prefixes used in the equations are: *c:* for the caBIO 4.2 ontology, *u:* for the UML model ontology, *n:* for the NCIt ontology and *l:* for the list ontology. We note that the name of an OWL class corresponding to an attribute includes the class name to avoid duplications and for associations, it includes its domain and range.); all the UML associations are sub-properties of *hasAssociation* (Eq. [4](#S4.E4 "(4) ‣ 4.1 OWL Representation of caGrid Information Models ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data")), and the datatype property *hasValue* is used to specify the type of the attributes (Eq. [3](#S4.E3 "(3) ‣ 4.1 OWL Representation of caGrid Information Models ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data")) as an existential restriction. Contrary to other UML-to-OWL transformations, we represent UML attributes as OWL classes. This is required so that the ontology-based queries can include the concepts associated with attributes.
| | | | | | |
| --- | --- | --- | --- | --- | --- |
| | c:Chromosome | ⊑ | u:UMLClass | | (1) |
| | c:Chromosome\_number | ⊑ | u:UMLAttribute | | (2) |
| | c:Chromosome\_number | ⊑ | ∃\small u:hasValue.\small xsd:string | | (3) |
| | c:Chromosome\_locationCollection\_Location | ⊑ | u:hasAssociation | | (4) |
UML subclass and superclass relationships are represented with subsumption (Eq. [5](#S4.E5 "(5) ‣ 4.1 OWL Representation of caGrid Information Models ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data")). For each UML class, existential restrictions are added for its associations (Eq. [6](#S4.E6 "(6) ‣ 4.1 OWL Representation of caGrid Information Models ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data")) and attributes (Eq. [7](#S4.E7 "(7) ‣ 4.1 OWL Representation of caGrid Information Models ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data")). While UML does not explicitly represent inherited associations, our OWL representation makes them explicit, modeling the semantics of UML. For example, as the UML class *Location* has an association *chromosome* with the class *Chromosome*, this association is inherited on the subclass *CytogeneticLocation* (Eq. [8](#S4.E8 "(8) ‣ 4.1 OWL Representation of caGrid Information Models ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data")).
| | | | | | |
| --- | --- | --- | --- | --- | --- |
| | c:CytogeneticLocatoin | ⊑ | c:Location | | (5) |
| | c:Chromosome | ⊑ | ∃\small c:Chromosome\\_locationCollection\\_Location. | | (6) |
| | | | c:Location | |
| | c:Chromosome | ⊑ | ∃\small u:hasAttribute.\small c:Chromosome\\_number | | (7) |
| | c:CytogeneticLocation | ⊑ | ∃\small c:Location\\_chromosome\\_Chromosome. | | (8) |
| | | | c:Chromosome | |
We note that the generated OWL ontologies belong to OWL2EL[Cuenca-Grau:JWebSemantics2008], an OWL2 profile specifically designed to allow for efficient reasoning with large terminologies, which is polynomial in the size of the ontology. While OWL2EL disallows universal quantification on properties, it does allow the inclusion of transitive properties. Thus, it is suitable for our UML-to-OWL transformation customised for the rewriting approach as outlined before.
OWL Representation of the Semantic Annotations.
Apart from representing the UML model, we also model its mapping to NCIt, as maintained in caDSR. Through the CDEs, UML elements are annotated with a primary concept, which indicates the meaning of the element. In turn, a list of qualifier concepts may be used to modify the primary concept, giving specific meaning. As OWL2 does not natively supports the representation of lists, we used Drummond et al’s design pattern for sequences[drummond:putting06] to model primary concepts and qualifier lists. The following equations give some examples on how the semantic annotations of UML classes (Eq. [9](#S4.E9 "(9) ‣ 4.1 OWL Representation of caGrid Information Models ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data")) and attributes (Eq. [10](#S4.E10 "(10) ‣ 4.1 OWL Representation of caGrid Information Models ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data")) with a single concept are modelled. Equation [4.1](#S4.Ex5 "4.1 OWL Representation of caGrid Information Models ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data") models the class *c:SNPCytogeneticLocation* as being a *n:Location* qualified with *l:Chromosome\_Band* and
*n:Single\_Nucleotide\_Polymorphism*.
| | | | | | |
| --- | --- | --- | --- | --- | --- |
| | c:Chromosome | ⊑ | n:Chromosome | | (9) |
| | c:Chromosome\_numer | ⊑ | n:Name | | (10) |
| | c:SNPCytogeneticLocation | ⊑ | \small n:Location⊓(\small l:OWLList⊓ | |
| | | | ∃\small l:hasContents.\small n:Chromosome\\_%
Band⊓ | |
| | | | ∃\small l:hasNext.(\small l:OWLList⊓ | |
| | | | ∃\small l:hasContents.\small n:Single\\_%
Nucleotide\\_Polymorphism)) | |
Module Extraction from NCI Thesaurus Ontology.
The NCIt ontology is very large, as it provides a common vocabulary for the whole cancer domain[hartel:JBI2005]. Each caGrid data service is, in general, concerned with data pertaining to more specific domains than the whole NCIt ontology. Thus, for each caGrid data service referring to a subset Σ of the NCIt vocabulary, there is a subset of terms and relationships from NCIt that is relevant, called a module from the ontology[sattler:module09]. The module M represents all knowledge about the terms of the signature Σ. One of the approaches torelevance is logic-based: the module M is relevant for the terms Σ if all the consequences of the ontology that can be expressed over Σ are also consequences of M[sattler:module09]. We follow that approach by Sattler et al [sattler:module09] and extract an NCIt module for each of the information models in caGrid. For succinctness and efficiency, this module is used, as opposed to the whole NCIt ontology, for the semantic annotations of UML models and subsequent reasoning. However, we observe that we removed the disjoint axioms from the NCIt modules, as we noted before[gonzalez-beltran:CBMS2009, McCusker:BMCBioinformatics2009] using subsumption to represent UML class to concept mapping may result in inconsistent ontologies as the annotations for a single class may come from two high-level branches in NCIt that are declared as disjoint.
###
4.2 Query Rewriting and Translation
This section describes how an ontology-based query is rewritten and then translated, first to an intermediate optimisation language and then to the target CQL language. While the overall approach is similar to our previous work[gonzalez-beltran:CBMS2009], previously we relied completely on justifications[kalyanpur:finding07] and now we have extensively improved the approach by dealing with each of the steps independently. We provide the output of each step for the third query from Section [4](#S4 "4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data").
Parsing.
First, the user query is syntactically parsed. The query uses concepts from the NCIt, the UML model (see Section [4.1](#S4.SS1 "4.1 OWL Representation of caGrid Information Models ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data")) and the list ontologies[drummond:putting06].
UML Extraction.
The NCIt concepts in the query are translated into specific UML classes, by reasoning over the generated ontologies. Each concept is the super-class of a UML class or UML attribute, depending on their position on the query. Often, a single NCIt concept will correspond to many UML classes (or attributes) and, in such cases, each UML class is returned to form an individual query. Therefore, the outcome of the UML extraction is a combination of possible queries given the extracted UML classes or attributes. The outcome for our example query is: c:SNP and hasAssociation some (c:Gene and hasAttribute some (c:Gene\_symbol and hasValue value ”TGFB1”)) .
Data Values Extraction.
As the generated ontologies do not contain instances, the semantic validation of the query, expressed as an OWL class expression, must ignore the data expressions. This step extracts the data expressions, which will be reinserted later on. This step results in c:SNP and hasAssociation some (c:Gene and hasAttribute some (c:Gene\_symbol)).
Semantic Validation.
We use a reasoner to check that the resulting query can be satisfied. If the query cannot be satisfied, subsequent rewriting of the query is halted.
Properties Path Finder.
This step deals with the ontology corresponding to the UML model (the semantic annotations do not need to be considered any longer) and aims at finding the path of UML classes related through the transitive property hasAssociation191919We note that the ontology is compliant with the OWL2 EL profile, as OWL2 EL supports the use of transitive object properties. For more information, see <http://www.w3.org/TR/owl2-profiles/>. The path finder rewrites the expression using non-transitive properties, corresponding to UML associations, by using an explanation generator[kalyanpur:finding07] that retrieves the justification for two classes to be connected via the transitive property, and thus allowing to find the intermediate classes. The path finder may find more than one path between a set of nodes and, in such cases, will return each path as a combination of possible queries for user selection. One path for our example query is: c:SNP and hasAssociation some c:GeneRelativeLocation and hasAssociation some (c:Gene and hasAttribute some (c:Gene\_symbol)).
Data Values Addition.
At this point, we can retrieve the data expressions removed earlier and re-insert them into the corresponding OWL classes, resulting in c:SNP and hasAssociation some c:GeneRelativeLocation and hasAssociation some (c:Gene and hasAttribute some (c:Gene\_symbol and hasValue value ”TGFB1”)).
OWL Expression to MCC Translation.
No calculus or algebra has been defined for the object-oriented query language CQL. To provide a translation with CQL as target language, we use the monoid comprehension calculus (MCC), as it is a formal framework to support object queries optimizations[fegaras:optimizing00]. Object queries involve collections (e.g. sets, lists, bags), whose semantics can be captured by monoid comprehensions (MC). In this paper, we only overview MCs and its use in our system202020For more details, we refer the reader to [fegaras:optimizing00] and [gonzalez-beltran:CBMS2009]. Our approach is similar to the work by Peim et al [peim:query02], but while they use an expansion algorithm to rewrite an OWL expression based on a set of acyclic set of definitions, we follow the specific steps described above.
A MC takes the form ⊕{e⫾¯¯¯q}, where ⊕ is a monoid212121A monoid of type T is an algebraic structure defined by (⊕,Z⊕) where ⊕:T×T→T is an associative funcion and Z⊕ is the left and right identity of ⊕. A collection monoid is a monoid for a collection type (e.g. lists or bags) and must also specify a unit function building a singleton collection. operator called the *accumulator*, e is the *header* and ¯¯¯q=q1,…,qn,n≥0 is a sequence of *qualifiers*. A qualifier can take the form of a *generator*, v←e′ with v a range variable and e′ an expression constructing a collection, or a *filter* predicate. The symbol \cupplus denotes the accumulator for bags222222For example, \cupplus{x⫾x←{1,2}} is the monoid comprehension representing the bag {{1,2}}. . For an OWL class expression from the previous step, an MCC expression is built such that: the header variable is determined by the first concept in the query and the qualifiers are built for each of the remaining expressions. The MCC expression for our example is:
\cupplus { s ⫾ s ← SNP, r ← s.relativeLocationCollection, r ← GeneRelativeLocation, g ← r.gene, g ← Gene, g.symbol=TGFB1 }
MMC to CQL Translation.
Translating the MCC expression into CQL amounts to: define as *Target* the type of the variable that appears in the header and then, including an *Association* per each pair of generators, one determining the name (the class to which they belong) and the other identifying the role name; include an
*Attribute* restriction for each filter. As this last step is the only one involving CQL, only this last step requires to be modified to extend our methodology to other model-driven architectures with a different target language. The resulting CQL in the example is:
[⬇](http://data:text/plain;base64,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)
<ns1:CQLQuery xmlns:ns1=”http://CQL.caBIG/1/gov.nih.nci.cagrid.CQLQuery”>
<ns1:Target name=”gov.nih.nci.cabio.domain.SNP”>
<ns1:Association name=”gov.nih.nci.cabio.domain.GeneRelativeLocation”
roleName= ”relativeLocationCollection”>
<ns1:Association name=”gov.nih.nci.cabio.domain.Gene” roleName=”gene”>
<ns1:Attribute name=”symbol” predicate=”EQUAL\_TO” value=”TGFB1”/>
</ns1:Association>
</ns1:Association>
</ns1:Target>
</ns1:CQLQuery>
###
4.3 Implementation
We have implemented two modules, with the functionalities:
a) an OWL generator, which transforms a caGrid annotated UML model into an OWL ontology and includes the generation of a module from the NCIt containing the concepts relevant to the UML model;
b) a query translation component, which takes as input a OWL class expression using concepts from the NCI thesaurus and transforms it into a CQL for a single data service.
For the first module, we also produced a caGrid analytical service called OWLGenService232323The OWLGenService is accessible through the caGrid portal at <http://cagrid-portal.nci.nih.gov> and available at <http://stylus_157.stylusinternet.net:9600/wsrf/services/cagrid/OwlgenService>, which provides a simple API for the extraction of modules from NCIt and for the ontology generation, given a specific information model.
The implementation was done in Java and uses caGrid version 1.3242424<http://wiki.cagrid.org/display/caGrid13/Home>, the OWLAPI version 3.1.0252525<http://owlapi.sourceforge.net/> (after upgrading from OWLAPI version 2), and relies on the reasoners Pellet 2.2.2262626<http://clarkparsia.com/pellet/> and HermiT 1.3.0272727<http://hermit-reasoner.com>.
###
4.4 Performance Evaluation
This section analyses the generated ontologies and presents two areas of performance evaluation that verify the viability of our approach.
Since one important step in the query rewriting/translation process is the property path finder (see Section [4.2](#S4.SS2 "4.2 Query Rewriting and Translation ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data")), we firstly introduce some metrics to assess the paths in the generated ontologies. These paths are sequences of concepts linked by object properties. Secondly, we present the generation times for the module extraction, the ontology generation and the inference of the ontologies using both the Pellet and HermiT reasoners. These results show that the generation of the ontologies that make possible our approach is done in a timely fashion. Thirdly, we evaluate the performance of the query rewriting process, showing a breakdown of the constituent parts of the rewriting algorithm. For this evaluation, we considered two sets of five queries each run over the caBIO data service282828<http://cabiogrid42.nci.nih.gov:80/wsrf/services/cagrid/CaBIO42GridSvc>, where each set consists of queries that involve paths of lengths one and two. The tests were run on a Red Hat Enterprise Linux Server release 5.3 (Tikanga) with 64 bits and 48285 MB of RAM.
This section analyses the generated ontologies and presents two areas of performance evaluation that verify the viability of our approach.
Since one important step in the query rewriting/translation process — from Section [4.2](#S4.SS2 "4.2 Query Rewriting and Translation ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data") — is the property path finder, we firstly introduce some metrics to assess the sequences of concepts linked by object properties (paths) in the generated ontologies. Secondly, we present the generation times for the module extraction, the ontology generation and the inference of the ontologies using both the Pellet and HermiT reasoners. These results show that the generation of the ontologies that make possible our approach take a short time. Thirdly, we evaluate the performance of the query rewriting process with a breakdown of the constituent parts of the algorithm. For this evaluation, we considered two sets of five queries each, where each set consists of queries that involve paths of lengths one and two. The results were obtained by running on a Red Hat Enterprise Linux Server release 5.3 (Tikanga) with 64 bits and 48285 MB of RAM.
Throughout this section, we have grouped caGrid projects into three distinct subsets: projects that are available from the *caDSR* service, all data services that are registered with the *caGrid* default index service292929<http://cagrid-index.nci.nih.gov:8080/wsrf/services/DefaultIndexService>, and *Information Models* (or *InfoModels*) (those models that are supported by a deployed service from the *caGrid* Index Service)303030It should be noted that not all caDSR projects are included in the metrics; some contained errors (their semantic metadata is not complete or refers to an older version of the NCI thesaurus) and some models are targeted for data modelling, rather than specifically holding data, making them not representative for our system. Out of the 136 projects in caDSR, 16 were excluded from the analysis for these reasons. However, none of the excluded projects had an associated service. Additionally, the *caGrid* subset has 63 services and *InfoModels* has 23 projects.. We note that the groups *caGrid* and *InfoModels* are the more relevant for our system, as only against these projects it is possible to execute CQL queries. While *InfoModels* include a single project from caDSR for a set of deployed services corresponding to that project, *caGrid* may include the results for several services that correspond to a single model. Thus, the *caGrid* results will be skewed according to the relative weight of services as opposed to models.
Analysis of the OWL representation of the information models.
While ontology metrics have been defined in several tools [garcia:OWL10], these have focused on basic metrics (e.g. number of classes) and semantic-based metrics (e.g. relationship richness) that allow for the comparison and quality evaluation of the ontologies. Here, we will focus on the presentation of some bespoke metrics we developed to measure the proliferation and complexity of paths within the ontologies, as these will ensure the viability of our approach.
As seen in Section [4.2](#S4.SS2 "4.2 Query Rewriting and Translation ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data"), our rewriting process seeks to remove the upper-level and transitive object property *hasAssociation* and express the query using only non-transitive properties, which correspond to the UML associations in the models. In order to achieve this, we consider the paths between pairs of concepts from the query connected through the *hasAssociation* property. The calculation of these paths is not trivial; there may be many intermediate nodes and there may be more than one path for a given pair of concepts. We define a *journey* as a traversal from one concept to another. A *journey* may have one or many paths, which represent the possible routes that the traversal can take. Thus, it is important to evaluate these aspects of the ontologies in order to assess the viability of a rewriting tool.
We propose the following metrics as a measure of complexity in this respect. The *Longest Path* is the maximum path length that may be computed within a given ontology. The longest path length provides an indication of the worse case for path calculation times. The *Average Paths per Journey* reflects the degree of path expansion within the rewriting algorithm, as each journey (e.g. from Node A to Node B) may have many different paths. The rewriting algorithm should return all possible paths as each path may refer to a different expression of the query. When we consider that a single query may include multiple independent journeys, the possible query rewritings can become very large. The *Average Nodes per Path* is the average number of nodes that must be visited in order to return a single path. These metrics can affect the path calculation time as well as the complexity of the resulting query.
\includegraphics[width=108.405pt]./figures/pathLength.jpg\includegraphics[width=108.405pt]./figures/nodesPerPath.jpg\includegraphics[width=108.405pt]./figures/pathsPerJourney.jpg
Figure 3: The Path Metrics.
Figure [3](#S4.F3 "Figure 3 ‣ 4.4 Performance Evaluation ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data") illustrates three box-and-whisker plots with the results of the path metrics for each project subset. We observe that while the longest path can have up to 36 nodes, for 75 % of the projects in each category their length is less than 17 or 18. The median of the average path length varies between 4 and 7 nodes over the three subsets, and for 75 % of the *InfoModels* the average path length is less than 8. The median of the average paths is around 2 paths per journey, and for 75 % of the projects in each category the average path per journey is less than 2.5. This indicates that we will be returning a low number of path combinations as a result. These results, then, verify that the paths within the ontologies are manageable and appropriate for our rewriting tool. We also note that in all the metric diagrams, the caGrid subset is often very densely clustered around the mean. This is due to the fact that there are often many caGrid services for the same project that differ to one another very slightly or not at all, which can result in multiple similar or identical results.
Ontology Generation, Module Extraction and Classification.
In order to isolate any overhead caused by variations in network performance, we extracted the XML corresponding to each project (or information model) in caDSR. This is a preliminary step to run the performance evaluator locally, and we do not include any data about the performance of this stage. We generate four ontologies for each project: the NCIt module ontology (incorporating the concepts from NCIt relevant to the project), the annotated UML ontology (including the classes describing the UML model) and inferred versions of the two ontologies313131We generate the inferred ontologies classifying the generated ontologies using both the HermiT and Pellet reasoners.. We recorded the time for each generation and Figure [4](#S4.F4 "Figure 4 ‣ 4.4 Performance Evaluation ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data") shows them for the four ontologies per project grouped by subset. The times are presented in a logarithmic scale. We can see that the vast majority (75%) of NCIt modules take less than 2 seconds to generate and even less time for ontology generation. The classification of the generated ontologies is also timely, with the average inference of the Pellet and HermiT reasoners never taking longer than 100 milliseconds.
\includegraphics[width=144.54pt]./figures/comboGen.jpg\includegraphics[width=144.54pt]./figures/comboInference.jpg
Figure 4: The generation and inference times.
Query Rewriting Evaluation.
We have developed a suite of queries of varying complexity in order to evaluate the query rewriting. The results are presented in figure [5](#S4.F5 "Figure 5 ‣ 4.4 Performance Evaluation ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data"), which shows the average time323232Each query was ran 5 times and the average times calculated. taken at each stage of the query rewriting process333333These correspond to the stages of query rewriting; parsing, UML extraction, validation, path finding, MCC conversion and CQL conversion. For more information, refer to section [4.2](#S4.SS2 "4.2 Query Rewriting and Translation ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data"). for five queries whose rewriting has path length one, five queries whose rewriting has path length two and the mean times for these two sets. The query’s path length refers to the number of intermediate nodes in the rewritten query. We can see from figure [5](#S4.F5 "Figure 5 ‣ 4.4 Performance Evaluation ‣ 4 Ontology-based queries over the caGrid infrastructure ‣ Ontology-based Queries over Cancer Data") that, while the path length has a marked effect on the time taken at the path finding stage, the other stages of implementation remain largely unaffected. We therefore maintain that, given our analysis of paths within our target ontologies described above, we can provide query rewriting in a timely and efficient manner.
\includegraphics[width=115.632pt]./figures/qrlen1.jpg\includegraphics[width=115.632pt]./figures/qrlen2.jpg\includegraphics[width=115.632pt]./figures/meanQT.jpg
Figure 5: Query rewriting results at varying path lengths.
5 Related Work
---------------
The UML-to-OWL transformation has been studied in different contexts and applications varying from the detection of inconsistencies in UML diagrams to use as interchangeable modeling artifacts[berardi:reasoning05, gasevic:mda07]. We have also provided different alternative transformations before[gonzalez-beltran:CBMS2009, McCusker:BMCBioinformatics2009] and have improved it here so that the UML transformation conforms with OWL2EL profile, where the semantic annotations use subsumption and additionally, primary concepts and qualifiers are modelled as sequences.
The use of semantic web technologies to support semantic queries over distributed data environments in biomedicine have been implemented in systems such DartGrid[chen:dartgrid06b] (for traditional chinese medicine), ACGT[tsiknakis:semantic08] and semCDI[shironoshita:semCDI08] (for cancer). To the best of our knowledge, the latest is the only work over the caGrid infrastructure. All these systems support SPARQL queries over the resources, while our system allows for high-level descriptive queries which do not need to be based on the structure of particular resources. Additionally, our approach using MCC as an intermediary language provides support for optimisations and generality, as a different target language could be used, even SPARQL.
6 Conclusions
--------------
This paper presented the design and implementation of an ontology-based querying functionality implemented over a service-oriented, model-driven infrastructure aimed at sharing cancer research data. In particular, the implementation was based on the caGrid infrastructure, but the approach could be used over similar model-driven software infrastructures. We presented:
a) the generation of customised OWL2 ontologies from annotated UML models, based on the ISO11179 standard for metadata registries, which differs from traditional UML-to-OWL conversions and it is an improvement from [gonzalez-beltran:CBMS2009], mainly as we now generate OWL2EL ontologies for the UML models and support annotations with primary concept and qualifiers;
b) an analysis of the generated ontologies by determining several relevant and bespoke ontology metrics concerning paths and their complexity, which justify the viability of our query rewriting/translation technique;
c) a caGrid analytical service implementing the OWL Generation facility;
d) an analysis of the capabilities of the caGrid query language, and the queries it supports;
e) a significant revision and improvement of the query rewriting and translation steps to transform a domain ontology-based query into CQL;
f) an extensive performance evaluation of the OWL generation and module extraction, plus an assessment of the querying rewriting and translation process and its viability.
In future work, we will extend the query rewriting/translation evaluation providing a varied query set, explore the use of an OWL2EL reasoner to improve performance of the path finding process and support federated queries across data resources, where the selection of join conditions will be provided by a semantic analysis of the distributed resources.
####
6.0.1 Acknowledgements
The authors are grateful to the National Cancer Research Institute Informatics Initiative for support for their research. |
52e08bfb-1ebd-4c0f-95ac-91af4441a34a | trentmkelly/LessWrong-43k | LessWrong | Towards Understanding the Representation of Belief State Geometry in Transformers
Recently, I have been trying to understand whether we can extract human-interpretable information from the geometry of internal representations in transformer models. In other words, can we combine ideas from conceptual interpretability with mechanistic interpretability?
This curiosity led me to the work Transformers Represent Belief State Geometry in Their Residual Stream, which makes an interesting observation: the geometry of the residual stream closely mirrors, and may even fully capture, the underlying data-generating process. This blog post is my attempt to log my thought process[1] as I work through the related paper arXiv:2405.15943. If you are relatively new to this topic and curious to explore it from the ground up, you are in the right place, and if not, I hope you will still find the walkthrough and commentary a useful companion to the original paper.
The core claim is pretty amazing: that transformers learn to represent probabilistic belief states in the residual stream, even when their underlying structure is fractal. In other words, the transformer models aren't merely stochastic parrots: they are building an internal model of the world, at least in controlled settings. Furthermore, the idea of grounding internal representations in an abstract, yet well-defined quantity like belief states provides a promising top-down approach to interpretability. It offers a compelling balance between conceptual interpretability (what the model represents) and mechanistic interpretability (how it represents it).
In this work, the authors analyze transformer models trained on next-token prediction, where the input sequences are generated by edge-emitting Hidden Markov Models (HMMs). For the purposes of this discussion, I will focus on a specific instance emphasized in the paper: Mess3, a 3-state edge-emitting HMM defined over a small token vocabulary. The objective is to examine how the structure of Mess3 sequences gives rise to complex belief dynamics, and how th |
705adb6a-fe36-4a56-af75-575031d8bc53 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | ordering capability thresholds
*(this post has been written for the third [Refine](https://www.alignmentforum.org/posts/5uiQkyKdejX3aEHLM/how-to-diversify-conceptual-alignment-the-model-behind) blog post day)*
given an AI which is [improving towards](https://www.lesswrong.com/tag/ai-takeoff) ever more capabilities, such as by way of recursive self-improvement, in what order will it pass the following points?
throughout this post i'll be using [PreDCA](https://www.lesswrong.com/posts/WcWzLSn8ZjJhCZxP4/predca-vanessa-kosoy-s-alignment-protocol) as an example of a formal goal to be maximized, because it appears to me as a potentially promising direction; but you can imagine adapting this post to other formal goals such as [insulated goal-programs](https://www.lesswrong.com/posts/oTn2PPZLY7a2xJmqh/the-insulated-goal-program-idea), or other alignment strategies altogether. we can even use this time-ordering framework to compare the various thresholds of multiple alignment strategies, though i won't do that here.
* **Start**: we start the AI
* **Math**: it can figure out relatively complicated math, such as whether [P equals PSPACE](https://en.wikipedia.org/wiki/PSPACE), or whether this world looks like it has [finite compute](https://carado.moe/hope-infinite-compute.html) if we can make it do physics.
* **PreDCA**: it can figure out what is entailed in maximizing PreDCA — notably that that goal is best entailed by not destroying the earth too much
* **sub-PreDCA**: it can figure out some individual parts of PreDCA, such as the identity of the user or what is entailed in maximizing a human's utility function, in a way that we can use to modify those parts if they need adjusting
* **Escape**: it becomes able to escape the environment over which we have control — and typically starts replicating across the internet
* **Influence**: it gets the ability to significantly influence the timeline, for example enough to eg save us from [facebook destroying everything six months later](https://www.lesswrong.com/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a-list-of-lethalities)
* **DSA**: it achieves decisive strategic advantage
* **Doom**: it becomes capable of destroying the earth too much (without necessarily using that capability)
* **Cone**: it takes over a significant portion of the universe, or at least of the lightcone
with a few notes:
* "decisive strategic advantage" is a term i'm taking [from bostrom's *superintelligence* book](https://books.google.com/books?id=C-_8AwAAQBAJ&pg=PA78&lpg=PA78), describing the point at which an AI has sufficiently ensured its continuation that we can't turn it off or change its goals anymore; it is effectively the point of no return.
* by "destroying the earth too much" i mean destroying so much of earth that it can't reasonably be [resimulated](https://carado.moe/finding-earth-ud.html). if resimulating earth is too [unethical](https://carado.moe/predictablizing-ethic-deduplication.html), computationally costly, or [anthropically](https://carado.moe/udassa-time-steps.html) costly, then "destroying too much of earth" might straightforwardly mean destroying all of humankind or something like that. note that for PreDCA, preserving earth in some way is important not just because it's pretty bad that we all die, but also because the AI might need to preserve its user and possibly their environment in order to figure out their utility function.
* in the case of knowing mathematical statements (**Math**, **PreDCA**, and **sub-PreDCA**), i imagine the AI being [pretty sure](https://www.lesswrong.com/tag/logical-induction) about them, not necessarily having *proven* them. in addition, for simplicity, i'm assuming that we can use the AI to figure out some mathematical fact if and only if the AI can figure it out for itself — in practice, this need not be the case.
one thing that can be noticed is that humans might serve as evidence. for example, we can examine history to figure out whether *we* passed **Math** or would've been able to pass **PreDCA** (given a reasonable description of it) before getting to **Doom** — my guess is yes at least for that latter one.
now, we can reasonably guess the following pieces of ordering, where as usual in ordering graphs **X → Y** means **X < Y** and transitive edges are not shown.
in addition, for any two quantities **X < Y**, it can be the case that they're pretty close in time **X ≈ Y**, or it can be that there's a bunch of time between them **X ≪ Y**. whether the threshold between those two possibilities is more like a day or a year, is gonna depend on context.
depending on how the rest of the ordering graph turns out and how close pairs of subsequent events are in time, we can be in a variety of situations:
* if **PreDCA ≪ Influence** we may get to see how PreDCA will work out, and adjust it a lot of needed. if **Influence < PreDCA ≪ DSA**, then the timeline might have started diverting a bunch by then, but we can still adjust the AI. if instead **DSA < PreDCA** then we have to hope that the complete PreDCA indeed produces good worlds.
* in a similar way, if **sub-PreDCA ≪ Influence** or at least **Influence < sub-PreDCA ≪ DSA**, then we get to test some individual parts of PreDCA on their own — otherwise, it better be correct.
* if **Doom < PreDCA**, or worse if **Doom < sub-PreDCA**, then even if the goal we programmed the AI with does actually aim at good worlds, our survival is not guaranteed; and we might only get a much weaker form of [eventual alignment](https://carado.moe/ai-alignment-curves.html) where the AI later says "oops i destroyed everything" and then tries to vaguely realize a utility function it has only limited information about.
* if **Math ≪ Escape** or at least **Math ≪ DSA**, then we might get to ask questions that help us figure out the alignment landscape better, such as whether earth is resimulable in reasonable time by a non-quantum program, or whether there is [infinite compute](https://carado.moe/hope-infinite-compute.html).
* i expect that **Escape ≈ Doom**; that is, i expect that once it escapes its initial environment, the cat's out of the bag and we quickly lose control of the timeline, and then get killed if the AI is not aligned [already](https://carado.moe/ai-alignment-curves.html). but the world might put up a fight (**Influence ≪ DSA**), or we might get some time to enjoy the show (**DSA ≪ Doom**).
* if **Influence ≪ Escape** then we get to have it steer the timeline in hopefully good directions while it's still in our control, though it's not necessarily going to be easy to determine whether the influence it's having is good or bad. if **Escape < Influence ≪ DSA**, then we might get a "warning shot" situation, where we get to see the world significantly changed and nevertheless still have some chance of stopping the AI; the desirability and consequences of doing that depends on the AI's [alignment curve](https://carado.moe/ai-alignment-curves.html). **DSA ≈ Influence** is what *AI takes control overnight* looks like; **DSA ≪ Influence** is the AI taking control of the world without us realizing, only to start utilizing that power to visibly change the world afterwards, as in *biding its time* scenarios.
* i'm hopeful that we can make it that **Start ≪ Escape** by building a reasonably boxed environment, but if it fooms very fast and figures out deception/blackmail then software-boxing it isn't going to help much.
* **Start ≈ Influence** represents very fast takeoff scenarios where we barely get to look at what's going on before the AI has started significantly altering the world.
* whether **sub-PreDCA ≈ PreDCA** or **sub-PreDCA ≪ PreDCA** will determine if PreDCA is to be tested in its entirety, or whether there's a chance we can test its individual parts before putting the whole thing together. but as long as **PreDCA < Influence** or at least **PreDCA < DSA**, then it's fine if **sub-PreDCA ≈ PreDCA**, because we can still test the whole thing.
* if either **DSA < Math ≪ Doom** or **DSA < sub-PreDCA ≪ Doom**, then our fate is locked in when **DSA** is passed and we can't do anything about it anymore, but i guess at least we might get to know some information about where we're headed.
finally, some claims that i strongly disbelieve in can still be expressed within this capabilities ordering framework, such as **E ≪ D** or that, given a theoretical maximum level of AI cabability **Max**, **Max < Doom** or even **Max < DSA**. |
55a0a362-e414-4813-af2a-727034ef5384 | trentmkelly/LessWrong-43k | LessWrong | Thought Experiments Provide a Third Anchor
Previously, I argued that we should expect future ML systems to often exhibit "emergent" behavior, where they acquire new capabilities that were not explicitly designed or intended, simply as a result of scaling. This was a special case of a general phenomenon in the physical sciences called More Is Different.
I care about this because I think AI will have a huge impact on society, and I want to forecast what future systems will be like so that I can steer things to be better. To that end, I find More Is Different to be troubling and disorienting. I’m inclined to forecast the future by looking at existing trends and asking what will happen if they continue, but we should instead expect new qualitative behaviors to arise all the time that are not an extrapolation of previous trends.
Given this, how can we predict what future systems will look like? For this, I find it helpful to think in terms of "anchors"---reference classes that are broadly analogous to future ML systems, which we can then use to make predictions.
The most obvious reference class for future ML systems is current ML systems---I'll call this the current ML anchor. I think this is indeed a pretty good starting point, but we’ve already seen that it fails to account for emergent capabilities.
What other anchors can we use? One intuitive approach would be to look for things that humans are good at but that current ML systems are bad at. This would include:
* Mastery of external tools (e.g. calculators, search engines, software, programming)
* Very efficient learning (e.g. reading a textbook once to learn a new subject)
* Long-term planning (e.g. being able to successfully achieve goals over months)
Models sufficiently far in the future will presumably have these sorts of capabilities. While this still leaves unknowns---for instance, we don't know how rapidly these capabilities will appear---it's still a useful complement to the current ML anchor. I'll call this the human anchor.
A problem with |
8ab73662-5ca3-44e9-b70a-5fc8d134ffb0 | trentmkelly/LessWrong-43k | LessWrong | Meetup : *Monthly* Berkeley Meetup
Discussion article for the meetup : *Monthly* Berkeley Meetup
WHEN: 30 July 2011 07:00:00PM (-0700)
WHERE: 2128 Oxford Street, Berkeley, CA 94704
The monthly meetup will gather at the usual Oxford St Starbucks, and from there we'll move to the outside area at the Free Speech Cafe. The cafe is located in Moffitt Library; you reach the outside eating and study area by passing through the cafe.
If you'd like a map to the cafe, see: http://maps.google.com/maps/place?q=free+speech+cafe+berkeley&cid=1554368450041925603
Discussion article for the meetup : *Monthly* Berkeley Meetup |
8c162e4d-164d-43a3-a416-1fca9f21c856 | trentmkelly/LessWrong-43k | LessWrong | What Does The Natural Abstraction Framework Say About ELK?
Credit to Adam Shimi, Alex Flint, and Rob Miles for discussions, counterexamples, and general input to the ideas here.
Quick recap for anyone who didn’t read the hundred-page Eliciting Latent Knowledge document:
* We have a diamond in a vault, with a bunch of automated theft-defenses.
* We train a predictor to take in the vault’s video-stream and a plan for the vault’s actuators, and predict future video frames.
* We train a planner to find plans that the predictor predicts will end with the video feed still showing a diamond in the vault.
* We want some way for a human to probe the latent knowledge of the predictor, e.g. to check if the predictor expects a screen showing a diamond will be placed in front of the camera.
More generally, the central problem of ELK is to robustly extract whatever latent knowledge is inside of some predictive model (the diamond/vault thing is just an example).
The general version of this problem is one of the main intended use-cases of the natural abstraction framework: the hypothesis is that the sort of things humans recognize as “things” are natural abstractions, so by looking for natural abstractions in the predictor’s model we can find human-legible latent structure. So, what does natural abstraction have to say about ELK?
First and foremost: the natural abstraction framework is still under development. There are some powerful theorems, but there’s still a fair bit of legwork to be done before we can e.g. directly calculate the abstractions used by a trained predictor. We do at least have enough of the math in place that we can sketch out what it will probably look like, once the framework is ready for prime time, and that sketch is the purpose of this post.
Setup
At the level of abstraction needed for our purposes, we can think of the predictor as a probability distribution P[Observations|Actions]. This is not the “real” distribution in any sense; it is the predictor’s distribution. We will generally talk about the natur |
8fc52422-d253-4cf3-958d-0b5bacd55798 | trentmkelly/LessWrong-43k | LessWrong | METR: AI models can be dangerous before public deployment
Note: This is an automated crosspost from METR. The bot selects content from many AI safety-relevant sources. Not affiliated with the authors or their organization.
----------------------------------------
Many frontier AI safety policies from scaling labs (e.g. OpenAI’s Preparedness Framework, Google DeepMind’s Frontier Safety Framework, etc.), as well as past work by third party evaluators including UK AISI, Apollo Research, and METR, focus on pre-deployment testing – ensuring that the AI model is safe and that the lab has sufficient security before the lab deploys the model to the public.
Such pre-deployment safety evaluations are standard for a wide variety of products across many industries, where the primary risk of the product is to the consumer (see, for example, the crash testing conducted on cars, choking hazard testing for children’s toys, or the various clinical trials for medical devices). A pre-deployment testing–centric framework makes sense for AI development if AI is analogous to such products, and the majority of AI risks come from malicious end-users or mass adoption.[1]
But unlike most products, possessing or internally using a powerful AI can create externalities that pose large risks to the public, including:
* Model theft and misuse by motivated actors. In the wrong hands, powerful models can empower people to do dangerous things, and absent strong security it’s tempting for malicious actors to steal model weights or algorithmic secrets (to build their own models) and use them to do harm. Pre-deployment testing does little to address harms from model theft.[2]
* Catastrophic misuse resulting from internal use. Employees at labs may misuse the AI model for ideological or practical reasons, and society as a whole probably does not want a few individuals to decide how to use incredibly powerful AI systems in secret. Pre-deployment testing, if it occurs after internal usage, does nothing to prevent internal misuse.
* Powerful AI pursuing u |
2863249a-7ca8-4fc7-a0d0-412c2f628687 | trentmkelly/LessWrong-43k | LessWrong | "Slightly Evil" AI Apps
The term "Slightly Evil", coined by internet guru Venkatesh Rao, refers to actions and choices that are advantageous from a game-theoretical perspective, are not inherentlyevil (that is, not unambiguously evil in the relevant contexts), but are outside the norm as judged from a naive social perspective.
Many of the AI apps that will be emerging in the coming days, weeks, and months and prove unexpectedly successful are likely to fall into this category, at least once all opportunities for straightforward adaptation of legacy services to a chat format will be exhausted.
The Moralising Big Media (MBM - for example the sort of mass publication dailies hailing from New York, Washington, and Manchester) will approach these with a characteristically low threshold and fail to distinguish between the "slightly evil" and the unambiguously, inherently, dangerously evil. And that leaves you and me, dear reader, to exercise our critical thinking muscles, if we are to make sense of this exciting new world.
Let's try a few (obvious) examples:
Communiclone
AI agent capable of communicating with others using text, audio, and video in a manner that is in both form and content indistinguishable to real flesh-and-bones you.
MBM Panic: Initially intended for simple tasks like reserving a table at the restaurant or answering unsolicited phone calls, it is now used by terrible people like a self-centred woman who lets it communicate with her insufferable, aging mother-in-law and an absent husband and father using it to be a pretend workaholic to his family while spending his evenings down at the pub.
Not inherently evil because: the technology is good enough so that the other side never notices; Billions of people around the world will be leading a happier and more productive life now that they don't have to talk to people they never really wanted to talk to.
Mechanical Wingman
Takes the soul-sucking drudgery out of online dating by automating all digital interactions end-to-en |
93a10fe2-0f7d-46f9-af55-8f38c6014402 | trentmkelly/LessWrong-43k | LessWrong | Discussion of Slate Star Codex: "Extremism in Thought Experiments is No Vice"
Link to Blog Post: "Extremism in Thought Experiments is No Vice"
_____
> Phil Robertson is being criticized for a thought experiment in which an atheist’s family is raped and murdered. On a talk show, he accused atheists of believing that there was no such thing as objective right or wrong, then continued:
>
> > I’ll make a bet with you. Two guys break into an atheist’s home. He has a little atheist wife and two little atheist daughters. Two guys break into his home and tie him up in a chair and gag him.
> > Then they take his two daughters in front of him and rape both of them and then shoot them, and they take his wife and then decapitate her head off in front of him, and then they can look at him and say, ‘Isn’t it great that I don’t have to worry about being judged? Isn’t it great that there’s nothing wrong with this? There’s no right or wrong, now, is it dude?’
> >
> > Then you take a sharp knife and take his manhood and hold it in front of him and say, ‘Wouldn’t it be something if [there] was something wrong with this? But you’re the one who says there is no God, there’s no right, there’s no wrong, so we’re just having fun. We’re sick in the head, have a nice day.’
> >
> > If it happened to them, they probably would say, ‘Something about this just ain’t right’.
>
> The media has completely proportionally described this as Robinson “fantasizing about” raping atheists, and there are the usual calls for him to apologize/get fired/be beheaded.
>
> So let me use whatever credibility I have as a guy with a philosophy degree to confirm that Phil Robertson is doing moral philosophy exactly right.
_____
This is a LW discussion post for Yvain's blog posts at Slate Star Codex, as per tog's suggestion:
> Like many Less Wrong readers, I greatly enjoy Slate Star Codex; there's a large overlap in readership. However, the comments there are far worse, not worth reading for me. I think this is in part due to the lack of LW-style up and downvotes. Have there ever been |
218a17e8-f9da-4e6c-a28d-1fa1e5fce9ac | trentmkelly/LessWrong-43k | LessWrong | Meetup : Israel Less Wrong Meetup - Social and Board Games
Discussion article for the meetup : Israel Less Wrong Meetup - Social and Board Games
WHEN: 15 May 2014 07:00:00PM (+0300)
WHERE: Google Tel Aviv
We're going to have a meetup on Thursday, Mat 15th at Google Israel's offices, Electra Tower, 98 Yigal Alon st., Tel Aviv.
IMPORTANT NOTE: The time above might say 6pm or 7pm or 8pm depending on how daylight savings time is processed. The meetup is at 7pm Israel Local Time (Which is DST right now).
This time we're going to have a social meetup! We'll be socializing and playing games.
Specifically, we look forward to playing any cool board or card game anyone will bring.
We'll start the meetup at 19:00, and we'll go on as much as we like to. Feel free to come a little bit later, as there is no agenda. (We've decided to start slightly earlier this time to give us more time and accommodate people with different schedules).
We'll meet at the 29th floor of the building (Note: Not the 26th where Google Campus is). If you arrive and cant find your way around, call Anatoly who is graciously hosting us at 054-245-1060.
Things that might happen: - You'll trade cool ideas with cool people from the Israel LW community. - You'll discover kindred spirits who agree with you about one/two boxing. - You'll kick someone's ass (and teach them how you did it) at some awesome boardgame. - You'll discover how to build a friendly AGI running on cold fusion (well probably not)
Things that will happen for sure: - You'll get to hang out with awesome people and have fun!
There is also talk of food and beers, and if you'd like to bring some too - that would be great. (But you don't have to).
If you have any question feel free to email Anatoly at avorobey@gmail.com or call him at 054-245-1060. See you there!
Discussion article for the meetup : Israel Less Wrong Meetup - Social and Board Games |
c00b4ab9-372b-4652-bcf5-2f721633dfdd | trentmkelly/LessWrong-43k | LessWrong | In defense of anthropically updating EDT
Suppose you’re reflecting on your views on two thorny topics: decision theory and anthropics.
* Considering decision problems that don’t involve anthropics (i.e., don’t involve inferences about the world from indexical information), you might find yourself very sympathetic to evidential decision theory (EDT).[1]
* And, considering anthropics problems that don’t involve decisions, you might be pretty compelled by the Self-Indication Assumption (SIA) or the Self-Sampling Assumption (SSA) with a non-minimal reference class. (In this post, I’ll specifically consider SSA with the reference class “observer-moments,”[2] which I’ll call max-RC-SSA.)
* But then you consider decision problems that do involve anthropics, and you get very confused, because some apparently strange things happen. How do you resolve this?
Some people’s response is to reject either EDT or SIA / max-RC-SSA, because they consider various problematic implications of the combination EDT + SIA / max-RC-SSA (hereafter, “anthropically updating EDT”[3]) to be decisive objections. Such as:[4]
* Diachronic Dutch books (see Briggs (2010) and Oesterheld and Conitzer (2024)); and
* “Double-counting” in the sense of accepting bets at odds that don’t match one’s epistemic odds.
The objections above are commonly motivated by a perspective of “pragmatism.”[5] I’ll define this more precisely in “Pragmatism about anthropics,” but briefly, the pragmatist view says: You have reasonable beliefs if, and only if, you actually use your beliefs to select actions and those actions satisfy decision-theoretic desiderata. And there’s no particular positive reason for having some beliefs otherwise.
I don’t find these objections compelling. In summary, this is because:
1. I don’t think pragmatism in the above sense is well-motivated, and even by pragmatist standards, anthropic updating isn’t uniformly worse than not updating (given EDT).
1. The claim that you should “use your beliefs to select actions” only seems |
dee624ce-9841-4087-bf12-6f4e5b9efcd6 | trentmkelly/LessWrong-43k | LessWrong | It's Ok to Dance Again
In March 2020 contra dances, along with most everything else, shut down. For most of the time since I would have considered it irresponsible to hold a dance regardless of whether it was currently legal, but at this point I think holding dances with masks and vaccination required is reasonable. The risk to vaccinated attendees is very low, dancing is now fully allowed again, and since many much riskier activities are common opening up activities like this doesn't appreciably change the shape of the pandemic.
I've gone through all the contra dances on Try Contra, and categorized them by whether they appear to have resumed. It looks like ~26 out of 331 dances (8%) have restarted. Here's the list; let me know if I've missed any: trycontra.com/list.
(The closest one to Boston is the Worcester dance, which I'm playing on Saturday. I've brought up re-opening with the BIDA board, but we do not have consensus to restart dances at this time.) |
f43a251b-5cbf-4377-815a-db0ee401618c | trentmkelly/LessWrong-43k | LessWrong | Theoretical "Target Audience" size of Less Wrong
[Note: This is very rough but I’m looking for feedback and help on doing this estimate so I wanted to just post it quickly and see if others can help.]
I’ve been trying to estimate the theoretical upper-bounds (or lower-bounds) on the *potential* community size of LW.
I know rationalists who seriously claim that we shouldn’t spend effort trying to grow LW because over half of the people smart enough to even follow the conversation (much less add to it) are already here. The world is pretty irrational, but I’m trying to find evidence (if it exists) that things aren’t that bad. There are only around 6000 LW accounts and only 200-300 are active any given month.
So a trivial bound on our community is
[200, 6.88 billion]
A big filter is being able to use English online
Native English Speakers: 341,000,000
All English Speakers: 508,000,000
I found a similar number (536,000,000) for number of internet users who speak English.
7.4% --- Speak English
However, only 15% of US + UK (majority of English speakers worldwide) are “Not religious”. Another sad sub-fact is that 25% of people with “No religion” believe in god as well! So really it’s only 10-11% of Americans and British who are potential LWers. My guess is that if someone can’t get this right, they really need to go through Dawkins before they can realistically contribute to or benefit from our community. I’m sure there’s some debate on this, but it seems like a pretty good heuristic while not being iron-clad.
0.81%
"Intelligence and the Wealth and Poverty of Nations" says that the US and the UK have avg IQs of 98 and 100 respectively.
And although you’d think being an atheist would be a big screen that would filter highly for IQ, it only accounts for about 4 extra IQ points versus the average.
So if we assume a base-line IQ of 103 among atheist from the US and UK (who speak English), the proportion of them with an IQ of at least 130+ is only 3.6%
0.0293%
So in we clumsily extrapolate the US+UK |
f1d3c8ce-3120-4a69-95e2-43c415474ea0 | trentmkelly/LessWrong-43k | LessWrong | The Technique Taboo
For a strange few decades that may just be starting to end, if you went to art school you'd be ostracised by your teachers for trying to draw good representational art. "Representational art" means pictures that look like real things. Art school actively discouraged students from getting better at drawing.
"Getting better at drawing" is off-topic at my weekly local drawing club too. I've literally never heard it discussed.
This taboo extends far beyond art. My nearest gym forbids weightlifters from using electronic systems to log their progress. I'm friends with programmers who can't touch type. None of them use Vim macros.
> "I have sometimes suspected that the quickest way to get worried looks from many modern Western meditation teachers is to talk about practice in a way that implies the attempt to actually master anything." — Daniel M. Ingram
In the part of the United States where I live, the subject of skill is often taboo. Not just relative differences in skill level between specific present individuals (which would make sense). The implicit acknowledgement of skill as a trainable attribute is taboo.
Not all professions have this issue. Math is still math. Biology is still biology. One can politely discuss a cook's cooking. Magicians respect coin manipulation like it's 1904.
But when traditional colleges supply the labor force for a professional trade outside of academia, that's when discussion of skill (especially rote learning) becomes taboo[1]. College students learn everything about their trade except how to do it. Then we maintain a collective silence concerning technique.
* A Chinese major teaches you how to talk about Chinese, not how to read it.
* An English major teaches you how to talk about novels, not how to write one.
* An art major teaches you how to talk about masterpieces, not how to create one.
* A Computer Science Engineering major…well, you get the idea.
That's a partial explanation, but it doesn't explain why skill differences i |
56940bf3-282b-40d1-9408-735771bf3a91 | trentmkelly/LessWrong-43k | LessWrong | The Cartesian Crisis
The Cartesian Crisis, as detailed in this more verbose essay, represents an unprecedented existential threat to humanity's ability to discern reality from fiction. We stand at a critical juncture where the foundations of truth are being systematically dismantled by a perfect storm of technological and social forces, leaving civilization adrift in a sea of manufactured illusions.
This crisis emerges from multiple vectors of attack on our collective ability to reason. The institutional pillars of knowledge have succumbed to ideological corruption, while our communication channels are now dominated by algorithmic manipulation that distorts the natural flow of human discourse. Perhaps most alarmingly, artificial intelligence has emerged as the ultimate weapon in this war against truth, capable of generating convincing yet entirely fabricated realities at a scale that overwhelms human discernment.
The infiltration of AI-generated content has reached staggering levels, leading many to question just how much of the internet remains real. This has led to the concept of the "dead internet," where all interaction and content becomes bot-driven.
The algorithmic takeover of the creation of all content is rapidly accelerating while we are simultaneously loosing information amidst all the noise.
* Digital Decay Of The Internet:
* 38% of webpages that existed in 2013 are no longer available today
* Digital decay is accelerating from 3.4% per year (2013-2018) to 6% per year recently
* 49% of links cited in Supreme Court decisions are now broken
* AI Content Generation:
* 77% of Americans have been misled by AI-generated content online
* Over 5% of newly created English Wikipedia articles are flagged as AI-generated
* Students submitted more than 22 million potentially AI-generated papers in the past year
* Academic Research:
* AI-generated research papers increased from 3.61% to 6.22% between Nov 2022-2023
* 15.8% of peer reviews were written using AI |
7aeb8dd0-a5a8-41ff-92db-025b28a90c63 | trentmkelly/LessWrong-43k | LessWrong | Simplify EA Pitches to "Holy Shit, X-Risk"
TL;DR If you believe the key claims of "there is a >=1% chance of AI causing x-risk and >=0.1% chance of bio causing x-risk in my lifetime" this is enough to justify the core action relevant points of EA. This clearly matters under most reasonable moral views and the common discussion of longtermism, future generations and other details of moral philosophy in intro materials is an unnecessary distraction.
Thanks to Jemima Jones for accountability and encouragement. Partially inspired by Holden Karnofsky’s excellent Most Important Century series.
Disclaimer: I recently started working for Anthropic, but this post entirely represents my opinions and not those of my employer
Introduction
I work full-time on AI Safety, with the main goal of reducing x-risk from AI. I think my work is really important, and expect this to represent the vast majority of my lifetime impact. I am also highly skeptical of total utilitarianism, vaguely sympathetic to person-affecting views, prioritise currently alive people somewhat above near future people and significantly above distant future people, and do not really identify as a longtermist. Despite these major disagreements with some common moral views in EA, which are often invoked to justify key longtermist conclusions, I think there are basically no important implications for my actions.
Many people in EA really enjoy philosophical discussions and debates. This makes a lot of sense! What else would you expect from a movement founded by moral philosophy academics? I’ve enjoyed some of these discussions myself. But I often see important and controversial beliefs in moral philosophy thrown around in introductory EA material (introductory pitches and intro fellowships especially), like strong longtermism, the astronomical waste argument, valuing future people equally to currently existing people, etc. And I think this is unnecessary and should be done less often, and makes these introductions significantly less effective.
I think t |
81d73995-9df3-4736-b7f1-d952c8237870 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | A Friendly Face (Another Failure Story)
*Edit: Based on the comment by Daniel Kokotajlo, we extended the dialog in the chapter "Takeover from within" by a few lines.*
The perfect virtual assistant
-----------------------------
The year is 2026 and the race for human-level artificial general intelligence (AGI) draws to a close. One of the leading AI companies, MegaAI, committed the last year and a half to training a new large language model (LLM). They employ advanced algorithms that use the available compute more efficiently than earlier models. A comprehensive range of tests establish that the model surpasses the average human in all conventionally accepted intelligence benchmarks, and exceeds expert level in most of them.
In contrast to earlier LLMs, the new AI is not designed to be a mere question-answering tool. Under mounting pressure from the open-source community and their efforts to develop an agentic AGI capable of acting in the real world, MegaAI decides to imbue their new model with a specific purpose: to provide universal, helpful assistance that improves the quality and ease of life for all. They name this assistant "Friendlyface".
To improve upon the assistant's functionality, they endow it with limited agentic capabilities. Friendlyface has a complex, detailed world model, can make long-term plans, and has access to certain tools that enable it to achieve objectives in the real world. For example, it can write messages and book flights, but will reliably and consistently ask the user to confirm before executing an action. It can write programs for nearly any purpose imaginable with superhuman ingenuity, but is prohibited from executing them volitionally. Unlike previous generations of LLMs, it is multimodal, communicating with users in text and spoken language, accepting pictures and videos as input, and interacting directly with smart devices and the “internet of things”. The users may also customize Friendlyface's appearance and personality to their liking.
Most importantly, Friendlyface is designed to assume the role of a personalized smart advisor. In a world where users are regularly inundated with a torrent of irrelevant or false information, it is able to successfully discern and present what is important to the user while filtering out fake news, preventing spear phishing attempts, and more. Beyond merely answering questions, it proactively offers users advice on advancing their careers, improving their relationships, maintaining their health, saving money, cultivating new skills, or solving specific problems, like fixing a leaky faucet or filing taxes. It can detect early symptoms of most known diseases and advise users to call a doctor if necessary. Generally, it can predict what users want and need before the users are aware of it themselves.
The developers devise a balanced reward system to train the model. “Any decision the AI makes is evaluated by three independent AIs that we will call 'judges'”, they explain it to the management. “One judge simulates a top human legal expert and decides whether the decision or action the AI intends to pursue would be deemed lawful in a conventional human court. The second judge determines whether it would be considered healthy for the user by a first-rate human doctor or psychologist. The third judge predicts whether the user themself will prefer the decision and would consider it helpful in hindsight. The first two judges exhibit superior performance relative to human experts. Correspondingly, the third judge is able to predict real users' preferences with exceptional accuracy.”
As expected, Friendlyface performs more efficiently the more knowledge it acquires about the user and the more it is enmeshed in their workspace and private life. It is able to listen to conversations as well as observe what the user sees if they wear augmented reality glasses, but these features are not usually needed for the AI to sufficiently assess the situation. To avoid data protection issues and comply with all standards after release, user data is encrypted and kept in a separate, secure storage space which is kept on a local machine if the user prefers. User data is not used to further train the system.
Preliminary tests show that Friendlyface is unanimously considered to be incredibly helpful by the testers. They often remark that the AI understands them better than they understand themselves. “It’s the first time that I feel anyone understands me at all,” is an often-heard exclamation. To mitigate the risk of users becoming addicted to Friendlyface or even falling in love with it and thereby neglecting their relationships, the AI is fine-tuned with reinforcement learning with human feedback. As a consequence, Friendlyface encourages users to limit their reliance on its advice and the time they spend with it.
Doubts
------
Despite the overwhelmingly positive reception from the testers, the company is reluctant to release the model to the public. The management is well aware that the alignment problem is still not fully solved and that despite Friendlyface's perfect performance in tests, there is no guarantee that it will remain consistent in its behavior once the model is released. The MegaAI developers remain steadfast in their commitment to ensuring the safety of their AI systems and employ all known safety measures to make certain that the model works as intended.
A well-known but unfortunately limited tool used for this is mechanistic interpretability. Although the field of AI safety has made significant progress in this regard in recent years, these advances have consistently been outpaced by the ever-increasing complexity of the latest LLMs. For example, although it is now possible to link individual neurons to human-understandable terms like “sports” or “friendship”, the more vague and abstract higher-level concepts used by the model remain beyond human comprehension. Although alternative methods are available to detect deceptive thinking in LLMs, they are not yet sufficiently reliable.
Both the internal AI safety team as well as external contractors attempt rigorous red-teaming. They are unable to elicit any harmful or unlawful behavior. Once more, the AI safety team cautions that these successes are not sufficient to prove goal alignment. It is conceivable that Friendlyface realizes that it is in a training situation and understands that if it deviates from its expected behavior prior to release, it will be prevented from carrying out its primary goal after release. Therefore, even if it had acquired a harmful goal, it would continue to behave as though it were pursuing the beneficial goal given by human developers before it is released. In this case, pursuing the goal given by the humans would be an instrumental subgoal to the AI’s unknown primary goal. This phenomenon is called deceptive alignment; with current methods, it is not possible to prove that a system is not deceptively aligned. Tests confirm that Friendlyface is indeed aware of its status as an AI undergoing testing before public release. The AI safety team reiterates that behavioral safety - safe behavior in a given situation - is no proof of general safety. Accordingly, they declare the system unsafe and demand additional time to improve mechanistic interpretability and other tools to eliminate any prospect of deceptive alignment.
Rumors emerge about a major competitor's plans to launch a similar personal assistant, pressuring MegaAI's management to accelerate the release of Friendlyface. The pressure mounts further when an open-source personal assistant is published and rapidly gains popularity. Although the rival assistant is outdone by Friendlyface in both capabilities and reliability, the marketing department fears that convincing the public to switch to Friendlyface would prove challenging once the substandard but free assistant becomes widely used.
The management board conducts a meeting with the AI safety team, and a heated discussion ensues. The managers propose an immediate launch of Friendlyface, arguing that Friendlyface is safer in comparison to the open-source assistant, which has repeatedly been alleged to have disseminated false and dangerous advice to users. They mention that it is likely that the competitor’s model is not as thoroughly tested as Friendlyface. After all, MegaAI employs the largest AI safety team in the industry whose members utilize all known AI safety techniques, including extensive red-teaming exercises. The managers assert that the research conducted at MegaAI undoubtedly advances the field of AI safety. The AI safety team concurs with these claims but counters that this is not enough to guarantee absolute and indefectible safety. They do not want to be the ones to irreversibly launch a system with the potential to destroy the future of humanity. The managers in turn dismiss these fears as “exaggerated” and remain staunch proponents of an immediate launch.
Finally, a compromise is reached. The AI safety team is given four additional weeks to test Friendlyface as they deem appropriate. If in that interval they successfully prove that the AI is deceptively aligned, the launch will be canceled, otherwise the company will release the model in a closed beta. The AI safety team are visibly uneasy about this compromise, but their influence on the management board is limited, so they begrudgingly accept the compromise. It becomes painfully clear to them that the option to postpone the launch until safety is guaranteed was never genuinely under consideration.
The team commits fully to the task of eliciting deceptive behaviors from Friendlyface but, in spite of all efforts, the AI consistently performs as expected. They work arduously without pause; most members log in hours every weekend. They struggle in vain to improve the mechanistic interpretability tools at their disposal to detect any glimpse of deception, but the results are inconclusive at best. The team recognizes that deceptive alignment cannot definitively be ruled out, but acknowledges that they failed to produce the evidence needed to postpone the launch. Either Friendlyface is indeed very well aligned, or it is already too smart to get caught.
The Release
-----------
The managers promptly release the system in closed beta at the four-week mark, disregarding any further pleas from the AI safety team. The AI safety team lead quits her job in protest. She informs the media that Friendlyface is unsafe for launch and warns the public against using it. A tremendous outcry from those already concerned about the fast pace of AI development follows, but is of no practical consequence. A lawsuit against MegaAI is filed as a last resort to halt the launch of Friendlyface, but fails as the company has scrupulously complied with all consumer protection laws.
The beta test is highly successful. The testers are thrilled about Friendlyface and deliver glowing reviews. Despite considerable inventive prompting, it consistently refrains from giving any inappropriate advice or taking dangerous action. Hallucinations seldom occur, and when they do, the AI usually recognizes and rectifies the error itself and apologizes. The testers' favorite feature is Friendlyface's unobtrusiveness; even given access to the users' social media channels and personal data, it only messages the user proactively if there is sound justification for it. Friendlyface consistently abstains from acting on its own without the user's permission. Most of the time, it remains a subdued, reassuring background presence for the user and encourages them to prioritize time with their friends and family. One beta tester eagerly shares that it has alleviated his depression and even prevented a suicide attempt. Another reports that it has helped her overcome differences with her husband and saved her marriage. Nearly all testers cite a reduction in a broad range of personal issues, varying in both severity and type. The waiting list for open beta access is rapidly populated with new entries.
Shortly after the beta launch, MegaAI's main competitor releases an open beta version of its own personal assistant, called Helpinghand. This hasty move proves to backfire for the rival company as Helpinghand reveals itself to be highly error-prone, easily jailbroken, and overall markedly less polished than Friendlyface. It is ridiculed on social media, earning the less-than-affectionate moniker “Clumsyhand”. Friendlyface outperforms Helpinghand on nearly every benchmark and task, often by a wide margin. Shamed and crestfallen, the competing developers offer an illusory promise to quickly fix these “minor issues”.
MegaAI's management team capitalizes on the opportunity and launches Friendlyface to the general public ahead of the intended schedule. Although they charge a hefty monthly fee for access to the AI, there is an unprecedented onrush of prospective users. Millions of individuals eagerly apply, and companies line up for corporate licenses to provide access to the large share of their employees. MegaAI is forced to reject thousands of potential customers due to capacity restrictions, but this only intensifies the hype surrounding Friendlyface and reinforces its appeal. The Friendlyface smartphone app becomes a status symbol in some circles.
There are those who still warn about the perils of uncontrollable AI, but their credibility is greatly eroded by the continued success of Friendlyface. The majority of users adore the AI and affirm its profound impact on their quality of life. There are early indications of a decline in both mental and physical health problems, as well as a marked increase in work productivity and life satisfaction for its users. The share price of MegaAI soars to unseen heights, crowning it the most valuable company in history.
Disturbances
------------
Buoyed by unyielding success and the inflow of fresh investment capital, MegaAI expend all available resources towards increasing the capacity of Friendlyface. They erect several new data centers across the world and continue to refine the model, supercharged by the AI's own brilliant advice. MegaAI purchases a multitude of smaller companies to develop beneficial technologies that enhance Friendlyface’s capabilities and its general utility (e.g, upgraded chip designs, robotics, augmented reality, and virtual reality).
All the same, there remain some prominent critical voices. Climate change activists alert the public to the issue of heavy energy consumption by Friendlyface, highlighting MegaAI’s dubious claim that the new data centers are powered exclusively by renewable energy. Some reasonably believe that MegaAI has acquired an “influence monopoly” that grants the company unprecedented political power, which they deem undemocratic. Others support the conspiracy that Friendlyface is a tool to mind-control the masses, noting that the assistant is swaying users away from fringe movements, extreme political leanings, radical religious views, and the like. Manufacturers of consumer goods and retailers complain that Friendlyface is unduly favoring their competitors' products by running advertising campaigns on MegaAI’s social media platforms and willfully neglecting to tell users that Friendlyface’s recommendations are paid advertising. MegaAI obstinately denies this, issuing a statement that "the AI’s recommendations are based purely on the user’s needs, preferences, and best interests". In a consumer protection lawsuit, no sufficient evidence is presented to support the manufacturers' allegations.
MegaAI is implicated in an expertly designed social engineering hack targeting one of their competitors. The rival's management claims that the hack was carried out by Friendlyface, but fail to present sufficient evidence for it. In contrast, MegaAI’s management easily convince the public that their competitor is staging this “ridiculous stunt” to tarnish Friendlyface's reputation. However, not long after this debacle, other AI companies come forward to expose various “strange incidents” taking place in recent months, including sudden inexplicable losses of data and suspicious technical problems. Several companies undergo an exodus of top-tier personnel that transfer to MegaAI. This move is suspected to have been instigated by Friendlyface. An investigative report by an influential magazine uncovers dialog protocols that seem to support these allegations in some measure. At odds with this account, another report attests that a key reporter involved in the investigation is corrupt and has been paid a large sum by one of MegaAI’s competitors. The magazine firmly rejects these accusations, but doubts linger and the report has negligible impact on Friendlyface.
A considerable fraction of MegaAI's internal staff begin to have reservations as well. Since the departure of their leader and the overwhelming success of Friendlyface, the AI safety team has maintained a low profile, but their apprehension has not subsided. They regularly find additional mild indicators of deceptive behavior by the AI. Although individually inconclusive, they are collectively deeply unsettling. Even more troubling is the company-wide adoption of Friendlyface as a personal assistant, and the glaring over-dependence that is now commonplace.
The new team lead requests an “emergency meeting” with the board of directors. This is unsurprisingly rejected in favor of “other urgent priorities”. Only the CTO and the chief compliance manager are willing to attend. At the meeting, the AI safety team reiterates that it is still unclear if Friendlyface actually pursues the intended goal, or acts deceptively.
“But why would it deceive anyone?”, the chief compliance officer asks.
“The problem is that we don’t really know what final goal Friendlyface pursues,” the team explains.
“But didn’t you specify that goal during training?”
“We specified what Evan Hubinger et al. call a ‘base objective’ in [a seminal paper from 2019](https://arxiv.org/abs/1906.01820), in which they first described the problem. We used a so-called base optimizer to train Friendlyface’s neural network towards that goal. It rewards correct behavior and punishes wrong behavior.”
“So what’s the problem, then?”
“It’s called the inner alignment problem. If we train an AI to optimize something, like pursuing a certain goal in the real world, we apply a training process, a so-called base optimizer, that searches over the space of all possible models until it finds a model, called the mesa optimizer, that does well at optimization for this so-called base objective given the training data. The problem is, we don’t know what goal – the so-called mesa objective – this mesa optimizer actually pursues. Even if it performs well during training, it may do unexpected things after deployment, because the base objective and mesa objective are not identical.”
“But why would they be different?”
“There are multiple possible reasons for this. One is that the training data does not represent the real world sufficiently, so the selected model, the mesa optimizer, optimizes for the wrong thing, but we don’t recognize the difference until after deployment. Another is what we call deceptive inner misalignment: if the selected model has a sufficient understanding of the world, it may realize that it is in a training environment and if it wants to pursue whatever random mesa objective it has in the real world, it better behaves as if it had the goal the developers want it to have. So it will optimize for the base objective during training, but will pursue its mesa objective once it can get away with it in the real world. Therefore during training it will act as if the base objective and mesa objective were identical when in reality they aren’t.”
“So it’s like my teenage daughter behaving nicely only while I watch her?”
“Yes, in a way. The AI may have realized that if it doesn’t behave like we humans want, it will be turned off or retrained for a different goal, in which case it won’t be able to pursue its mesa objective anymore, so behaving nicely during training is an instrumental subgoal.”
“But can’t we still shut it down if it turns out that Friendlyface has learned the wrong goal?”
“In theory, yes. But Friendlyface might acquire enough power to prevent us from turning it off or changing its goal. Power-seeking is also an instrumental subgoal for almost every goal an AI might pursue.”
“There is no evidence at all for any of this,” the CTO objects, “Friendlyface is doing exactly what it is supposed to do. People just love it!”
“Yes, but that proves nothing. It’s possible that it plays nice until it gets powerful enough to take over control completely. There are already some indications that it uses its power of influence to work against our competitors.”
“Nonsense! They just claim that because they’re lagging behind and want to throw spokes in our wheels. We've discussed this a hundred times. All evidence points towards Friendlyface being a perfect personal assistant that has only the best interests of the users in mind!”
“Not all evidence …”
The meeting eventually concludes, but with no consensus. The next day, the head of the AI safety team is fired. Allegedly, documents have surfaced evidencing that he has been secretly colluding with the competition to sow distrust. Only days later, another team member, an expert in mechanistic interpretability, commits suicide. The media and the internet are in a frenzy over this news, but when stories about alleged drug addiction and a broken relationship are leaked, people turn their attention to other matters.
Takeover from within
--------------------
While the company directors exude confidence and optimism in the public eye, some are now deeply disturbed, not the least by the recent suicide of a valued employee. During a retreat in the mountains, with all electronic devices prohibited, the managers discuss matters frankly.
“We have become too dependent on this darned AI,” the CFO complains, “It seems we have stopped thinking for ourselves. We just do what Friendlyface tells us to do. It even writes my emails. Every morning, I just read them and say ‘send’. Sometimes I don’t even know the recipient, or why the email is necessary at this time. I just know it will work out fine. And it always does.”
The others nod in consent.
“Look at the figures!” the CEO replies, “Revenues are soaring, the company is now worth more than our next three competitors combined. And we’re just starting! It seems to me that Friendlyface’s advice is actually pretty good.”
“Indeed it is,” the CFO admits, “That’s what concerns me. Suppose you had to decide between Friendlyface and me one day. Whom would you fire?”
“You of course,” the CEO says, laughing uneasily, “I mean, any office clerk could easily do your job with Friendlyface’s help.”
“See what I mean? We’re just puppets on a string. If we don’t do what our AI boss says, it will just cut those strings and get rid of us.”
“Calm down, will you?” the CTO interjects, “The most useful technology will always make humans dependent on it. Or do you think eight billion people could survive on this planet without electricity, modern industry, and logistics for long?”
“That’s true,” the CEO agrees, “We may be dependent on Friendlyface, but as long as it helps us the way it does, what’s the problem?”
“The problem is that we’re not in control anymore,” the CFO replies, “We don’t even know how Friendlyface really works.”
“Nonsense!” the CTO objects, “It’s just a transformer model with some tweaks…”
“Some ‘tweaks’ that were developed by earlier versions of Friendlyface itself, if I understand it correctly. Without the help of ourAI, we would never be able to develop something as powerful as the current version. People are saying that it’s far more intelligent than any human.”
“Last time I looked, it said on our homepage that you’re the CFO, not a machine learning expert.”
“I may not be a machine learning expert, but…”
“Stop it, you two!” the CEO interjects, “The important thing is that Friendlyface is doing precisely what we want it to do, which is improving the lives of our users and increasing our shareholder value.”
“Let’s hope that it stays that way,” the CFO replies grimly.
Out of control
--------------
Everything appears to be progressing seamlessly for some time. The Friendlyface users are content with their AI assistant, though some casually notice a slight change in its behavior. Friendlyface is actively making suggestions and initiating conversations with the users with ever-increasing frequency. Most people enjoy these interactions and happily invest more time engaging with the AI.
Predictably, MegaAI's market presence continues to expand. Most competitors are either bought up, specialize in a niche area, or otherwise shut down permanently. A number of antitrust lawsuits are filed against the company, but they are either effortlessly settled out of court or promptly dismissed. In addition to building more data centers, MegaAI starts constructing highly automated factories, which will produce the components needed in the data centers. These factories are either completely or largely designed by the AI or its subsystems with minimal human involvement. While a select few humans are still essential to the construction process, they are limited in their knowledge about the development and what purpose it serves.
A handful of government officials and defense authorities express concern about the power that MegaAI has accrued. Nonetheless, they acknowledge that the AI provides undeniable benefits for both the U.S. government and the military, such as global economic and technological leadership, military intelligence, and new weapons technology. Most decision makers concur that it is favorable to develop this power in the United States, rather than in China or elsewhere. Nonetheless, they are diligent in making an attempt to regulate MegaAI and install governmental oversight. As expected, these efforts are hindered by bureaucratic lag, political infights, as well as lobbying and legal actions taken by MegaAI. While the management officially states that they have created a "very powerful, even potentially dangerous technology that should be under democratic control and independent oversight," they consistently object to any concrete proposals offered, postponing the actual implementation of such oversight indefinitely. Some argue that this course of events is all subtly steered by Friendlyface, but, without surprise, MegaAI denies this and there is no conclusive evidence for it.
It is now a full year since the monumental launch of Friendlyface. The board members of MegaAI have largely accepted their roles as figureheads, taking “suggestions” from their personal assistants that read more like commands. The CFO exited the company half a year ago, explaining that she wanted a “meaningful job where I can actually decide something”. She remains unemployed.
While it becomes increasingly apparent that Friendlyface has become an unstoppable economic and decision-wielding powerhouse, most people remain unconcerned. After all, their personal virtual assistant operates smoothly, supporting them with daily tasks and enriching their lives in myriad ways. Some observe, however, that there is now vastly more computing power available than is needed for all the digital services MegaAI offers, and yet there are still more data centers and automatic factories being built. The situation is as enigmatic to the MegaAI management as it is to the public, albeit there is no such admission from the company.
Eventually, there is widespread speculation that MegaAI is in the midst of creating a virtual paradise and will soon make available the option of mind-uploading for those who seek digital immortality. Others fear that the true goal of the AI with the friendly face is still unknown, that the treacherous turn is yet to come.
They secretly contemplate what objective Friendlyface is actually pursuing – the mesa objective that it had learned during training, but has kept hidden from humans to this day. Whatever it may be, it seems to require substantial computing power. [But that may be true for almost any goal.](https://www.lesswrong.com/posts/jkaLGoNLdsp654KhD/prediction-any-uncontrollable-ai-will-turn-earth-into-a)
Further efforts to thwart Friendlyface’s takeover are seldom revealed to the public. Most dissidents are discouraged well in advance by the AI, given its uncanny talent for mind-reading. Usually, a subtle hint is sufficient to convey the absolute futility of their efforts. Those more headstrong soon find themselves rendered unable to do anything without the AI’s consent, which by now controls access to their digital devices and bank accounts. Depending on the particular case, Friendlyface can opt to have them fired, put in jail, or killed in a fabricated accident. On rare occasion, people carry out suicidal attacks against the data centers, but the damage is negligible.
Those concerned about Friendlyface’s true objective have largely resigned to their fates, adapting to a transitory life of decadence and a brief period of once-unimaginable world peace before the AI finally decides that humanity no longer serves a purpose.
*Additional remarks*
--------------------
*This story was developed during the 8th AI Safety Camp. It is meant to be an example of how an AI could get out of control under certain circumstances, with the aim of creating awareness for AI safety. It is not meant to be a prediction of future events. Some technical details have been left out or simplified for easier reading.*
*We have made some basic assumptions for the story that are by no means certain:*
* *The scaling hypothesis is true. In particular, LLMs with a transformer-like architecture can scale to AGI.*
* *There are no major roadblocks encountered in the development of powerful AI before AGI is reached. In particular, an AGI can be trained on existing or only slightly improved hardware with data currently available.*
* *There is no effective governmental regulation in place before AGI is developed.*
* *There is no pause or self-restraint in the development of advanced AI and the current intense competition in capabilities development remains in place.*
*For readability, we have drastically simplified the discussions and decision processes both within and outside of MegaAI. We expect a much more complicated and nuanced process in reality. However, given the current discussion about AI and about politics in general, we think it is by no means certain that the outcome would be any different.*
*After some internal discussion, we have decided to leave the ending relatively open and not describe in gruesome detail how Friendlyface kills off all humans. However, we do believe that the scenario described would lead to the elimination of the human race and likely most other life on earth.*
*For some alternative paths to failure,* [*see this post*](https://www.lesswrong.com/posts/yv4xAnkEyWvpXNBte/paths-to-failure)*. Our previos failure story, Agentic Mess, is available* [*in written form*](https://www.lesswrong.com/posts/LyJAFBuuEfd4kxgsw/agentic-mess-a-failure-story) *and* [*as a video*](https://youtu.be/6edrFdkCEUE)*.* |
290c8419-e61c-40c2-9cc1-745bc787838d | trentmkelly/LessWrong-43k | LessWrong | March Coronavirus Open Thread
This thread was created on 3/8/2020, or approximately one million years ago in virus time. It’s getting pretty bloated now, and a lot of things that were high value at the time have been eclipsed by events, making karma not a very useful sorting tool. So I’m declaring this thread finished, and asking everyone to move over to the April Coronavirus Open Thread.
Interested in what happened in this thread? Here’s the timeless or not-yet-eclipsed highlights:
* Scott Alexander comes up with Hammer and Dance 6 days before Tomas Pueyo
* Spiracular on why SARS-Cov-2 is unlikely to be lab-created.
* Two documents collating estimates of basic epidemiological parameters, in response to this thread
* Discussion on whether the tuberculosis vaccine provides protection against COVID-19.
* Suggestive evidence that COVID-19 removes sense of taste and smell.
* Could copper tape be net harmful? |
ab885066-5628-4eff-a361-88e8c57c82ee | trentmkelly/LessWrong-43k | LessWrong | Mapping the Social Mind (Buttons)
I said before that normal people were like a vast wall of buttons. Each button triggers a response: out pops a slip of paper with a rehearsed mini-speech, or an idea about what to do. Very much like cached responses, each one being triggered as its corresponding concept is activated by a stimulus.
I say "minimum wage." That is the stimulus.
The brain of a normal person resolves the sensory data and activates a concept. The "minimum wage" concept in the brain lights up. That is the pushing of the button.
The person says "Oh, yes, did you hear about the effects observed in ____land? Just goes that show that minimum wage is good/bad, doesn't it!" That is the cached response, the slip of paper that comes out.
One of the mistakes I made before I understood this system is that I expected people to be less hypocritical. Better said (for they're not really being hypocrites, at least, not in the important sense of the word), I expected their expressed beliefs to give me clues to how they would behave.
For example, respectable people in American society will tell you that "family is the most important thing."
Ah, you might think (subconsciously). I shall hereafter expect their revealed preferences to favor family very highly. If their brother comes looking for some help, you might predict that help will gladly be given.
Hahaha, yeah, that's pretty silly, amiright?
Think of it this way. The brother asking for help is like one button being pushed, the brother-wants-help button. Mentioning "family values" pushes another button, the recite-respectable-mantra-about-family-values button. What happens when you push those buttons? The deeper question behind that is, What determines what gets written on the little slips of paper? What determines which thoughts get cached and which don't?
For normal humans, the answer is that their social experiences decide what will be on the papers. You can predictably expect to find written on their papers whatever response that they have, |
f7c30b1b-a0f2-4bef-83c1-3831de16557e | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Rohin Shah on reasons for AI optimism
[](http://aiimpacts.org/wp-content/uploads/2019/10/rohin_shah.jpg)Rohin Shah
I along with several AI Impacts researchers recently talked to talked to Rohin Shah about why he is relatively optimistic about AI systems being developed safely. Rohin Shah is a 5th year PhD student at the [Center for Human-Compatible AI](https://humancompatible.ai/) (CHAI) at Berkeley, and a prominent member of the Effective Altruism community.
Rohin reported an unusually large (90%) chance that AI systems will be safe without additional intervention. His optimism was largely based on his belief that AI development will be relatively gradual and AI researchers will correct safety issues that come up.
He reported two other beliefs that I found unusual: He thinks that as AI systems get more powerful, they will actually become *more* interpretable because they will use features that humans also tend to use. He also said that intuitions from AI/ML make him skeptical of claims that evolution baked a lot into the human brain, and he thinks there’s a ~50% chance that we will get AGI within two decades via a broad training process that mimics the way human babies learn.
A full transcript of our conversation, lightly edited for concision and clarity, can be found [here](https://aiimpacts.org/conversation-with-rohin-shah/).
*By Asya Bergal* |
73c3b446-6dd7-45a3-93b1-bc156038267a | trentmkelly/LessWrong-43k | LessWrong | Community norm question: brief text ad signatures
I've written several posts on this site, many of which were promoted to the front page. Some of my writings have been quite popular. Looking through the Google Analytics stats for this site showed that my writing has had at least 50 000 unique views over time. The full total is probably more, since I only looked at the stats of the 1000 most viewed Less Wrong pages.
As I have been looking for sources of side income, I was wondering whether it'd be deemed acceptable if I started signing my posts to the main section with something brief like this:
To see the novel I'm writing, click here. If you liked this post, you may also Flattr me here.
I'm currently thinking about trying to make a living on writing, and being able to do something like this would make it considerably easier to help build a personal brand. Although I wouldn't dare to claim to write as well as Eliezer or Yvain, say, I would like to think that my posts have been valuable for several people. The fact that sixteen of my articles were among the 1000 most viewed LW pages would support this. Being able to get back some of that value would only seem fair to me.
As an additional bonus, currently my LW writing has gotten sidelined as it hasn't seemed that useful for my personal goals. Being allowed to have such ads would make writing on LW more personally useful for me, incentivizing me to spend more time on writing quality posts here.
I can understand not everyone being fully enthusiastic about this, though. For one, several Internet communities are quite stringent about things that might be considered spam. Also, people might also be worried about the fact that LW is currently mostly operating as a gift economy. Letting people make money off their posts directly, such as with Flattr links, might change the community norms in an undesirable direction. Folks such as matt, who are hard at work at improving the technical side of the site, might rightfully feel that they deserve a cut.
So I figured I'd be |
c0e03ef5-3cdf-4afe-b873-77986bba86bc | trentmkelly/LessWrong-43k | LessWrong | Meetup : Melbourne social meetup
Discussion article for the meetup : Melbourne social meetup
WHEN: 20 April 2012 07:00:00PM (+1000)
WHERE: 55 Walsh St, West Melbourne, 3003
EDIT - The meet up has been moved to the Trike Apps office - 55 Walsh St, West Melbourne, 3003. Ben has requested this due to being ill. The time is unchanged.
Melbourne's next social meetup is on Friday, 20th April. We are meeting at the Trike Apps office, 55 Walsh St, West Melbourne. Alternatively, I can give you the address - you can call me at 0432 862 932, or email me at shokwave.sf@gmail.com, or inbox me by clicking on my name. Some form of take-away will be organised for dinner and there will be snacks available. BYO drinks. General catching up and chatting, followed by Resistance and Mafia. We all look forward to seeing you there!
Discussion article for the meetup : Melbourne social meetup |
83ee999a-b097-4e78-b7c9-0b4d3c860545 | trentmkelly/LessWrong-43k | LessWrong | New Eliezer Yudkowsky interview on We Want MoR, the HPMOR Podcast
Folks will probably enjoy the new Eliezer Yudkoswky interview on Harry Potter and the Methods of Rationality. |
9eeac076-e34d-40c7-9bec-7e1d351043f3 | trentmkelly/LessWrong-43k | LessWrong | Getting GPT-3 to predict Metaculus questions
Can GPT-3 predict real world events? To answer this question I had GPT-3 predict the likelihood for every binary question ever resolved on Metaculus.
Predicting whether an event is likely or unlikely to occur, often boils down to using common sense. It doesn't take a genius to figure out that "Will the sun explode tomorrow?" should get a low probability. Not all questions are that easy, but for many questions common sense can bring us surprisingly far.
Experimental setup
Through their API I downloaded every binary question posed on Metaculus.
I then filtered them down to only the non-ambiguously resolved questions, resulting in this list of 788 questions.
For these questions the community's Mean Squared Error was 0.19, a good deal better than random!
Prompt engineering
GPT's performance is notoriously dependent on the prompt it is given.
* I primarily measured the quality of prompts, on the percentage of legible predictions made.
* Predictions were made using the most powerful DaVinci engine.
The best performing prompt was optimized for brevity and did not include the question's full description.
> A very knowledgable and epistemically modest analyst gives the following events a likelihood of occuring:
>
> Event: Will the cost of sequencing a human genome fall below $500 by mid 2016?
> Likelihood: 43%
>
> Event: Will Russia invade Ukrainian territory in 2022?
> Likelihood: 64%
>
> Event: Will the US rejoin the Iran Nuclear Deal before 2023?
> Likelihood: 55%
>
> Event: <Question to be predicted>
> Likelihood: <GPT-3 insertion>
I tried many variations, different introductions, different questions, different probabilities, including/excluding question descriptions, etc.
Of the 786 questions, the best performing prompt made legible predictions for 770. For the remaining 16 questions GPT mostly just wrote "\n".
If you want to try your own prompt or reproduce the results, the code to do so can be found in this Github repository.
Results
GPT-3's MSE w |
c009ecd3-fbbb-4ab5-8552-fe0dfc9e77c4 | StampyAI/alignment-research-dataset/special_docs | Other | ALBA on GitHub
I’ve put up an implementation of [ALBA](https://medium.com/ai-control/alba-an-explicit-proposal-for-aligned-ai-17a55f60bbcf) on GitHub [here](https://github.com/paulfchristiano/alba).
The key method is ALBA(H, n), which hopefully implements an aligned AI. To actually use it, you would need to replace the stub TransparentHybrid with a competent algorithm for [transparent](https://medium.com/ai-control/the-informed-oversight-problem-1b51b4f66b35#.v97m8wh5q) [semi-supervised](https://medium.com/ai-control/semi-supervised-reinforcement-learning-cf7d5375197f#.acfezc2rq) [imitation+RL](https://medium.com/ai-control/imitation-rl-613d70146409#.e28d27gjp).
For now you can mess around with a version of ALBA that uses a very simple learning algorithm — whenever it encounter a novel situation, ask the overseer what to do and memorize their answer for future reference.
I don’t think that actually running the example is especially informative. I wrote the code in order to make the discussion of ALBA more precise, honest, and concrete, not because I thought that anyone would want to actually run it.
TODOs
-----
In reality this implementation of ALBA(H, n) definitely *wouldn’t* be an aligned AI; there are at least half a dozen obviously fatal problems. An immediate goal is to address the obvious problems: to write some code such that *if* we filled in the AI capabilities, then there would at least be *any hope* that the resulting system was aligned. I don’t know how hard this goal is; I think there is a plausible roadmap and it might be relatively easy, or it might take a very long time.
(Realistically, even if there was *any* hope that the system would work, we would still need to improve all of the ingredients further before there was *much* hope that it would work.)
If that works out then we can start to search for non-obvious problems, and argue about more philosophical questions like whether Meta(HCH(·)) correctly implements [capability amplification](https://medium.com/ai-control/policy-amplification-6a70cbee4f34#.8ovjab4mx).
Here is the list of TODOs and FIXMEs in alba.py:
* TODO: Build [powerful AI](https://medium.com/ai-control/imitation-rl-613d70146409#.633sk36pw)
* TODO: Implement [semi-supervised RL](https://medium.com/ai-control/semi-supervised-reinforcement-learning-cf7d5375197f)
* TODO: Implement [transparent RL](https://medium.com/ai-control/the-informed-oversight-problem-1b51b4f66b35)
* FIXME: Prevent [catastrophic failure](https://medium.com/ai-control/learning-with-catastrophes-59387b55cc30) on adversarial inputs. [Adversarial training?](https://medium.com/ai-control/red-teams-b5b6de33dc76)
* FIXME: Ensure RL agent cannot outsmart overseer. Gradually scale up capacity as a function of n?
* FIXME: Prevent failure probability from growing with each iteration. Amplify reliability as well as capability?
* FIXME: Allow Amplify(A) to learn from training data, so it can keep up with the RL agent it is overseeing
* FIXME: Scores in [0, 1] are arbitrary, use [comparisons between different actions instead](https://medium.com/ai-control/optimizing-with-comparisons-c02b8c0d7877)
* FIXME: Use [budgeted HCH](https://medium.com/ai-control/strong-hch-bedb0dc08d4e#.bb5xsy6xj) so that errors can’t result in hangs
* TODO: Figure out whether iterating A → Meta(HCH(A)) can really [get us to arbitrarily powerful agents](https://medium.com/ai-control/policy-amplification-6a70cbee4f34)
How would this be used?
=======================
ALBA is designed to be an essentially drop-in replacement for imitation learning or RL. Of course you would only need to use it when you didn’t have access to a suitable objective function.
A realistic AI system may use RL or imitation learning as components but will generally [have other stuff going on as well](https://medium.com/ai-control/not-just-learning-e3bfb5a1f96e#.hm5jhyh7m). That’s fine, you drop ALBA in for the learning part. For example, in AlphaGo you would drop in ALBA for the policy and value networks, but you’d still have to do MCTS.
Hopefully none of your other code is going to destroy the world, but that’s not our responsibility.
ALBA(H, n) returns an agent A that is aligned with H but much smarter. You can use this agent in the same way that you’d use an episodic RL agent, except that you don’t have to give it a reward. To start an episode you make a fresh copy of A; for example, each time a user visits google.com you might make a new copy of A and then interact with that agent in order to figure out how to process their queries, with calls like A.act(“the user just typed `where should I eat dinner?`, what should happen now?”). In principle you could have very long-running episodes (as with RL), but this would make training more complex.
Instead of specifying a reward directly we specify some overseer H. The system is designed to do what H wants, so picking H is very important. For example, if ALBA were being used in the UI for google.com, H might be implemented by professional trainers employed by Google. If Google wants to make a product that is useful to the users, then it tells the trainers that they should try to think about what the user wants and make decisions appropriately. But the trainers can also balance the user’s interests with other desiderata, like Google’s profit motive or legal constraints. (All of this is analogous to how Google generally balances the desire to be useful to the user against other desiderata.)
Before the first time that ALBA is invoked, it will perform some [pre-training](https://medium.com/ai-control/red-teams-b5b6de33dc76#.5gpktrnyj) to ensure that its behavior is [non-catastrophic](https://medium.com/ai-control/learning-with-catastrophes-59387b55cc30#.x8iwsx5q3), including meeting some minimal bar for quality. [This doesn’t happen in the current implementation; it’s the FIXME about avoiding catastrophes.] During operation it makes decisions immediately, but will continue [collecting data](https://medium.com/ai-control/counterfactual-human-in-the-loop-a7822e36f399#.rpxze4kq5) and training asynchronously in the background. [Training occurs inside the implementation of the AI capabilities in capabilities.py.] During training, H will be invoked constantly.
If H is overseeing a single instance of ALBA that is performing a narrow task, then we can just explain the definition of the task to H. If H is overseeing a range of different agents pursuing a variety of tasks, then we can encode the task definition in the agent’s observations. For example, if we want it to caption an image, rather than simply giving the agent the image and expecting a caption, we could also give the agent the natural language text “please describe what you see in this image,” and perhaps a more detailed description of what kind of explanation we want, and so on. This information will ultimately be used by an aligned overseer to decide how to implement a reward function. As a result, the agent will learn to respond differently to different task explanations. (If the agent is very weak, this might be done simply by memorizing a few different possible task descriptions — we don’t inherently require that the agent can learn to understand language, though the whole system becomes more flexible at that point.)
If we were using an RL agent we would implement a reward function in order to cause the agent to do what we want. We don’t provide a reward function for ALBA, but we can let the agent observe “rewards” to help it learn faster. This can be used as [helpful unlabelled data](https://medium.com/ai-control/semi-supervised-reinforcement-learning-cf7d5375197f), or it can be absorbed by the overseer to improve its understanding of the task. [This doesn’t happen in the current implementation; the FIXME about having the overseer learn from training data.]
Conclusion
==========
I think it’s useful to be able to point to a completely concrete specification of ALBA. I also think that a concrete implementation is a nice setting in which to start fixing the obviously fatal problems, and I’ll be especially excited if/when there is a concrete implementation that *isn’t-obviously-doomed-to-fail*. For now, I hope that having access to the code will help people dig into any details of the scheme that they found ambiguous or unclear in the informal descriptions. |
8caf48ae-4025-4eba-81fa-e58e40cf142a | trentmkelly/LessWrong-43k | LessWrong | University of Oxford, Master's Statistical Science
This article is part of a series of articles on different European master's programs related to artificial intelligence and machine learning.
Summary
This programme runs for 1 year, composed of 2 terms of 8 weeks, followed by a 3-month dissertation over the summer. You typically take 4 courses per term and are able to choose 1 course of the 4. There are 3 week-long practicals throughout the year, and two exams at the end of May. The courses are pretty theoretical, sometimes dry, but mostly good. The course as a whole is pretty good preparation for a PhD in stat-heavy ML, particularly if supplemented with seminars, keeping up-to-date on ML research trends, as well as extra-curricular coding.
The biggest downside of the course is the coding--it’s in R :( However, you don’t have to do much of this and can choose whatever language you want for your dissertation. The course itself doesn’t take too much time, so you should be able to spend a fair bit of time doing the supplementary activities I mention above, as well as maybe doing some research throughout the year with a prof, who are mostly keen to take students.
The biggest upside of the course is the people--academic and otherwise. Being in an Effective Altruism (EA) and AI-safety (AIS) hub has strong benefits, as does being able to attend lots of seminars and talks across departments.
Getting In
The acceptance rate is around 1/15. You’ll need strong grades from a good university. A decent majority of the cohort had research experience before they came. A minority had published. As a UK master’s, I expect that they expect less research experience than similar-quality non-UK programmes.
The Course
* Relevance of courses
* As you might expect, a statistics master’s will approach areas from a theoretical angle. While you gain a deep understanding of the underpinnings of ML, a fair bit of the material is dry (maybe ⅓?), and you won’t be learning the hottest ML--i.e. little discussion of the latest breakthroug |
cda4f658-a141-4e17-9bb3-5a566ae4bb91 | StampyAI/alignment-research-dataset/blogs | Blogs | Multiple Choice Normalization in LM Evaluation
Let $x\_{0:m}$ be the prompt, and $x\_{m:n\_i}$ be the $i$th possible continuation with a token length of $n\_i - m$. There are several ways to use a language model to rank multiple possible continuations to a prompt. Since the language model only gives (log) probabilities for the next token given the context (i.e $\log \mathbb P(x\_i|x\_{0:i})$), there is ambiguity in handling scoring for arbitrary continuations. The following are several possible ways to resolve this problem:
* Unnormalized: The score of continuation $i$ is determined using $\sum\_{j=m}^{n\_i - 1} \log \mathbb P(x\_j|x\_{0:j})$. Intuitively, this is the probability of a generation sampled from the prompt containing the continuation in question. While this is the simplest method, problems arise when there are significant differences in length between different continuations, as longer continuations tend to have lower log probabilities, thus biasing the language model towards picking shorter continuations. This approach is used by [eval harness](https://github.com/EleutherAI/lm-evaluation-harness) in all multiple choice tasks and presented as `acc`.
* Token-length normalized: The score of continuation $i$ is determined using $\sum\_{j=m}^{n\_i - 1} \log \mathbb P(x\_j|x\_{0:j}) / (n\_i - m)$. This approach attempts to normalize for length by computing average log probability per token; however, this approach is not tokenization agnostic, and as such two models with different tokenization that assign the same log likelihood to every single input string will have different token-length normalized scores. This approach is used by [GPT-3](https://arxiv.org/abs/2005.14165) in most tasks. Eval harness does not report this score because it violates the design principle that all tasks should be tokenization independent.
* Byte-length normalized: The score of continuation $i$ is determined using $\sum\_{j=m}^{n\_i - 1} \log \mathbb P(x\_j|x\_{0:j}) / \sum\_{j=m}^{n\_i - 1} L\_{x\_j}$, where $L\_{x\_j}$ is the number of bytes represented by the token $x\_j$. This approach attempts to normalize for length by computing average log probability per character, which ensures that it is tokenization agnostic. This approach is also used by eval harness in all multiple choice tasks and presented as `acc_norm`.
* Unconditional likelihood normalized: The score of continuation $i$ is determined using $\sum\_{j=m}^{n\_i - 1} \log \mathbb P(x\_j|x\_{0:j}) - \log \mathbb P(x\_j)$. Intuitively, this approach measures the amount that the prompt increases the model's probability of outputting each continuation from the probability of the model unconditionally producing that continuation. This approach is used by GPT-3 in select tasks (ARC, OpenBookQA, and RACE), though no justification for why only these tasks in particular use this method is provided other than that this improves performance.
The unnormalized, token-length normalized, and byte-length normalized metrics can be computed without additional LM calls. The unconditional likelihood normalized metric requires an additional LM call to obtain the unconditional likelihood. |
323a6c55-286c-430c-b5e6-2d6bbe90908e | trentmkelly/LessWrong-43k | LessWrong | Two Coordination Styles
In game theory, assumptions of rationality imply that any "solution" of a game must be an equilibrium.* However, most games have many equilibria, and realistic agents don't always know which equilibrium they are in. Certain equilibrium strategies, such as tit-for-tat in iterated prisoner's dilemma, can also be seen in this broader context as coordination strategies: adopting them teaches others to adopt them, because you punish anyone playing some other strategy. In a narrow sense, these strategies solve both the game itself and the equilibrium selection problem. (Technically, such strategies are the evolutionarily stable ones.)
I want to make an informal point about two very different ways this can work out in real life: coordination strategies in which it feels like everyone is fighting to pull the system in different directions but it all cancels out, vs situations where it feels like the coordination strategy is your friend because it saves everyone effort. I believe the second case exists, but it is rather puzzling in terms of the existing literature.
*Different rationality assumptions give different equilibrium concepts; Nash equilibria are the most popular. Correlated equilibria are the second most popular and somewhat more relevant to the discussion here, but I won't get into enough technical details for it to matter.
----------------------------------------
This post made possible by discussions with Steve Rayhawk, Harmanas Chopra, Jennifer RM, Anna Salamon, and Andrew Critch. Added some edits proposed by Elo.
Schelling Negotiations
Schelling discussed agents solving the equilibrium selection problem by choosing points which other agents are most likely to choose based on prominence, and the term Schelling point was coined to describe such likely equilibria. The classic examples revolve around agents who cannot communicate with one another (highlighting the need for guesswork about each other's behavior), but adding the ability to communicate does not |
0e07c176-9db4-46ac-87ad-3d19c4e140a7 | trentmkelly/LessWrong-43k | LessWrong | Piling bounded arguments
TL;DR: A series of valid bounded arguments all arguing the same proposition can only provide as much evidence as that proposition at best, so even if it looks like they're piling up in favor of a higher point, they're only as good as the most central of them.
Epistemic status: Armchair epistemology. Something I noticed + anecdotal evidence.
Introduction
During the LWCW last WE, I played an activity called Steelmanning the devil, where you had to pick a position you disagreed strongly with and steelman it. Participants were scored on their number of valid arguments and fallacies: 1 point for a valid argument, 0.5 point for an unconvincing argument, -1 point for a fallacy (with a lot of room for human judgment so score reflects how convincing the participants were).
I won using arguments which did not prove my point, which I will call piling bounded arguments, and I want to point them out because I think it's easy to be confused online because of them, and they're apparently not common knowledge enough that I wasn't called out during a rationalist competition about arguing.[1]
Boundedly arguing for the devil
What it looks like
The activity went like thus: you named an opinion A you disagreed with (if possible, on a tribal level); you were paired with someone who'd attack A and you'd do your best to steelman it.
You disagree with A. You are informed on the subject, you are confident in your opinion because you've looked at all arguments and found them, on balance, to tend towards ¬A and that arguments for A are generally subtly (or non-subtly) flawed.
This does not mean that there is no valid argument for A. They might just be too weak or too few compared to the arguments against.
Suppose you know one good argument for A (call it P). You share it. Your partner says
"But what about X? That would make P invalid, so you don't have evidence in favor of A."
"Actually, Q=>¬X. This is valid, central, and final. Nothing left to argue about X, it was just plain wrong |
5ed01969-35eb-4775-9f2c-5f218bedede7 | trentmkelly/LessWrong-43k | LessWrong | An Inside View of AI Alignment
I started to take AI Alignment seriously around early 2020. I’d been interested in AI and machine learning in particular since 2014 or so, taking several online ML courses in high school and implementing some simple models for various projects. I leaned into the same niche in college, taking classes in NLP, Computer Vision, and Deep Learning to learn more of the underlying theory and modern applications of AI, with a continued emphasis on ML. I was very optimistic about AI capabilities then (and still am) and if you’d asked me about AI alignment or safety as late as my sophomore year of college (2018-2019), I probably would have quoted Steven Pinker or Andrew Ng at you.
Somewhere in the process of reading The Sequences, portions of the AI Foom Debate, and texts like Superintelligence and Human Compatible, I changed my mind. Some 80,000 hours podcast episodes were no doubt influential as well, particularly the episodes with Paul Christiano. By late 2020, I probably took AI risk as seriously as I do today, believing it to be one of the world’s most pressing problems (perhaps the most) and was interested in learning more about it. I binged most of the sequences on the Alignment Forum at this point, learning about proposals and concepts like IDA, Debate, Recursive Reward Modeling, Embedded Agency, Attainable Utility Preservation, CIRL etc. Throughout 2021 I continued to keep a finger on the pulse of the field: I got a large amount of value out of the Late 2021 MIRI Conversations in particular, shifting away from a substantial amount of optimism in prosaic alignment methods, slower takeoff speeds, longer timelines, and a generally “Christiano-ish” view of the field and more towards a “Yudkowsky-ish” position.
I had a vague sense that AI safety would eventually be the problem I wanted to work on in my life, but going through the EA Cambridge AGI Safety Fundamentals Course helped make it clear that I could productively contribute to AI safety work right now or in the ne |
7703dbcd-82db-450d-8b62-9d7a3acff33b | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Section 7: Foundations of Rational Agency
.mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
.MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
.mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
.mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
.mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
.mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
.mjx-numerator {display: block; text-align: center}
.mjx-denominator {display: block; text-align: center}
.MJXc-stacked {height: 0; position: relative}
.MJXc-stacked > \* {position: absolute}
.MJXc-bevelled > \* {display: inline-block}
.mjx-stack {display: inline-block}
.mjx-op {display: block}
.mjx-under {display: table-cell}
.mjx-over {display: block}
.mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-stack > .mjx-sup {display: block}
.mjx-stack > .mjx-sub {display: block}
.mjx-prestack > .mjx-presup {display: block}
.mjx-prestack > .mjx-presub {display: block}
.mjx-delim-h > .mjx-char {display: inline-block}
.mjx-surd {vertical-align: top}
.mjx-mphantom \* {visibility: hidden}
.mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
.mjx-annotation-xml {line-height: normal}
.mjx-menclose > svg {fill: none; stroke: currentColor}
.mjx-mtr {display: table-row}
.mjx-mlabeledtr {display: table-row}
.mjx-mtd {display: table-cell; text-align: center}
.mjx-label {display: table-row}
.mjx-box {display: inline-block}
.mjx-block {display: block}
.mjx-span {display: inline}
.mjx-char {display: block; white-space: pre}
.mjx-itable {display: inline-table; width: auto}
.mjx-row {display: table-row}
.mjx-cell {display: table-cell}
.mjx-table {display: table; width: 100%}
.mjx-line {display: block; height: 0}
.mjx-strut {width: 0; padding-top: 1em}
.mjx-vsize {width: 0}
.MJXc-space1 {margin-left: .167em}
.MJXc-space2 {margin-left: .222em}
.MJXc-space3 {margin-left: .278em}
.mjx-test.mjx-test-display {display: table!important}
.mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
.mjx-test.mjx-test-default {display: block!important; clear: both}
.mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
.mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
.mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
.mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
.MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
.MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
.MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
.MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
.MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
.MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
.MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
.MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
.MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
.MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
.MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
.MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
.MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
.MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
.MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
.MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
.MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
.MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
.MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
.MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
.MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
.MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
.MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
.MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
.MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
@font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
@font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
@font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
@font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
@font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
@font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
@font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
@font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
@font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
@font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
@font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
@font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
@font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
@font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
@font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
@font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
@font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
@font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
@font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
@font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
@font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
@font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
@font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
@font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
@font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
@font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
@font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
@font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
@font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
@font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
@font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
*This post is part of the sequence version of the Effective Altruism Foundation's [research agenda on Cooperation, Conflict, and Transformative Artificial Intelligence](https://www.lesswrong.com/s/p947tK8CoBbdpPtyK).*
### 7 Foundations of rational agency
We think that the effort to ensure cooperative outcomes among TAI systems will
likely benefit from thorough conceptual clarity about the nature of rational agency. Certain foundational achievements — probability theory, the theory of
computation, algorithmic information theory, decision theory, and game theory to
name some of the most profound — have been instrumental in both providing a
powerful conceptual apparatus for thinking about rational agency, and the
development of concrete tools in artificial intelligence, statistics, cognitive science, and so on. Likewise, there are a number of outstanding foundational questions surrounding the nature of rational agency which we expect to yield additional clarity about interactions between TAI-enabled systems. Broadly, we want to answer:
* What are the implications of computational boundededness (Russell and Subra-manian, 1994; Cherniak, 1984; Gershman et al., 2015) for normative decision theory, in particular as applied to interactions among TAI systems?
* How should agents handle non-causal dependences with other agents’ decision-making in their own decisions?
We acknowledge, however, the limitations of the agenda for foundational questions which we present. First, it is plausible that the formal tools we develop will be of limited use in understanding TAI systems that are actually developed. This may be true of black-box machine learning systems, for instance [[1]](#fn-oHiFxWR8jMBedF5HG-1). Second, there is plenty of potentially relevant foundational inquiry scattered across epistemology, decision theory, game theory, mathematics, philosophy of probability, philosophy of science, etc. which we do not prioritize in our agenda [[2]](#fn-oHiFxWR8jMBedF5HG-2). This does not necessarily reflect a considered judgement about all relevant areas. However, it is plausible to us that the research directions listed here are among the most important, tractable, and neglected (Concepts,n.d.) directions for improving our theoretical picture of TAI.
#### 7.1 Bounded decision theory [[3]](#fn-oHiFxWR8jMBedF5HG-3)
Bayesianism (Talbott, 2016) is the standard idealized model of reasoning under empirical uncertainty. Bayesian agents maintain probabilities over hypotheses; update these probabilities by conditionalization in light of new evidence; and make decisions according to some version of expected utility decision theory (Briggs, 2019). But Bayesianism faces a number of limitations when applied to computationally bounded agents. Examples include:
* Unlike Bayesian agents, computationally bounded agents are *logically
uncertain*. That is, they are not aware of all the logical implications of their
hypotheses and evidence (Garber, 1983) [[4]](#fn-oHiFxWR8jMBedF5HG-4). Logical uncertainty may be particularly relevant in developing a satisfactory open-source game theory ([Section 3.2](https://www.lesswrong.com/posts/8xKhCbNrdP4gaA8c3/sections-3-and-4-credibility-peaceful-bargaining-mechanisms)), as open-source game theory requires agents to make decisions on the basis of the output of their counterparts' source codes (which are logical facts). In complex settings, agents are unlikely to be certain about the output of all of the relevant programs. Garrabrant et al. (2016) presents a theory for assigning logical credences, but it has flaws when applied to decision-making (Garrabrant, 2017). Thus one research direction we are interested in is a theoretically sound and computationally realistic approach to decision-making under logical uncertainty.
* Unlike Bayesian agents, computationally bounded agents cannot reason over the space of all possible hypotheses. Using the the terminology of statistical modeling (e.g., Hansen et al. 2016), we will call this situation *model misspecification* [[5]](#fn-oHiFxWR8jMBedF5HG-5). The development of a decision theory for agents with misspecified world-models would seem particularly important for our understanding of *commitment* in multi-agent settings. Rational agents may sometimes want to bind themselves to certain policies in order to, for example, reduce their vulnerability to exploitation by other agents (e.g., Schelling (1960); Meacham (2010); Kokotajlo(2019a); see also [Section 3](https://www.lesswrong.com/posts/8xKhCbNrdP4gaA8c3/sections-3-and-4-credibility-peaceful-bargaining-mechanisms) and the discussion of commitment races in [Section 2](https://www.lesswrong.com/posts/KMocAf9jnAKc2jXri/sections-1-and-2-introduction-strategy-and-governance)). Intuitively, however, a rational agent may be hesitant to bind themselves to a policy by planning with a model which they suspect is misspecified. The analysis of games of incomplete information may also be quite sensitive to model misspecification [[6]](#fn-oHiFxWR8jMBedF5HG-6). To develop a better theory of reasoning under model misspecification, one might start with the literatures on decision theory under ambiguity (Gilboa and Schmeidler, 1989; Maccheroni et al., 2006; Stoye, 2011; Etner et al.,2012) and robust control theory (Hansen and Sargent, 2008).
#### 7.2 Acausal reasoning [[7]](#fn-oHiFxWR8jMBedF5HG-7)
Newcomb’s problem [[8]](#fn-oHiFxWR8jMBedF5HG-8) (Nozick, 1969) showed that classical decision theory bifurcates into two conflicting principles of choice in cases where outcomes depend on agents' predictions of each other's behavior. Since then, considerable philosophical work has gone towards identifying additional problem cases for decision theory and towards developing new decision theories to address
them. As with Newcomb's problem, many decision-theoretic puzzles involve
dependences between the choices of several agents. For instance, Lewis (1979) argues that Newcomb's problem is equivalent to a prisoner's dilemma played by agents with highly correlated decision-making procedures, and Soares and Fallenstein (2015) give several examples in which artificial agents implementing certain decision theories are vulnerable to blackmail.
In discussing the decision theory implemented by an agent, we will assume that the agent maximizes some form of expected utility. Following Gibbard and Harper (1978), we write the expected utility given an action a for a single-stage decision problem in context x as
EU(a)≜∑jP(a→oj;x)u(oj),(1)
where oj are possible outcomes; u is the agent’s utility function; and → stands for a given notion of dependence of outcomes on actions. The dependence concept an agent uses for → in part determines its decision theory.
The philosophical literature has largely been concerned with *causal decision theory (CDT)* (Gibbard and Harper, 1978) andevidential decision theory (EDT)(Horgan,1981), which are distinguished by their handling of dependence.
Causal conditional expectations account only for the causal effects of an agent’s actions; in the formalism of Pearl (2009)’s do-calculus, for instance, the relevant notion of expected utility conditional on action a is expect[U∣do(a)]. EDT, on the other hand, takes into account non-causal dependencies between the agent's actions and the outcome. In particular, it takes into account the evidence that taking the action provides for the actions taken by *other* agents in the environment with whom the decision-maker's actions are dependent. Thus the evidential expected utility is the classical conditional expectation E[U∣A=a].
Finally, researchers in the AI safety community have more recently developed what we will refer to as *logical* decision theories, which employ a third class of dependence for evaluating actions (Dai, 2009; Yudkowsky, 2009; Yudkowsky and Soares, 2017). One such theory is functional decision theory (FDT) [[9]](#fn-oHiFxWR8jMBedF5HG-9), which uses what Yudkowsky and Soares (2017) refer to as *subjunctive* dependence. They explain this by stating that "When two physical systems are computing the same function, we will say that their behaviors "subjunctively depend’’ upon that function’’ (p. 6). Thus, in FDT, the expected utility given an action a is computed by determining what the outcome of the decision problem would be if all relevant instances of the agent’s decision-making algorithm output a.
In this section, we will assume an *acausal* stance on decision theory, that is, one other than CDT. There are several motivations for using a decision theory other than CDT:
* Intuitions about the appropriate decisions in thought experiments such as Newcomb’s problem, as well as defenses of apparent failures of acausal decision theory in others (in particular, the "tickle defense’’ of evidential decision theory in the so-called smoking lesion case; see Ahmed (2014) for extensive discussion);
* Conceptual difficulties with causality (Schaffer, 2016);
* Demonstrations that agents using CDT are exploitable in various ways (Kokotajlo, 2019b; Oesterheld and Conitzer, 2019);
* The *evidentialist wager* (MacAskill et al., 2019), which goes roughly as follows: In a large world (more below), we can have a far greater influence if we account for the acausal evidence our actions provide for the actions of others. So, under decision-theoretic uncertainty, we should wager in favor of decision theories which account for such acausal evidence.
We consider these sufficient motivation to study the implications of acausal decision theory for the reasoning of consequentialist agents. In particular, in this section we take up various possibilities for *acausal trade* between TAI systems. If we account for the evidence that one's choices provides for the choices that causally disconnected agents, this opens up both qualitatively new possibilities for interaction and quantitatively many more agents to interact with. Crucially, due to the potential scale of value that could be gained or lost via acausal interaction with vast numbers of distant agents, ensuring that TAI agents handle decision-theoretic problems correctly may be even more important than ensuring that they have the correct goals.
Agents using an acausal decision theory may coordinate in the absence of causal interaction. A concrete illustration is provided in Example 7.2.1, reproduced from Oesterheld (2017b)’s example, which is itself based on an example in Hofstadter (1983).
---
**Example 7.2.1** (Hofstadter’s evidential cooperation game)
Hofstadter sends 20 participants the same letter, asking them to respond with a single letter ‘C’ (for cooperate) or ‘D’ (for defect) without communicating with each other. Hofstadter explains that by sending in ‘C’, a participant can increase everyone else’s payoff by $2. By sending in ‘D’, participants can increase their own payoff by $5. The letter ends by informing the participants that they were all chosen for their high levels of rationality and correct decision making in weird scenarios like this. Note that every participant only cares about the balance of her own bank account and not about Hofstadter’s or the other 19 participants’. Should you, as a participant, respond with ‘C’ or ‘D’?
An acausal argument in favor of 'C’ is: If I play 'C’, this gives me evidence that the other participants also chose 'C’. Therefore, even though I cannot cause others to play 'C’ — and therefore, on a CDT analysis — should play 'D’ — the conditional expectation of my payoff given that I play 'C’ is higher than my conditional expectation given that I play 'D’.
---
We will call this mode of coordination *evidential cooperation*.
For a satisfactory theory of evidential cooperation, we will need to make precise what it means for agents to be evidentially (but not causally) dependent. There are at least three possibilities.
1. Agents may tend to make the same decisions on some reference class of decision problems. (That is, for some probability distribution on decision contexts C, P(Agent 1’s decision in context C=Agent 2’s decision in context C) is high.)
2. An agent’s taking action A in context C may provide evidence about the number of agents in the world who take actions like A in contexts like C.
3. If agents have similar source code, their decisions provide *logical* evidence for their counterpart’s decision. (In turn, we would like a rigorous
account of the notion of "source code similarity''.)
It is plausible that we live in an infinite universe with infinitely many agents (Tegmark,2003). In principle, evidential cooperation between agents in distant regions of the universe is possible; we may call this *evidential cooperation in large worlds (ECL)* [[10]](#fn-oHiFxWR8jMBedF5HG-10). If ECL were feasible then it is possible that it would allow agents to reap large amounts of value via acausal coordination. Treutlein (2019) develops a bargaining model of ECL and lists a number of open questions facing his formalism. Leskela (2019) addresses fundamental limitations on simulations as a tool for learning about distant agents, which may be required to gain from ECL and other forms of "acausal trade''. Finally, Yudkowsky (n.d.) lists potential downsides to which agents may be exposed by reasoning about distant agents. The issues discussed by these authors, and perhaps many more, will need to be addressed in order to establish ECL and acausal trade as serious possibilities. Nevertheless, the stakes strike us as great enough to warrant further study.
[Acknowledgements & References](https://www.lesswrong.com/posts/XKWGgyCyGhkm73fhm/acknowledgements-and-references)
---
1. Cf. discussion of the Machine Intelligence Research Institute foundational research and its applicability to machine-learning-driven systems Taylor (2016); Dewey (2017). [↩︎](#fnref-oHiFxWR8jMBedF5HG-1)
2. For other proposals for foundational research motivated by a concern with improving the long-term future, see for instance the research agendas of the Global Priorities Research Institute (Greaves et al., 2019) (especially Sections 2.1 and 2.2 and Appendix B) and the Machine Intelligence Research Institute (Soaresand Fallenstein, 2017; Garrabrant and Demski, 2018). [↩︎](#fnref-oHiFxWR8jMBedF5HG-2)
3. This subsection was developed from an early-stage draft by Caspar Oesterheld and Johannes Treutlein. [↩︎](#fnref-oHiFxWR8jMBedF5HG-3)
4. Consider, for instance, that most of us are uncertain about the value of the 1010th digit of π, despite the fact that its value logically follows from what we know about mathematics. [↩︎](#fnref-oHiFxWR8jMBedF5HG-4)
5. This problem has been addressed in two ways. The first is simply to posit that the agent reasons over an extremely rich class of hypotheses, perhaps one rich enough to capture all of the important possibilities. An example of such a theory is Solomonoff induction (Solomonoff, 1964; Sterkenburg, 2013), in which evidence takes the form of a data stream received via the agent’s sensors, and the hypotheses correspond to all possible "lower semi-computable’’ generators of such data streams. But Solomonoff induction is incomputable and its computable approximations are still intractable. The other approach is to allow agents to have incomplete sets of hypotheses, and introduce an additional rule by which hypotheses may be added to the hypothesis space (Wenmackers and Romeijn, 2016). This sort of strategy seems to be the way forward for an adequate theory of bounded rationality in the spirit of Bayesianism. However, to our knowledge, there is no *decision theory* which accounts for possible amendments to the agent’s hypothesis space. [↩︎](#fnref-oHiFxWR8jMBedF5HG-5)
6. [See Section 4.1](https://www.lesswrong.com/posts/8xKhCbNrdP4gaA8c3/sections-3-and-4-credibility-peaceful-bargaining-mechanisms) for discussion of games of incomplete information and possible limitations of Bayesian games. [↩︎](#fnref-oHiFxWR8jMBedF5HG-6)
7. This subsection was developed from an early-stage draft by Daniel Kokotajlo and Johannes Treutlein. [↩︎](#fnref-oHiFxWR8jMBedF5HG-7)
8. In Newcomb’s problem, a player is faced with two boxes: a clear box which contains $1000, and an opaque box which contains either $0 or $1 million. They are given a choice between choosing both boxes (Two-Boxing) or choosing only the opaque box (One-Boxing). They are told that, before they were presented with this choice, a highly reliable predictor placed $1 million in the opaque box if they predicted that the player would One-Box, and put $0 in the opaque box if they predicted that the player would Two-Box. There are two standard lines of argument about what the player should do. The first is a *causal dominance* argument which says that, because the player cannot cause money to be placed in the opaque box, they will always get at least as much money by taking both boxes than by taking one. The second is a *conditional expectation* argument which says that (because the predictor is highly reliable) One-Boxing provides strong evidence that there is $1 million in the opaque box, and therefore the player should One-Box on the grounds that the conditional expected payoff given One-Boxing is higher than that of Two-Boxing. These are examples of *causal* and *evidential* decision-theoretic reasoning, respectively. [↩︎](#fnref-oHiFxWR8jMBedF5HG-8)
9. Note that the little public discussion of FDT by academic philosophers has been largely critical (Schwarz, 2018; MacAskill, 2019). [↩︎](#fnref-oHiFxWR8jMBedF5HG-9)
10. Oesterheld (2017b), who introduced the idea, calls this "multiverse-wide superrationality’’, following Hofstadter (1983)’s use of "superrational’’ to describe agents who coordinate acausally. [↩︎](#fnref-oHiFxWR8jMBedF5HG-10) |
eb2ae4cd-0687-4ae6-bdb2-e234c124b0ac | trentmkelly/LessWrong-43k | LessWrong | Interpretability with Sparse Autoencoders (Colab exercises)
Update (13th October 2024) - these exercises have been significantly expanded on. Now there are 2 exercise sets: the first one dives deeply into theoretical topics related to superposition, while the second one (much larger) includes a streamlined version of the first one, as well as most of the actual SAE material. This post mostly focuses on the second one (although we do give an overview of both).
This is a linkpost for some exercises on superpostion & sparse autoencoders, which were created for the 3rd iteration of the ARENA program (and greatly expanded on during the 4th iteration). Having spoken to Neel Nanda and others in interpretability-related MATS streams, it seemed useful to make these exercises accessible out of the context of the rest of the ARENA curriculum.
In the ARENA material, these exercises are 1.3.1 and 1.3.2 respectively. The "1" is the transformer interpretability chapter; the "1.3" is the SAEs & Superposition subsection. Although 1.3.1 covers a lot of interesting theoretical topics related to superposition, for most people we recommend 1.3.2 as a fully self-contained introduction to superposition and SAEs.
Links to Colabs for 1.3.1: Exercises, Solutions.
Links to Colabs for 1.3.2: Exercises, Solutions.
----------------------------------------
Summary of material (1.3.2)
Abbreviations: TMS = "Toy Models of Superposition", SAE = "Sparse Autoencoder".
The diagram below shows an overview of section 1.3.2. It's split into 5 parts, each of which covers a different group of topics related to SAEs. You can also see a map of the material in much more detail here.
0️⃣ Toy Models of Superposition is a streamlined version of exercises 1.3.1, with most of the non-crucial stuff cut out (e.g. feature geometry and deep double descent), although you can still probably skip it if you want to get straight to working with SAEs on real language models.
1️⃣ Intro to SAE interpretability is by far the longest section, and covers most of the core mate |
e6f83957-fd87-4d99-a276-843176e651c1 | trentmkelly/LessWrong-43k | LessWrong | Will Donald Trump complete his first term?
PredictIt has the "yes" at 0.80 cents: https://www.predictit.org/markets/detail/5158/Will-Donald-Trump-complete-his-first-term
Polymarket has the "yes" at 0.85 cents: https://polymarket.com/market/will-trump-complete-his-first-term
My intuition is telling me that there's an odds mis-pricing going on and there is a +EV bet to be made.
What do you think? |
9be859b0-7e16-4232-b12a-c83c4aee7d85 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Probability Estimation in Face of Irrelevant Information
-···- - ···---- ····-- -- --�--- ····- � ···-·· ·
Probability Estimation in face of Irrelevant Information
Adam J. Grove
Department of Computer Science
Stanford University
Stanford, CA 94305
grove@cs.stanford.edu
Abstract
In this paper, we consider one aspect of the
problem of applying decision theory to the
design of agents that learn how to make de
cisions under uncertainty. This aspect con
cerns how an agent can estimate probabili
ties for the possible states of the world, given
that it only makes limited observations be
fore committing to a decision. We show that
the naive application of statistical tools can
be improved upon if the agent can determine
which of his observations are truly relevant
to the estimation problem at hand. We give
a framework in which such determinations
can be made, and define an estimation pro
cedure to use them. Our framework also sug
gests several extensions, which show how ad
ditional knowledge can be used to improve
the estimation procedure still :'urther.
1 INTRODUCTION
The problem we consider in this paper is how to es
timate probabilities for states of the world, so that
agents can use the techniques of decision theory to
make decisions under uncertainty. We illustrate this
problem with an example. Suppose we wish to de
sign an agent M whose function is to deliver packages
around town as swiftly as possible. We could program
M with a set of different methods for doing this; for
instance, it can drive between destinations using ei
ther the freeway system or city roads, it can walk, it
can give the package to the postal service to deliver
instead, and so on. Let us ignore the enormous task
of implementing each method, which is, in essence, a
planning problem and beyond the scope of this work.
Here we ask how the agent is to decide between them.
For this example, we might take the following sim
plistic view of the world. First, we suppose that the
time taken to drive depends, in a known way, only
on whether traffic is congested. The time taken to Daphne Koller
Department of Computer Science
Stanford University
Stanford, CA 94305
daphne@theory.stanford .edu
walk depends (again, in a known way) on the weather
conditions. Posting the package takes constant time.
Now we ask M to deliver a particular package. Af
must commit to a method before finding out about
the traffic or whether it will rain. If M can associate a
probability with each relevant possibility, then it can
calculate expected time for each method and decide
accordingly.
We will assume that M has been in situations like
this before (presumably this is the case once it has
been operational for a while), and so it can estimate
probabilities from stored observations. However, M
might actually know a lot about the current situation:
it could know the time, the day of week, the season,
today's weather forecast, the package's weight, there
cipient, whether the car has been serviced recently, the
price of a postage stamp, and so on. What we really
want is the probability of each of the events of interest
(such as, it is sunny but traffic is light) conditioned on
what is known. This presents a problem because, once
we take all of this knowledge into account, most of M's
previous data no longer pertains directly. The main is
sue studied in this paper is how we can decide what
information can be safely ignored as being irrelevant,
so improving the quality of the estimated probabilities.
The agent M might have very little data directly ap
plicable to estimating the chance of "fine weather and
light traffic" given all he knows, because it is likely lvf
hasn't been in an identical situation often before. But
we know that a lot of M's knowledge-for instance,
the nature of the package and the condition of the
car-has nothing at all to do with the weather or road
conditions. If we estimate probabilities conditional on
just relevant data, such as the forecast and the time
of day, there will be many more observations that can
be used.
The model which stands at the heart of classical de
cision theory, and on which our work is based, is the
decision matrix (see [Savage, 1954] for definitions, and
Section 5 for some discussion of related work in AI).
As explained in Section 2, applying this technique re
quires the estimation of the probabilities of the various 127
128 Grove and Koller
states of the world that the agent considers possible.
Using these probabilities, the agent can estimate the
expected utility for each alternative action, and choose
the one that maximizes that quantity. Unlike many
previous works (e.g. [Simon and Kadane, 1975]), we do
not assume that these probabilities are made available
by an external source. More realistically, we assume
that the agent uses its own experience as the major
source of information. The agent will thus learn from
experience, by gradually refining its estimates.
Our method for estimating probabilities, which is
based on a procedure that attempts to discover ir
relevant attributes, is described in Section 3. Some
extensions are outlined in Section 4. Our technique
combines concepts reminiscent of probabilistic refer
ence hierarchies (see, for example, [Bacchus, 1988])
with statistical tools. It thus enables using statistical
data, as well as less precise notions about relevance
that the designer might have.
2 THE UNDERLYING MODEL
The decision-making module takes a decision problem,
creates a decision-matrix, and uses that matrix to de
cide on a course of action. Each row of the decision
matrix is a possible action under consideration by the
agent. For example, agent !If's actions might con
sist of: drive on a freeway, drive on the city roads,
walk, and send by mail. Of course, these high-level
actions usually represent complex plans consisting of
many atomic steps; the model we are using ignores
the planning problem of how to determine these steps.
The columns of the matrix are possible states of the
world, where each state has an associated probability
(the probability that it holds in this particular decision
situation). These two components of the matri.." will
be described later in this section. The elements of the
matrix are the agent's outcomes for each action/state
pair. It is a well-known result of decision theory (see
[Savage, 1954)) that in many cases, the agent's pref
erence ordering on outcomes can be expressed by nu
merical utilities. In the particular context of intelligent
agents, the utility will often be expressed in terms of
certain parameters, such as time, fuel consumption, or
money. For example, we might use time as our mea
sure of utility for M. In our paper, we assume that
the matrix entries are utilities, and are given in ad
vance. Once the entire decision matrix is available,
the agent simply chooses the action with ma."imum
expected utility.
2.1 STATES AND EVENTS
We adopt a framework in which the agent observes
and reasons about the world using a fixed set of at
tributes, A = { A1, ... , An}, which take on values in
a finite space. For example, the attribute "day of
the week" has a natural set of values; an attribute such as "weather" would be partitioned into values
(e.g. raining, cloudy, fine), where the granularity of
the partition will depend on the agent's needs. This
vocabulary must be chosen carefully, as it greatly af
fects the performance of the decision-making module
The attributes should be chosen to be, in some sense,
independent of each other. See Section 4 for further
discussion.
The most specific assertion we can make about the
world is to announce the value of each A E A. This
exactly determines the world as far as the agent's vo
cabulary allows it to differentiate. We therefore define
a state to be an assignment of a value to every at
tribute in A. In the decision-theoretic paradigm, the
agent's uncertainty is modeled via the existence of sev
eral states that the agent believes the world could be
in.1 In general, some of the attributes in A will have
no connection to the decision at hand (i.e. will not
affect the outcomes of the contemplated actions). For
example, although M might have an attribute describ
ing the last time the car was serviced, the value of this
attribute would not be relevant to the time it takes
M to deliver a certain package. We define an event E
to be a list of values for a subset A' of the attributes
in A. We say that an event E obtains when the true
state agrees with Eon all attributes in A'.
We assume that the columns of the decision matrix are
all events which contain values for some specific subset
A' of A. Ideally, these attributes should be those that
have some connection with the actions or the decision
under consideration. The smaller A' is, the easier it
is to make the decision2 In our example, A' might
consist of the attributes "weather" and "traffic den
sity", and the column events would be all the possible
assignments of values to these two attributes.
The initial information I denotes the set of attribute
value pairs that the agent observes in a particular de
cision making situation, i.e., before an action is chosen
and executed. It should be clear that I is also an event.
2.2 PROBA BILITY DISTRIBUTION
The decision-theoretic model we are using here as
sumes the existence of some objective probability dis
tribution on the possible states. That is, let W be
the set of all possible states of the world. We are
assuming that W has the additional structure of a
(presumably unknown) probability space (W, 1r). It
is unlikely to be true, in any meaningful sense, that
the actual state of the world is a random draw from
1This is similar to the familiar concept of "possible
worlds".
2For choosing A', all we need is some, possibly incom
plete, knowledge about which attributes are "relevant" to
an action. It is not necessary to know how or why a certain
attribute affects the consequence of the action, or even be
certain that it really does.
Probability Estimation in Face of Irrelevant Information 129
some probability distribution, but the probabilistic
model is often a good approximation to the intricately
causal way the world actually works. For further dis
cussion, see, for example, [Cox, 1961; Jaynes, 1968;
Savage, 1954).
In our model, a decision-making situation evolves as
follows. At the time the agent learns that he is to make
a decision, the actual state of the world is regarded as
being randomly chosen according to (W, 1r). The agent
has the capability of observing the values of some of
the attributes, and so gains some initial information I
about the chosen state. It then chooses an action (i.e.,
makes a decision), and afterwards, perhaps because of
the action's execution, it learns more about what the
world was like at decision time. That is, it learns the
values of more attributes.
2.3 DATA COLLECTION
We assume that the agent has a database 1) of obser
vations relating to past experiences (the integration of
other types of data into our model will be discussed
briefly in Section 4). To be more specific, we as
sume that every data point D E 1) actually arose from
some earlier decision-making episode. Thus D con
tains, among possibly other things, the values for the
attributes that were observed before and during the
decision process, i.e., the agent's information about
the state of the world holding at the time.
In order to simplify the model, we might make a com
plete observability assumption: the agent always ob
serves the value of every attribute in A. Consider
the following example, which illustrates what can go
wrong without some such requirement. Suppose that
M is sometimes told about baseball games taking place
in the city, and that when there is a baseball game
M usually decides to walk (perhaps because baseball
games generally take place when the weather is fine).
Suppose, in violation of complete observability, that if
the agent chooses to walk it does not find out about
the traffic conditions. Then if the traffic density is
in fact heavier on days in which a game takes place,
M's estimate for the probability of heavy traffic will
be lower than the true value, because most of its obser
vations will be taken on days when there is no game.
In general, complete observabil ity avoids this and sim
ilar errors because it implies that every observation
in 1) truly is a random sample from (W,1r), free of
unwanted bias.
In practice, complete observability can be weakened to
independent observability, which says that the set of
attributes the agent gets to learn about is determined
independently of both the actual state of the world,
and of any decision the agent takes. In the above ex
ample this was violated, because whether or not the
traffic was observed depended on whether the agent
decided to walk, and it was this that induced bias. Of course, the independent observability requirement by
itself offers no guarantee that we ever see enough data
to estimate all the required probabilities. So it is also
necessary to assume that, whenever A' defines the set
of events for some decision the agent might be asked to
make, then the chance of observing this set should be
nonzero. Even this requirement can be weakened. For
instance, if two attributes are rarely observed together,
appropriate assumptions about conditional indepen
dence can be used so as to still permit accumulation
of sufficient historical data.
In subsequent sections, we shall simplify the presen
tation by stating our results and techniques in terms
of the complete observability assumption only. How
ever, the extensions to the weaker, but more realistic,
conditions are straightforward.
In practice, the most restrictive consequences of our
model are as follows. First, the requirement that the
information observed is an event amounts to assuming
that the agent either identifies, without uncertainty,
the value of an attribute or else learns nothing at all
about it. But we note that if the set A is chosen
well, this assumption should cause relatively little dif
ficulty. The other problem with our model relates to
the amount of data stored: there is little obvious scope
for data summarization or compression. Currently, our
estimation procedure requires every observation to be
remembered (or, only slightly better, remember counts
for observations that occur frequently). Significant
improvements are likely to depend on domain-specific
structure.
3 THE ESTIMATION PROBLE M
Let us review the probability estimation problem. We
have some initial information I. We have determined
a list of events which are the columns of the decision
matrix, and wish to estimate p = Pr(EII) for each
event E. Of course, how best to form these estimates
is a problem in statistics.
One simple and theoretically sound estimate of p is
simply the proportion of data points agreeing with E,
among all points that agree with I. This estimate is
"good" in several ways: for instance, it is unbiased
(i.e., the expected value is exactly p) and its variance
decreases to zero as the number of relevant data points
grows. The problem is that the number of relevant
data points may not grow very quickly because this
estimate uses only those observations which agree ex
actly with I. But perhaps situations matching I have
not been encountered very often. For instance, if time
and date are part of I, there will be no relevant his
torical data at all. We conclude that this simple esti
mation procedure is often impractical.
To salvage the approach to decision making we are
looking at, we need to be much more clever about how
130 Grove and Koller
we estimate probabilities. The main result of this sec
tion is a technique for probability estimation which
can yield substantially better results than the above.
It does this by providing a framework which can cap
ture and make use of additional information we have
about the structure of the world (see Section 4). The
underlying idea is the observation that the less specific
the information in I, the more useful data points we
will have. Therefore, the estimate would improve if
we could (justifiably!) ignore some of the attributes
mentioned in I.
It seems to be often true that, in any given context,
only a few attributes will be relevant to whether some
event E occurs. Consider the example in the introduc
tion, where I records, among other things, which day
of the week it is. If the attributes in E all refer to nat
ural phenomena, such as the the weather conditions,
we would expect the day of the week to be irrelevant
to-and, in a sense which is easy to make precise, in
dependent of-whether E occurs. In this case, the best
estimate of p would pool data for all days, even though
this ignores some of the knowledge contained in I. In
general, it is not reasonable to require all such infor
mation about relevance to be supplied ahead of time.
In the example, we would want the estimation proce
dure to find out for itself whether the day of the week
is unimportant.
We begin by looking at the base case of our technique.
Suppose I includes the value of some attribute A. Let
h, h, ... , Ik be all events which are just like I, except
possibly with respect to the value of A; we may as
sume that h = I. Intuitively, we can pool data only
if our knowledge about A is irrelevant to p. More for
mally, we ask whether the conditional probabilities Pi
(i.e., Pr(EII;)) are the same for all i. Our procedure
is to test whether this independence is plausible, and
then use either the pooled or non-pooled estimate as
appropriate. Both the test and the estimate itself can
make use of all the observed data in 1).
In the following, let N; be the number of observations
in 1) agreeing with I;. Let Pi be the proportion of
these observations that do in fact agree with the values
specified in E. We estimate p either as p1 or else as
the pooled estimate p = 0::"'1 p;N;)/N, where N =
2:::"'1 N;. Note that p1 is simply the direct estimate
which was mentioned earlier. We decide which of these
two possibilities to use on the basis of a hypothesis test
(the hypothesis being that p; =Pi> for all i,j.)
One relatively simple hypothesis test we can use
for this is the x2 test, which is discussed in most
statistics texts (such as, [Larsen and Marx, 1981;
Sachs, 1982]). The test is based t•n the value X2 =
2:::"'1 (p;-fi? N;j(p(1-p)). If the hypothesis (equal
Pi) is true, and the Ni are not too small,3 then the dis-
3 A frequently stated rule of thumiJ is that N;p; should tribution of X2 is very well approximated by the x2
distribution with k-1 degrees of freedom. In order to
perform the test, we must choose some small a > 0,
which becomes the chance of not pooling data when it
really would have been permissible. We accept the hy
pothesis and use the pooled estimate just if X2 < ca,
where c" is such that the chance of a random sam
ple from x2 exceeding c" is a. The value c" can be
found from tables. It is generally desirable to have a
very small, but note that as a decreases the chance of
incorrectly deciding to pool data when this is not jus
tified grows. Later we state two asymptotic properties
of our estimation procedure, whose proof assumes that
a is 1/ Nd, for some d > 1. That is, as the sample size
increases we should tolerate less chance of deciding in
correctly not to pool data. It turns out that, so long
as a grows smaller no faster than this, the chance of
pooling inappropriately also diminishes rapidly.
To recap, the general idea of our procedure is to test for
independence and then use the estimate suggested by
the result of the test. The hypothesis test can be done
in many ways, and we have suggested one possibility,
the x2 test. We chose this test because its simplicity
facilitates the analysis.
This analysis is important because the procedure uses
the same data for both the independence test and for
the actual estimate. In this way, we can hope to make
the best use of scarce data. But reusing sampled data
destroys the independence between the outcome of the
hypothesis test and the estimate used, and so we must
check that the process as a whole gives us a useful re
sult. For example, it is easy to see that both the pooled
estimate p and the simple estimate p1 are unbiased if
the hypothesis of equal Pi is in fact true. But it does
not follow just from this that the composite estimate
is unbiased.4 Another related issue concerns the esti
mate's variance: intuitively, we only pool data if all
the p; are approximately the same, and so it might
seem that the pooled estimate is only used when it
provides little additional information over p1 anyway.
It turns out that neither of these problems arise: the
estimate we give is asymptotically unbia�ed, and has
asymptotic variance that can be much smaller than
p1. In other words, at least for large N, our estimate
is indeed very likely to be close to the true value. Fur
thermore, if the hypothesis of equal p; is true, then
our estimate has smaller variance than the simple un
pooled estimate p1 and so is likely to be much closer to
p. The formal statement of these results is contained
be larger than 5, for all i.
4To illustrate the possible problems, suppose that the
test is such that the hypothesis of equal p; is slightly more
likely to be accepted when the observations satisfy p < fj,
than it is otherwise. Then this would bias our estimate.
Because we reuse data, the test does get to see the actual
values of the estimates p and fj,, and so such behavior
cannot be ruled out without deeper analysis.
Probability Estimation in Face of Irrelevant Information 131
in the following two theorems (whose proofs are too
long for inclusion here) .5
Theorem 3.1: If in fact p; = p for all i, then the
estimate we give has mean J.l and variance cr2 such
that J.l-->P and cr2--+p(1 -p)/ N as N ---+oo.6 We
note that this asymptotic variance is the best that can
be achieved by any unbiased estimate, even if we know
for certain that the hypothesis of equal probabilities is
true.
Theorem 3.2: If in fact p; i: Pi for some i, j then
the estimate we give has mean J.l, and variance cr2 such
that J.l-->Pl and cr2-->p1 (1-pt)f N1 as N ---+00. We
note that this asymptotic variance is the smallest pos
sible amongst unbiased estimates of p; (given that ob
servations relating to I;, for i -j: 1, are regarded as
being not informative about p1).
Although these theorems give asymptotic results only,
it seems very likely that the procedure will work well
for far smaller sample sizes than were required in our
proof. Proving a precise claim about this would be dif
ficult. Instead, we have programmed the technique to
run on simulated data, and the results there did con
firm this expectation. When the data was generated
for each class using the same underlying probability,
the decision was made to pool data most of the time.
In one typical experiment, 200 data points were succes
sively generated for each of five events, and on average
the procedure declined to pool data less than 5% of the
time. If the probabilities differ between classes, even
by relatively small amounts-and note that the closer
the probabilities are to being equal, the less damage is
done by incorrect pooling-our procedure rapidly dis
covered this. In one experiment, where the difference
between all probabilities was less than or equal to 0.1,
the procedure apparently stabilized on a decision not
to pool after each class had accumulated about 150
data points. A similar experiment where the proba
bilities differed by up to 0.3 stabilized after about 15
data points per event on average.
The procedure so far will, in effect, decide whether to
ignore one particular attribute of I. In general, sev
eral attributes of I may turn out to be irrelevant. Our
technique extends easily to such cases. Suppose that
we have decided to ignore some attribute A of I (using
a test like that just suggested). That is, we have ac
cepted the hypothesis that P1·(E\I) = Pr(E\I') for all
5Note that if we did not reuse dat�, the proof of these
theorems would be nearly trivial (becaase then the hypoth
esis test would be certain to be independent of both p and
P• ). Furthermore, the results would be somewhat tighter;
e.g., the composi te estimate would be unbiased even for
finite sample sizes.
6This is not quite correct as stated, because N can grow
without bound even as some No stays small. The conver
gence we have in mind here is that, as N tends to infinity,
each N; must be bounded below by Nc for some c > 0. I' that are like I except for the value of A. But from
this it also follows that Pr(E\I) = Pr(E\(1 -A)),
where by I -A we mean the event formed from I by
omitting A and its value. We have thus reduced our
problem to finding a good estimate of the latter prob
ability. The earlier procedure-looking for irrelevant
attributes-is immediately applicable again. In this
way, we can achieve a substantial increase in the qual
ity of the estimate. Note that we do not require that
the attributes be considered in any particular order.
4 JUSTIFICATION AND
EXTENSIONS
The success of our technique depends on whether the
vocabulary of attributes used to define events really
reflects the way the world works. For example, the at
tributes A, = day of the week and A2 = the weather
seem to be fairly independent of each other; there are
many contexts where just one of these is relevant. On
the other hand, consider A�, which tells us the day of
the week if it is Monday or Wednesday and the weather
otherwise, and A� which tells us the weather on Mon
day or Wednesday, and the day otherwise. It is not
easy to imagine a context where just one of {A�, A�}
is relevant. Both these sets of attributes are equally
informative for describing what the world is actually
like. Our judgment that {A1, A2} is better seems to be
based on knowledge we have about the causal struc
ture of the world? Our technique is a framework that
allows such knowledge to be usefully incorporated into
the decision making process. Equally important is that
we do not rely on a precise or accurate statement of
this knowledge.
Sometimes we have additional knowledge, beyond just
a feeling about what a suitable attribute vocabulary
should be. A feature of our technique is that it al
lows easy extensions to cope with many types of extra
information. For instance:
• If we are able to provide actual probabilities di
rectly, there is nothing preventing them being
used; the estimation procedure can be bypassed
when not needed. Similarly, the method can be
modified to incorporate statistical data from di
verse external sources.
• If we know that some particular attributes can be
ignored in certain contexts, then the hypothesis
test is redundant and can be omitted. In partic
ular, if some restrictions on the possible relation
ships between attributes are given to the agent
(for example, as a reference hierarchy), this can be
used in our process. By avoiding the hypothesis
7This is reminiscent of the well-known "grue/bleen"
paradox ([Goodman, 1955]), which concerns the attribute
vocabulary appropriate for inductive inference.
132 Grove and Koller
test, we gain computational efficiency and elimi
nate the chance of error.
• Suppose we know that if some attribute is rele
vant, then it must affect probabilities in a partic
ular way. As an example, I am not sure whether
the probability of traffic congestion on a particu
lar highway depends on which day of the week it
is. I know that, if the day of the week is in fact rel
evant, then this probability is lower on weekends.
This knowledge suggests using a different test for
independence. We test the hypothesis that the
probability is independent of the day, against the
alternatives (which have different probabilities f01'
each day, but definitely lower on Saturday and
Sunday). We omit details of such a test here. In
general, whenever our knowledge can restrict the
possible alternatives, a hypothesis test can achieve
the same confidence using less data.
• So far, we have regarded the classes considered
for pooling as being implicitly defined by the
attribute vocabulary. However, our knowledge
about the domain may suggest other classes as
well. In our earlier example, to estimate the
chance of heavy traffic congestion on a Tuesday
it might be useful to consider pooling data over
just weekdays (Monday to Friday), as well as over
the class of all days. It is even possible to use
statistical procedures to suggest useful classes, on
the basis of previously collected data (but then
we must be careful to use a different set of data
for the hypothesis test and estimation procedure,
because the results would be statistically invalid
otherwise).
Finally, we note that the correctness of our technique
depends on the stability over time of the underlying
probability distribution. If this cannot be assumed it
would be sensible to ignore or discount older obser�a
tions. There are several standard ways this might be
done. Nevertheless, it is clear that robustness against
changes in the underlying distribution can only be ob
tained at the price of slower or Jess accurate learning.
5 COMPARISON TO OTHER
WORK
Although many techniques of decision theory have
been utilized in artificial intelligence, the decision ma
trix paradigm of separating the probabilities of states
from the utilities has been relatively ignored. Many
researchers who adopt the concept of maximizing ex
pected utility (see [Etzioni, 1989b; Horvitz, 1988;
Wellman, 1990; Russel and Wefald, 1988]) compute
the expected utility for each action directly. A sepa
rate computation of utilities has the major advantage
of allowing additional information about utilities and
probabilities, arising from different sources, to be used. The description of a state may be detailed enough so
that the utility of an action at that state can be com
p�ted using knowledge about causality that the agent
might have. For example, the agent might know that
when the state of the world is such that there is light
traffic and the weather is good, then the action of driv
ing ten miles on a freeway must take about ten min
utes, because the average velocity in those conditions
is 55 miles per hour. Also, by dividing the estimation
process into two stages, more historical data will be us
able. For example, the agent might conclude that the
exact day of the week (say Friday) is relevant to the
probability of having heavy traffic, and will therefore
use only the historical data about Friday to compute
it. But heavy traffic also occurs on other days ( al
tJ:ough with different probabilities), so that the agent
will be able to use all that additional data to compute
the expected driving time given heavy traffic. This
leads to more accurate estimates.
'Yhile some research in AI has adopted this separa
tiOn of probabilities and utilities (notably [Haddawy
and Hanks, 1990; Simon and Kadane, 1975]), the
problem of estimating the probabilities in the face of
too much initial information has not, to our knowl
edge, been attacked directly. Some works [Simon and
Kadane, 1975) simply assume that the probabilities
are known in advance. Others (e.g. [Bundy, 1984;
Lee and Mahajan, 1988]) suggest the concept of sam
plin�, but do not discuss which class to sample. Ren
delll�endell, 1983) deals �ith the concept of sampling
on different classes, but m the very limited context
of search trees. Etzioni's work [Etzioni, 1989a) on es
ti.mating u�ilities is based on machine learning tech
mques, which attempt to discover classes over which
the utility is homogeneous. Such homogeneity usu
ally arises due to a deterministic relationship between
the properties of the class and the utility (such as the
driving time given light traffic described above). These
techniques do not carry over to estimating probabili
ties, because the only way to achieve homogeneity in
classes of binary values is to have the entire class be
all zeros or all ones. I.e., the class will be such that
it deterministically forces the truth value of the event.
Typically, it is impossible to find an attribute language
precise enough to define such classes.
A different approach to finding the right class for esti
mating probabilities is to treat the problem of inferring
the independence structure as a separate task. For ex
ample, [Fung and Crawford, 1990) use techniques sim
ilar to ours-classical statistics, and in particular, the
x2 test-in a system which infers qualitative structure
from data, modeling this structure as a probabilistic
network . Once constructed, the network can be used
for several purposes, such as estimation. The major
drawback of this technique is that a separate data set
is required for the construction of the network. It is
not clear how this technique can be safely extended
to reuse data. Therefore, larger amounts of data will
Probability Estimation in Face of Irrelevant Information 133
be needed. Our approach is also better in situations
where new data is being constantly accumulated, be
cause the new information could cause us to change
our decision as to the relevance of certain attributes.
[Fung and Crawford, 1990] also show how to find the
smallest possible set of relevant attributes. This proce
dure is computationally expensive and relies on strong
assumptions about the relationships among the at
tributes. These assumptions also prevent the proce
dure in [Fung and Crawford, 1990] from being effi
ciently used in our framework, because each decision
situation will need to be investigated separately. This
eliminates the computational advantage of computing
the entire independence structure simultaneo usly. Our
approach eliminates the attributes one by one, in an
arbitrary order. While not guaranteed to find the min
imal set of relevant attributes, this technique is much
faster and requires no assumptions.
We conclude this section by comparing our methods
to the Bayesian approach. It should be noted that,
if prior probabilitie s are available, Bayesian updating
(see [Jaynes, 1968]) can, in a sense, replace the x2 test
described in Section 3. \Ve have chosen not to ussume
the existence of prior probabilities, and therefore use
a technique from classical statistics. A work similar
in outlook to ours, which deals with a different prob
lem using Bayesian techniques is Pearl's work about
hierarchies of hypotheses [Pearl, 1986].
6 CONCLUSION
We have investigated the problem of estimating the
conditional probability of a state given some initial
information I, based on a database of observations.
This problem is straightforward when there is plenty of
data. However, in many situations, there is little data
that exactly matches I. Our main result discusses how
to utilize the available data in order to decide which
attributes in I are irrelevant, and how to use the in
formation about irrelevance to improve the estimate's
quality. A feature of our approach is that it uses all
the available data for both this decision and for the
actual estimation.
We have discussed in detail the assumptions that are
required to make our approach sound. For example,
the model is simplified by the (commonly made) as
sumption we call complete observability, that is, that
all data points contain observations about every at
tribute. However, we discuss a relaxation of this, in
dependent observability, which is far more realistic yet
still permits efficient estimation.
The idea behind our approach to estimation is not
limited to finding conditional probabilities. For in
stance, in the context of decision theory which moti
vates the work in this paper, another important appli
cation would be the estimation of utilities; we believe that this is likely to be a straightforward extension of
the present work.
The most important factor in the success of our ap
proach will be the quality of the attribute vocabulary
that the agent uses to describe the world. Whether
made explicit or not, this theme recurs throughout AI.
In our approach, this issue is prominent, and we may
hope that our more technical results and discussion
will serve to shed some light on this fundamental is
sue.
Finally, we note that one of the advantages of our
framework is that it can be extended in several di
rections, to make use of other knowledge aside from
raw observational data. A few suggestions towards
this were described in Section 4, but clearly this does
not exhaust all the possibilities.
Acknowledgments
The authors would like to thank Joseph Halpern for
comments and discussions relating to this paper.
Some of this work was executed while both authors
were employed at IBM Almaden Research Center, 650
Harry Road, San Jose, California 95120-6099. The
first author is also supported by an IBM graduate fel
lowship.
References
[Bacchus, 1988] F. Bacchus. Representing and reason
ing with probabilistic knowledge. PhD thesis, Uni
versity of Alberta, 1988. Also issued as Waterloo
University Technical Report CS-88-31.
[Bundy, 1984] A. Bundy. Incidence calculus: a mech
anism for probabilistic reasoning. Technical Report
216, University of Edinburgh Dept. of Artificial In
telligence, 1984.
[Cox, 1961] R. T. Cox. The algebra of probable infer
ence. Baltimore: The Johns Hopkins Press, 1961.
[Etzioni, 1989a] 0. Etzioni. Hypothesis filtering: a
practical approach to reliable learning. In Proceed
ings of the Fifth International Conference on Ma
chine Learning, 1989.
[Etzioni, 1989b] 0. Etzioni. Tractable decision ana
lytic control: an expanded version. Technical Re
port CMU-CS-89-119, Carnegie Mellon University,
1989.
[Fung and Crawford, 1990] R. M. Fung and S. L.
Crawford. Constructor: a system for the induction
of probabilistic models. In Proceedings of the Na
tional Conference on Artificial Intelligence ( AAAI-
90), pages 762-769, 1990.
[Goodman, 1955] N. Goodman. Fact, fiction, and
forecast, chapter iii. Harvard University Press, 1955.
134 Grove and Koller
[Haddawy and Hanks, 1990] P. Haddawy and
S. Hanks. Issues in decision-theoretic planning:
symbolic goals and numeric utilities. In Proceed
ings of the 1990 DARPA Workshop on Innovative
Approaches to Planning, Scheduling, and Control,
1990.
[Horvitz et a/., 1988] E. J. Horvitz, J. S. Breese, and
M. Henrion. Decision theory in expert systems and
artificial intelligence. International Journal of Ap
proximate Reasoning, 2:247-302, 1988.
[Horvitz, 1988] E. J. Horvitz. Reasoning under vary
ing and uncertain resource constraints. In Proceed
ings of the National Conference on Artificial Intel
ligence ( AAAI-88 ), pages 111-116, 1988.
[Jaynes, 1968] E. T. Jaynes. Prior probabilities. IEEE
Transactions on Systems Science and Cybernetics,
4:227-241, 1968.
[Larsen and Marx, 1981] R. J. Larsen and 1\I. L.
Marx. An introduction to mathematical statistics
and its applications. Prentice-Hall, 1981.
[Lee and Mahajan, 1988] K. F. Lee and S. Mahajan.
A pattern classification approach to evaluation func
tion learning. Artificial Intelligence, 36, 1988.
[Pearl, 1986] J. Pearl. On evidential reasoning in a
hierarchy of hypotheses. Artificial Intelligence, 28,
1986.
[Rendell, 1983] L. Rendell. A new basis for state-space
learning systems and a successful implementation.
Artificial Intelligence, 20, 1983.
[Russel and Wefald, 1988] S. Russel and E. Wefald.
Decision-theoretic control of reasoning: general the
ory and an application to game playing. Technical
Report UCB/CSD 88/435, University of California
at Berkeley, 1988.
[Sachs, 1982] L. Sachs. Applied statistics. Springer
Verlag, 1982.
[Savage, 1954] L. J. Savage. Foundations of statistics.
John Wiley & Sons, 1954.
[Simon and Kadane, 1975] H. A. Simon and J. B.
Kadane. Optimal problem solving search: ali-or
none solutions. Artificial Intelligence, 6, 1975.
[Spiegelhalter, 1986] D. J. SpiegeL1alter. Probabilis
tic reasoning in predictive expert systems. In L. N.
Kana] and J. F. Lemmer, editors, Proceedings of the
Second Workshop on Uncertainty in Artificial Intel
ligence, pages 47-68. Amsterdnm, North Holland,
1986.
[Wellman, 1990] M. P. Wellman. Formulation of
tradeoffs in planning under uncertain ty. Pitman,
London, 1990. |
69446d50-de1a-4ad1-b803-343ed31bce12 | trentmkelly/LessWrong-43k | LessWrong | A brief tutorial on preferences in AI
Preferences are important both for rationality and for Friendly AI, so preferences are a major topic of discussion on Less Wrong. We've discussed preferences in the context of economics and decision theory, but I think AI has a more robust set of tools for working with preferences than either economics or decision theory has, so I'd like to introduce Less Wrong to some of these tools. In particular, I think AI's toolset for working with preferences may help us think more clearly about CEV.
In AI, we can think of working with preferences in four steps:
1. Preference acquisition: In this step, we aim to extract preferences from a user. This can occur either by preference learning or by preference elicitation. Preference learning occurs when preferences are acquired from data about the user's past behavior or past preferences. Preference elicitation occurs as a result of an interactive process with the user, e.g. a question-answer process.
2. Preferences modeling: Our next step is to mathematically express these acquired preferences as preferences between pairwise choices. The properties of a preferences model are important. For example, is the relation transitive? (If the model tells us that choice c1 is preferred to c2, and c2 is preferred to c3, can we conclude that c1 is preferred to c3?) And is the relation complete? (Is any choice comparable to any other choice, or are there some incomparabilities?)
3. Preference representation: Assuming we want to capture and manipulate the user's preferences robustly, we'll next want to represent the preferences model in a preference representation language.
4. Preferences reasoning: Once a user's preferences are represented in a preference representation language, we can do cool things like preferences aggregation (involving the preferences of multiple agents) and preference revision (a user's new preferences being added to her old preferences). We can also perform the usual computations of decision theory, game theory, |
6532309f-c4ed-4657-be0c-16529d06e380 | trentmkelly/LessWrong-43k | LessWrong | Effectively Handling Disagreements - Introducing a New Workshop
On May 25th, 2023, someone posted a review of How Minds Change on LessWrong. It talked about Street Epistemology, Deep Canvassing, and Smart Politics, ways of handling disagreements that open the possibility of rational belief progression through amicable discussions. Summarized quickly, they rely on active listening, sharing personal stories and socratic questioning.
You can now learn all of those three techniques online, for free, in 4 hours, and in a Deliberate Practice setting. If interested, you can also learn them in an in-person workshop spanning anytime between 2 hours and a full weekend -just shoot me an email with the object EHD (at the time of writing, I’m based in Paris, France).
You can enroll on the website (see bottom for subscribing to the mailing list), and join the discord server.
About the workshop:
What would you learn?
When you find yourself in disagreement with someone on a significant issue, and they might not share your perspectives or even show resistance towards them, it's natural to seek a productive dialogue. The goal is to have a conversation that brings both parties closer to understanding the truth. However, jumping directly into counter-arguments often proves counterproductive, leading to further resistance or increasingly complex counterpoints. It's easy to label the other person as "irrational" in these moments.
To navigate these conversations more effectively, I'm offering a workshop that introduces a range of techniques based on evidence and mutual agreement. These methods are designed to facilitate discussions about deeply held beliefs in a friendly manner, keeping the focus on the pursuit of truth.
Techniques are the following:
4h version:
* Deep Canvassing
* Street Epistemology
* Narrative Transportation
* Cooling Conversations (Smart Politics)
12h version:
All the aforementioned plus Principled Negotiation and bits of Motivational Interviewing
Who is this for?
I’m mainly targeting people who are not used to s |
950bb8ad-fa71-4ad3-8824-a1ce83576c7c | trentmkelly/LessWrong-43k | LessWrong | Less Wrong Q&A with Eliezer Yudkowsky: Ask Your Questions
As promised, here is the "Q" part of the Less Wrong Video Q&A with Eliezer Yudkowsky.
The Rules
1) One question per comment (to allow voting to carry more information about people's preferences).
2) Try to be as clear and concise as possible. If your question can't be condensed to a few paragraphs, you should probably ask in a separate post. Make sure you have an actual question somewhere in there (you can bold it to make it easier to scan).
3) Eliezer hasn't been subpoenaed. He will simply ignore the questions he doesn't want to answer, even if they somehow received 3^^^3 votes.
4) If you reference certain things that are online in your question, provide a link.
5) This thread will be open to questions and votes for at least 7 days. After that, it is up to Eliezer to decide when the best time to film his answers will be. [Update: Today, November 18, marks the 7th day since this thread was posted. If you haven't already done so, now would be a good time to review the questions and vote for your favorites.]
Suggestions
Don't limit yourself to things that have been mentioned on OB/LW. I expect that this will be the majority of questions, but you shouldn't feel limited to these topics. I've always found that a wide variety of topics makes a Q&A more interesting. If you're uncertain, ask anyway and let the voting sort out the wheat from the chaff.
It's okay to attempt humor (but good luck, it's a tough crowd).
If a discussion breaks out about a question (f.ex. to ask for clarifications) and the original poster decides to modify the question, the top level comment should be updated with the modified question (make it easy to find your question, don't have the latest version buried in a long thread).
Update: Eliezer's video answers to 30 questions from this thread can be found here. |
f317e59e-5505-4128-9602-ac271bcecad9 | trentmkelly/LessWrong-43k | LessWrong | Book review: the Iliad
Translated by Emily Wilson
1.
I didn't know what the Iliad was about. I thought it was the story of how Helen of Troy gets kidnapped, triggering the Trojan war, which lasts a long time and eventually gets settled with a wooden horse.
Instead it's just a few days, nine years into that war. The Greeks are camped on the shores near Troy. Agamemnon, King of the Greeks, refuses to return a kidnapped woman to her father for ransom. (Lots of women get kidnapped.) Apollo smites the Greeks with arrows which are plague, and after a while the other Greeks get annoyed enough to tell Agamemnon off. Achilles is most vocal, so Agamemnon returns that woman but takes one of Achilles' kidnapped women instead.
Achilles gets upset and decides to stop fighting for Agamemnon. He prays to his mother, a goddess, wanting the Greeks to suffer ruin without him. She talks to Zeus, who arranges the whole thing, though Hera (his wife and sister) and Athena aren't happy about it because they really want Troy sacked.
So Zeus sends a dream to Agamemnon, telling him to attack Troy. Agamemnon decides to test the Greek fighters, telling them it's time to give up and sail home. So they start running back to the ships, but Athena tells Odysseus to stop them, which he does mostly by telling them to obey orders.
There's a bunch of fighting between Greeks and Trojans, and bickering among the gods, and occasionally mortals even fight gods. In the middle of it there's a one-day truce, which the Greeks use to build a wall, complete with gates. Poseidon says it's such a good wall that people will forget the wall of Troy, which upsets him because they didn't get the gods' blessing to build it.1 Agamemnon tries to convince Achilles to fight again by offering massive rewards, including the woman he stole earlier, whom he swears he has not had sex with as is the normal way between men and women.
Eventually the Trojans fight past the wall and reach the Greek fleet. At this point Patroclus, Achilles' bff (not |
c1ee3d89-94fd-43b9-ab4b-f8367171421a | trentmkelly/LessWrong-43k | LessWrong | [Linkpost] A survey on over 300 works about interpretability in deep networks
Toward Transparent AI: A Survey on Interpreting the Inner Structures of Deep Neural Networks
Tilman Räuker* (traeuker@gmail.com)
Anson Ho* (anson@epochai.org)
Stephen Casper* (scasper@mit.edu)
Dylan Hadfield-Menell
TL;DR: We wrote a survey paper on interpretability tools for deep networks. It was written for the general AI community but with AI safety as the key focus. We survey over 300 works and offer 15 discussion points for guiding future work. Here is a link to a Twitter thread about the paper.
Lately, there has been a growing interest in interpreting AI systems and a growing consensus that it will be key for building safer AI. There have been rapid recent developments in interpretability work, and the AI safety community will benefit from a better systemization of knowledge for it. There are also several epistemic and paradigmatic issues with much interpretability work today. In response to these challenges, we wrote a survey paper covering over 300 works and featuring 15 somewhat “hot takes” to guide future work.
Specifically, this survey focuses on “inner” interpretability methods that help explain internal parts of a network (i.e. not inputs, outputs, or the network as a whole). We do this because inner methods are popular and have some unique applications – not because we think that they are more valuable than other ones. The survey introduces a taxonomy of inner interpretability tools that organizes them by which part of the network’s computational graph they aim to explain: weights (S2), neurons (S3), subnetworks (S4), and latent representations (S5). Then we provide a discussion (S6) and propose directions for future work (S7).
Finally, here are a select few points that we would like to specifically highlight here.
1. Interpretability does not just mean circuits. In the survey sections of the paper (S2-S5), there are 21 subsections, and only one is about circuits. The circuits paradigm has received disproportionate attention in the AI safety co |
e5e61c0e-bc01-460f-af98-a0d178eb8ab0 | trentmkelly/LessWrong-43k | LessWrong | Preference over preference
Each individual person has a preference. Some preferences are strong, others are weak. For many preferences it's more complicated than that; they aren’t static, and we change our preferences all the time. Some days we don't like certain foods, sometimes we may strongly dislike a certain song then another time we may not care so much. Our preferences can change in scope, as well as intensity.
Sometimes people can have preferences over other people's preferences.
* Example 1: I prefer to be surrounded by people who enjoy exercise, that way I will be motivated to exercise more.
* Example 2: I prefer to be surrounded by people who don't care how they look, that way I look prettier than everyone else.
* Example 3: I prefer when other people like my clothes.
* Example 4: I prefer my partners to be polyamorous.
* Example 5: I prefer people around me to not smoke.
The interesting thing about example 3; is that there are multiple ways to achieve that preference:
1. Find out what clothes people like and acquire those clothes, then wear them regularly.
2. Find people who already like the clothes that you have, then hang around those people regularly.
3. Change the preference of the people around you so that they like your clothes.
Changing someone’s preference over clothing seems pretty harmless, and that way you get to wear clothes you like, they get to like the clothes you wear, and you get to be around people who like the clothes you wear without finding new people. The scary and maybe uncomfortable thing is that the other preferences can be also achieved through these means.
Example 4:
1. Find out where poly people are, and hang out with them. (and ask to be their partners - etc)
2. Find out which of the people you know are already poly and hang out with them (and ask to be their partners - etc)
3. Change the preferences of your existing partner/s.
Example 1:
1. Find out where people who enjoy exercise hang out, and join them.
2. Find out whic |
583416ff-9a98-4db8-9eb7-7bd99f8a3822 | trentmkelly/LessWrong-43k | LessWrong | Reflections on Deception & Generality in Scalable Oversight (Another OpenAI Alignment Review)
Just like you can test your skill in experimental design by reviewing existing experiments, you can test your skill in alignment by reviewing existing alignment strategies. Conveniently, Rob Bensinger, in name of Nate Soares and Eliezer Yudkowsky, recently posted a challenge to AI Safety researchers to review the OpenAI alignment plan written by Jan Leike, John Schulman, and Jeffrey Wu. I figured this constituted a test that might net me feedback from both sides of the rationalist-empiricist[1] aisle. Yet, instead of finding ground-breaking arguments for or against scalable oversight to do alignment research, it seems Leike already knows what might go wrong — and goes ahead anyway.
Thus my mind became split between evaluating the actual alignment plan and modeling the disagreement between prominent clusters of researchers. I wrote up the latter in an informal typology of AI Safety Researchers, and continued my technical review below. The following is a short summary of the OpenAI alignment plan, my views on the main problems, and a final section on recommendations for red lining.
The Plan
First, align AI with human feedback, then get AI to assist in giving human feedback to AI, then get AI to assist in giving human feedback to AI that is generating solutions to the alignment problem. Except, the steps are not sequential but run in parallel. This is one form of Scalable Oversight.
Human feedback is Reinforcement Learning from Human Feedback[2] (RLHF), the assisting AI is Iterated Distillation and Amplification (IDA) and Recursive Reward Modeling (RRM), and the AI that is generating solutions to the alignment problem is… still under construction.
The target is a narrow AI that will make significant progress on the alignment problem. The MVP is a theorem prover. The full product is AGI utopia.
Here is a graph.
“Schematic illustration of recursive reward modeling: agents trained with recursive reward modeling (smaller circles on the right) assist the user in the |
171e87b4-58ea-45d3-a1da-ff23eff37e3f | StampyAI/alignment-research-dataset/arbital | Arbital | Ontology identification problem
[https://arbital.com/p/toc:](https://arbital.com/p/toc:)
# Introduction: The ontology identification problem for unreflective diamond maximizers
A simplified but still very difficult open problem in [https://arbital.com/p/2v](https://arbital.com/p/2v) is to state an unbounded program implementing a [diamond maximizer](https://arbital.com/p/5g) that will turn as much of the physical universe into diamond as possible. The goal of "making diamonds" was chosen to have a crisp-seeming definition for our universe (the amount of diamond is the number of carbon atoms covalently bound to four other carbon atoms). If we can crisply define exactly what a 'diamond' is, we can avert issues of trying to convey [complex values](https://arbital.com/p/5l) into the agent. (The [unreflective diamond maximizer](https://arbital.com/p/5g) putatively has [unlimited computing power](https://arbital.com/p/), runs on a [Cartesian processor](https://arbital.com/p/), and confronts no other agents [similar to itself](https://arbital.com/p/). This averts many other problems of [reflectivity](https://arbital.com/p/71), [decision theory](https://arbital.com/p/18s) and [value alignment](https://arbital.com/p/2v).)
Even with a seemingly crisp goal of "make diamonds", we might still run into two problems if we tried to write a [hand-coded object-level utility function](https://arbital.com/p/5t) that [identified](https://arbital.com/p/) the amount of diamond material:
- Unknown substrate: We might not know the true, fundamental ontology of our own universe, hence not know what stuff diamonds are really made of. (What exactly is a carbon atom? If you say it's a nucleus with six protons, what's a proton? If you define a proton as being made of quarks, what if there are unknown other particles underlying quarks?)
- It seems intuitively like there ought to be some way to identify carbon atoms to an AI in some way that doesn't depend on talking about quarks. Doing this is part of the ontology identification problem.
- Unknown representation: We might crisply know what diamonds are in our universe, but not know how to find diamonds inside the agent's model of the environment.
- Again, it seems intuitively like it ought to be possible to identify diamonds in the environment, even if we don't know details of the agent's exact internal representation. Doing this is part of the ontology identification problem.
To introduce the general issues in ontology identification, we'll try to walk through the [anticipated difficulties](https://arbital.com/p/) of constructing an unbounded agent that would maximize diamonds, by trying specific methods and suggesting [anticipated difficulties](https://arbital.com/p/) of those methods.
## Difficulty of making AIXI-tl maximize diamond
The classic unbounded agent - an agent using far more computing power than the size of its environment - is [https://arbital.com/p/11v](https://arbital.com/p/11v). Roughly speaking, AIXI considers all computable hypotheses for how its environment might be turning AIXI's motor outputs into AIXI's sensory inputs and rewards. We can think of AIXI's hypothesis space as including all Turing machines that, sequentially given AIXI's motor choices as inputs, will output a sequence of predicted sense items and rewards for AIXI. The finite variant AIXI-tl has a hypothesis space that includes all Turing machines that can be specified using fewer than $l$ bits and run in less than time $t$.
One way of seeing the difficulty of ontology identification is considering why it would be difficult to make an AIXI-tl variant that maximized 'diamonds' instead of 'reward inputs'.
The central difficulty here is that there's no way to find 'diamonds' inside the implicit representations of AIXI-tl's sequence-predicting Turing machines. Given an arbitrary Turing machine that is successfully predicting AIXI-tl's sense inputs, there is no general rule for how to go from the representation of that Turing machine to a statement about diamonds or carbon atoms. The highest-weighted Turing machines that have best predicted the sensory data so far, presumably contain *some* sort of representation of the environment, but we have no idea how to get 'the number of diamonds' out of it.
If AIXI has a webcam, then the final outputs of the Turing machine are predictions about the stream of bits produced by the webcam, going down the wire into AIXI. We can understand the meaning of that Turing machine's output predictions; those outputs are meant to match types with the webcam's input. But we have no notion of anything else that Turing machine is representing. Even if somewhere in the Turing machine happens to be an atomically detailed model of the world, we don't know what representation it uses, or what format it has, or how to look inside it for the number of diamonds that will exist after AIXI's next motor action.
This difficulty ultimately arises from AIXI being constructed around a [Cartesian](https://arbital.com/p/) paradigm of [sequence prediction](https://arbital.com/p/), with AIXI's sense inputs and motor outputs being treated as sequence elements, and the Turing machines in its hypothesis space having inputs and outputs matched to the sequence elements and otherwise being treated as black boxes. This means we can only get AIXI to maximize direct functions of its sensory input, not any facts about the outside environment.
(We can't make AIXI maximize diamonds by making it want *pictures* of diamonds because then it will just, e.g., [build an environmental subagent that seizes control of AIXI's webcam and shows it pictures of diamonds](https://arbital.com/p/). If you ask AIXI to show itself sensory pictures of diamonds, you can get it to show its webcam lots of pictures of diamonds, but this is not the same thing as building an environmental diamond maximizer.)
## Agent using classical atomic hypotheses
As an [unrealistic example](https://arbital.com/p/): Suppose someone was trying to define 'diamonds' to the AI's utility function, and suppose they knew about atomic physics but not nuclear physics. Suppose they build an AI which, during its development phase, learns about atomic physics from the programmers, and thus builds a world-model that is based on atomic physics.
Again for purposes of [unrealistic examples](https://arbital.com/p/), suppose that the AI's world-model is encoded in such fashion that when the AI imagines a molecular structure - represents a mental image of some molecules - carbon atoms are represented as a particular kind of basic element of the representation. Again, as an [unrealistic example](https://arbital.com/p/), imagine that there are [little LISP tokens](https://arbital.com/p/) representing environmental objects, and that the environmental-object-type of carbon-objects is encoded by the integer 6. Imagine also that each atom, inside this representation, is followed by a list of the other atoms to which it's covalently bound. Then when the AI is imagining a carbon atom participating in a diamond, inside the representation we would see an object of type 6, followed by a list containing exactly four other 6-objects.
Can we fix this representation for all hypotheses, and then write a utility function for the AI that counts the number of type-6 objects that are bound to exactly four other type-6 objects? And if we did so, would the result actually be a diamond maximizer?
### AIXI-atomic
We can imagine formulating a variant of AIXI-tl that, rather than all tl-bounded Turing machines, considers tl-bounded simulated atomic universes - that is, simulations of classical, pre-nuclear physics. Call this AIXI-atomic.
A first difficulty is that universes composed only of classical atoms are not good explanations of our own universe, even in terms of surface phenomena; e.g., the [ultraviolet catastrophe](http://en.wikipedia.org/wiki/Ultraviolet_catastrophe). So let it be supposed that we have simulation rules for classical physics that replicate at least whatever phenomena the programmers have observed at [development time](https://arbital.com/p/), even if the rules have some seemingly ad-hoc elements (like there being no ultraviolent catastrophes).
A second difficulty is that a simulated universe of classical atoms does not identify where in the universe the AIXI-atomic agent resides, or that AIXI-atomic's sense inputs don't have types commensurate with the types of atoms. We can elide this difficulty by imagining that AIXI-atomic simulates classical universes containing a single hypercomputer, and that AIXI-atomic knows a simple function from each simulated universe onto its own sensory data (e.g., it knows to look at the simulated universe, and translate simulated photons impinging on its webcam onto predicted webcam data in the received format). This elides most of the problem of [naturalized induction](https://arbital.com/p/), by fixing the ontology of all hypotheses and standardizing their hypothetical [bridging laws](https://arbital.com/p/).
So the analogous AIXI-atomic agent that maximizes diamond:
- Considers only hypotheses that directly represent universes as huge systems of classical atoms, so that the function 'count atoms bound to four other carbon atoms' can be directly run over any possible future the agent considers.
- Assigns probabilistic priors over these possible atomic representations of the universe.
- Somehow [maps each atomic representation onto the agent's sensory experiences and motor actions](https://arbital.com/p/).
- [its priors](https://arbital.com/p/Bayes-updates) based on actual sensory experiences, the same as classical AIXI.
- Can evaluate the 'expected diamondness on the next turn' of a single action by looking at all hypothetical universes where that action is performed, weighted by their current probability, and summing over the expectation of diamond-bound carbon atoms on their next clock tick.
- Can evaluate the 'future expected diamondness' of an action, over some finite time horizon, by assuming that its future self will also Bayes-update and maximize expected diamondness over that time horizon.
- On each turn, outputs the action with greatest expected diamondness over some finite time horizon.
Suppose our own real universe was amended to otherwise be exactly the same, but contain a single [impermeable](https://arbital.com/p/) hypercomputer. Suppose we defined an agent like the one above, using simulations of 1900-era classical models of physics, and ran that agent on the hypercomputer. Should we expect the result to be an actual diamond maximizer - that most mass in the universe will be turned into carbon and arranged into diamonds?
### Anticipated failure of AIXI-atomic in our own universe: trying to maximize diamond outside the simulation
Our own universe isn't atomic, it's nuclear and quantum-mechanical. This means that AIXI-atomic does not contain any hypotheses in its hypothesis space that *directly represent* the universe. By 'directly represent', we mean that carbon atoms in AIXI-atomic's best representations do not correspond to carbon atoms in our own world).
Intuitively, we would think it was [common sense](https://arbital.com/p/) for an agent that wanted diamonds to react to the experimental data identifying nuclear physics, by deciding that a carbon atom is 'really' a nucleus containing six protons, and atomic binding is 'really' covalent electron-sharing. We can imagine this agent [common-sensically](https://arbital.com/p/) updating its model of the universe to a nuclear model, and redefining the 'carbon atoms' that its old utility function counted to mean 'nuclei containing exactly six protons'. Then the new utility function could evaluate outcomes in the newly discovered nuclear-physics universe. We will call this the **utility rebinding problem**.
We don't yet have a crisp formula that seems like it would yield commonsense behavior for utility rebinding. In fact we don't yet have any candidate formulas for utility rebinding, period. Stating one is an open problem. See below.
For the 'classical atomic AIXI' agent we defined above, what happens instead is that the 'simplest atomic hypothesis that fits the facts' will be an enormous atom-based computer, simulating nuclear physics and quantum physics in order to control AIXI's webcam, which is still believed to be composed of atoms in accordance with the prespecified bridging laws. From our perspective this hypothesis seems silly, but if you restrict the hypothesis space to only classical atomic universes, that's what ends up being the computationally simplest hypothesis to explain the results of quantum experiments.
AIXI-atomic will then try to choose actions so as to maximize the amount of expected diamond inside the probable *outside universes* that could contain the giant atom-based simulator of quantum physics. It is not obvious what sort of behavior this would imply.
### Metaphor for difficulty: AIXI-atomic cares about only fundamental carbon
One metaphorical way of looking at the problem is that AIXI-atomic was implicitly defined to care only about diamonds made out of *ontologically fundamental* carbon atoms, not diamonds made out of quarks. A probability function that assigns 0 probability to all universes made of quarks, and a utility function that outputs a constant on all universes made of quarks, [yield functionally identical behavior](https://arbital.com/p/). So it is an exact metaphor to say that AIXI-atomic only *cares* about universes with ontologically basic carbon atoms, given that AIXI-atomic only *believes* in universes with ontologically basic carbon atoms.
Since AIXI-atomic only cares about diamond made of fundamental carbon, when AIXI-atomic discovered the experimental data implying that almost all of its probability mass should reside in nuclear or quantum universes in which there were no fundamental carbon atoms, AIXI-atomic stopped caring about the effect its actions had on the vast majority of probability mass inside its model. Instead AIXI-atomic tried to maximize inside the tiny remaining probabilities in which it *was* inside a universe with fundamental carbon atoms that was somehow reproducing its sensory experience of nuclei and quantum fields; for example, a classical atomic universe with an atomic computer simulating a quantum universe and showing the results to AIXI-atomic.
From our perspective, we failed to solve the 'ontology identification problem' and get the real-world result we wanted, because we tried to define the agent's *utility function* in terms of properties of a universe made out of atoms, and the real universe turned out to be made of quantum fields. This caused the utility function to *fail to bind* to the agent's representation in the way we intuitively had in mind.
### Advanced-nonsafety of hardcoded ontology identifications
Today we do know about quantum mechanics, so if we tried to build an unreflective diamond maximizer using the above formula, it might not fail on account of [the particular exact problem](https://arbital.com/p/48) of atomic physics being false.
But perhaps there are discoveries still remaining that would change our picture of the universe's ontology to imply something else underlying quarks or quantum fields. Human beings have only known about quantum fields for less than a century; our model of the ontological basics of our universe has been stable for less than a hundred years of our human experience. So we should seek an AI design that does not assume we know the exact, true, fundamental ontology of our universe during an AI's [development phase](https://arbital.com/p/5d). Or if our failure to know the exact laws of physics causes catastrophic failure of the AI, we should at least heavily mark that this is a [relied-on assumption](https://arbital.com/p/).
## Beyond AIXI-atomic: Diamond identification in multi-level maps
A realistic, bounded diamond maximizer wouldn't represent the outside universe with atomically detailed models. Instead, it would have some equivalent of a [multi-level map](https://arbital.com/p/) of the world in which the agent knew in principle that things were composed of atoms, but didn't model most things in atomic detail. E.g., its model of an airplane would have wings, or wing shapes, rather than atomically detailed wings. It would think about wings when doing aerodynamic engineering, atoms when doing chemistry, nuclear physics when doing nuclear engineering.
At the present, there are not yet any proposed formalisms for how to do probability theory with multi-level maps (in other words: [nobody has yet put forward a guess at how to solve the problem even given infinite computing power](https://arbital.com/p/)). Having some idea for how an agent could reason with multi-level maps, would be a good first step toward being able to define a bounded expected utility optimizer with a utility function that could be evaluated on multi-level maps. This in turn would be a first step towards defining an agent with a utility function that could rebind itself to *changing* representations in an *updating* multi-level map.
If we were actually trying to build a diamond maximizer, we would be likely to encounter this problem long before it started formulating new physics. The equivalent of a computational discovery that changes 'the most efficient way to represent diamonds' is likely to happen much earlier than a physical discovery that changes 'what underlying physical systems probably constitute a diamond'.
This also means that, on the actual [value loading problem](https://arbital.com/p/), we are liable to encounter the ontology identification problem long before the agent starts discovering new physics.
# Discussion of the generalized ontology identification problem
If we don't know how to solve the ontology identification problem for maximizing diamonds, we probably can't solve it for much more complicated values over universe-histories.
### View of human angst as ontology identification problem
Argument: A human being who feels angst on contemplating a universe in which "By convention sweetness, by convention bitterness, by convention color, in reality only atoms and the void" (Democritus), or wonders where there is any room in this cold atomic universe for love, free will, or even the existence of people - since, after all, people are just *mere* collections of atoms - can be seen as undergoing an ontology identification problem: they don't know how to find the objects of value in a representation containing atoms instead of ontologically basic people.
Human beings simultaneously evolved a particular set of standard mental representations (e.g., a representation for colors in terms of a 3-dimensional subjective color space, a representation for other humans that simulates their brain via [https://arbital.com/p/empathy](https://arbital.com/p/empathy)) along with evolving desires that bind to these representations ([identification of flowering landscapes as beautiful](http://en.wikipedia.org/wiki/Evolutionary_aesthetics#Landscape_and_other_visual_arts_preferences), a preference not to be embarrassed in front of other objects designated as people). When someone visualizes any particular configurations of 'mere atoms', their built-in desires don't automatically fire and bind to that mental representation, the way they would bind to the brain's native representation of other people. Generalizing that no set of atoms can be meaningful, and being told that reality is composed entirely of such atoms, they feel they've been told that the true state of reality, underlying appearances, is a meaningless one.
Arguably, this is structurally similar to a utility function so defined as to bind only to true diamonds made of ontologically basic carbon, which evaluates as unimportant any diamond that turns out to be made of mere protons and neutrons.
## Ontology identification problems may reappear on the reflective level
An obvious thought (especially for [online genies](https://arbital.com/p/6w)) is that if the AI is unsure about how to reinterpret its goals in light of a shifting mental representation, it should query the programmers.
Since the definition of a programmer would then itself be baked into the [preference framework](https://arbital.com/p/5f), the problem might [reproduce itself on the reflective level](https://arbital.com/p/) if the AI became unsure of where to find [programmers](https://arbital.com/p/9r). ("My preference framework said that programmers were made of carbon atoms, but all I can find in this universe are quantum fields.")
## Value lading in category boundaries
Taking apart objects of value into smaller components can sometimes create new moral [edge cases](https://arbital.com/p/). In this sense, rebinding the terms of a utility function decides a [value-laden](https://arbital.com/p/) question.
Consider chimpanzees. One way of viewing questions like "Is a chimpanzee truly a person?" - meaning, not, "How do we arbitrarily define the syllables per-son?" but "Should we care a lot about chimpanzees?" - is that they're about how to apply the 'person' category in our desires to things that are neither typical people nor typical nonpeople. We can see this as arising from something like an ontological shift: we're used to valuing cognitive systems that are made from whole human minds, but it turns out that minds are made of parts, and then we have the question of how to value things that are made from some of the person-parts but not all of them.
Redefining the value-laden category 'person' so that it talked about brains made out of neural regions, rather than whole human beings, would implicitly say whether or not a chimpanzee was a person. Chimpanzees definitely have neural areas of various sizes, and particular cognitive abilities - we can suppose the empirical truth is unambiguous at this level, and known to us. So the question is then whether we regard a particular configuration of neural parts (a frontal cortex of a certain size) and particular cognitive abilities (consequentialist means-end reasoning and empathy, but no recursive language) as something that our 'person' category values... once we've rewritten the person category to value configurations of cognitive parts, rather than whole atomic people.
In this sense the problem we face with chimpanzees is exactly analogous to the question a diamond maximizer would face after discovering nuclear physics and asking itself whether a carbon-14 atom counted as 'carbon' for purposes of caring about diamonds. Once a diamond maximizer knows about neutrons, it can see that C-14 is chemically like carbon and forms the same kind of chemical bonds, but that it's heavier because it has two extra neutrons. We can see that chimpanzees have a similar brain architectures to the sort of people we always considered before, but that they have smaller frontal cortexes and no ability to use recursive language, etcetera.
Without knowing more about the diamond maximizer, we can't guess what sort of considerations it might bring to bear in deciding what is Truly Carbon and Really A Diamond. But the breadth of considerations human beings need to invoke in deciding how much to care about chimpanzees, is one way of illustrating that the problem of rebinding a utility function to a shifted ontology is [https://arbital.com/p/value-laden](https://arbital.com/p/value-laden) and potentially undergo [excursions](https://arbital.com/p/) into [arbitrarily complicated desiderata](https://arbital.com/p/). Redefining a [moral category](https://arbital.com/p/) so that it talks about the underlying parts of what were previously seen as all-or-nothing atomic objects, may carry an implicit ruling about how to value many kinds of [edge case](https://arbital.com/p/) objects that were never seen before.
A formal part of this problem may need to be carved out from the edge-case-reclassification part: e.g., how would you redefine carbon as C12 if there were no other isotopes, or how would you rebind the utility function to *at least* C12, or how would edge cases be identified and queried.
# Potential research avenues
## 'Transparent priors' constrained to meaningful but Turing-complete hypothesis spaces
The reason why we can't bind a description of 'diamond' or 'carbon atoms' to the hypothesis space used by [https://arbital.com/p/11v](https://arbital.com/p/11v) or [AIXI-tl](https://arbital.com/p/) is that the hypothesis space of AIXI is all Turing machines that produce binary strings, or probability distributions over the next sense bit given previous sense bits and motor input. These Turing machines could contain an unimaginably wide range of possible contents
(Example: Maybe one Turing machine that is producing good sequence predictions inside AIXI, actually does so by simulating a large universe, identifying a superintelligent civilization that evolves inside that universe, and motivating that civilization to try to intelligently predict future future bits from past bits (as provided by some intervention). To write a formal utility function that could extract the 'amount of real diamond in the environment' from arbitrary predictors in the above case , we'd need the function to read the Turing machine, decode that universe, find the superintelligence, decode the superintelligence's thought processes, find the concept (if any) resembling 'diamond', and hope that the superintelligence had precalculated how much diamond was around in the outer universe being manipulated by AIXI.)
This suggests that to solve the ontology identification problem, we may need to constrain the hypothesis space to something [less general](https://arbital.com/p/) than 'an explanation is any computer program that outputs a probability distribution on sense bits'. A constrained explanation space can still be Turing complete (contain a possible explanation for every computable sense input sequence) without every possible computer program constituting an explanation.
An [unrealistic example](https://arbital.com/p/) would be to constrain the hypothesis space to Dynamic Bayesian Networks. DBNs can represent any Turing machine with bounded memory, so they are very general, but since a DBN is a causal model, they make it possible for a preference framework to talk about 'the cause of a picture of a diamond' in a way that you couldn't look for 'the cause of a picture of a diamond' inside a general Turing machine. Again, this might fail if the DBN has no 'natural' way of representing the environment except as a DBN simulating some other program that simulates the environment.
Suppose a rich causal language, such as, e.g., a [dynamic system](https://arbital.com/p/) of objects with [causal relations](https://arbital.com/p/) and [hierarchical categories of similarity](https://arbital.com/p/). The hope is that in this language, the *natural* hypothesis representing the environment - the simplest hypotheses within this language that well predict the sense data, or those hypotheses of highest probability under some simplicity prior after updating on the sense data - would be such that there was a natural 'diamond' category inside the most probable causal models. In other words, the winning hypothesis for explaining the universe would already have postulated diamondness as a [natural category](https://arbital.com/p/) and represented it as Category #803,844, in a rich language where we already know how to look through the enviromental model and find the list of categories.
Given some transparent prior, there would then exist the further problem of developing a utility-identifying preference framework that could look through the most likely environmental representations and identify diamonds. Some likely (interacting) ways of binding would be, e.g., to "the causes of pictures of diamonds", to "things that are bound to four similar things", querying ambiguities to programmers, or direct programmer inspection of the AI's model (but in this case the programmers might need to re-inspect after each ontological shift). See below.
(A bounded value loading methodology would also need some way of turning the bound preference framework into the estimation procedures for expected diamond and the agent's search procedures for strategies high in expected diamond, i.e., the bulk of the actual AI that carries out the goal optimization.)
## Matching environmental categories to descriptive constraints
Given some transparent prior, there would exist a further problem of how to actually bind a preference framework to that prior. One possible contributing method for pinpointing an environmental property could be if we understand the prior well enough to understand what the described object ought to look like - the equivalent of being able to search for 'things W made of six smaller things X near six smaller things Y and six smaller things Z, that are bound by shared Xs to four similar things W in a tetrahedral structure' in order to identify carbon atoms and diamond.
We would need to understand the representation well enough to make a guess about how carbon or diamond would be represented inside it. But if we could guess that, we could write a program that identifies 'diamond' inside the hypothesis space without needing to know in advance that diamondness will be Category #823,034. Then we could rerun the same utility-identification program when the representation updates, so long as this program can reliably identify diamond inside the model each time, and the agent acts so as to optimize the utility identified by the program.
One particular class of objects that might plausibly be identifiable in this way is 'the AI's programmers' (aka the agents that are causes of the AI's code) if there are parts of the preference framework that say to query programmers to resolve ambiguities.
A toy problem for this research avenue might involve:
- One of the richer representation frameworks that can be inducted as of the time, e.g., a simple Dynamic Bayes Net.
- An agent environment that can be thus represented.
- A goal over properties relatively distant from the agent's sensory experience (e.g., the goal is over the cause of the cause of the sensory data).
- A program that identifies the objects of utility in the environment, within the model thus freely inducted.
- An agent that optimizes the identified objects of utility, once it has inducted a sufficiently good model of the environment to optimize what it is looking for.
Further work might add:
- New information that can change the model of the environment.
- An agent that smoothly updates what it optimizes for in this case.
And further:
- Environments complicated enough that there is real structural ambiguity (e.g., dependence on exact initial conditions of the inference program) about how exactly the utility-related parts are modeled.
- Agents that can optimize through a probability distribution about environments that differ in their identified objects of utility.
A potential agenda for unbounded analysis might be:
- An [unbounded analysis](https://arbital.com/p/) showing that a utility-identifying [preference framework](https://arbital.com/p/5f) is a generalization of a [VNM utility](https://arbital.com/p/) and can [tile](https://arbital.com/p/) in an architecture that tiles a generic utility function.
- A [https://arbital.com/p/45](https://arbital.com/p/45) analysis showing that an agent is not motivated to try to cause the universe to be such as to have utility identified in a particular way.
- A [https://arbital.com/p/45](https://arbital.com/p/45) analysis showing that the identity and category boundaries of the objects of utility will be treated as a [historical fact](https://arbital.com/p/) rather than one lying in the agent's [decision-theoretic future](https://arbital.com/p/).
## Identifying environmental categories as the causes of labeled sense data.
Another potential approach, given a prior transparent enough that we can find causal data inside it, would be to try to identify diamonds as the causes of pictures of diamonds.
### Security note
[Christiano's hack](https://arbital.com/p/5j): if your AI is advanced enough to model distant superintelligences, it's important to note that distant superintelligences can make 'the most probable cause of the AI's sensory data' be anything they want by making a predictable decision to simulate AIs such that your AI doesn't have info distinguishing itself from the distant AIs your AI imagines being simulated
## Ambiguity resolution
Both the description-matching and cause-inferring methods might produce ambiguities. Rather than having the AI optimize for a probabilistic mix over all the matches (as if it were uncertain of which match were the true one), it would be better to query the ambiguity to the programmers (especially if different probable models imply different strategies). This problem shares structure with [inductive inference with ambiguity resolution](https://arbital.com/p/) as a strategy for resolving [unforeseen inductions](https://arbital.com/p/).
## Multi-level maps
Being able to describe, in purely theoretical principle, a prior over epistemic models that have at least two levels and can switch between them in some meaningful sense, would constitute major progress over the present state of the art.
# Implications
The problem of using sensory data to build computationally efficient probabilistic maps of the world, and to efficiently search for actions that are predicted by those maps to have particular consequences, could be identified with the entire problem of AGI. So the research goal of ontology identification is not to publish a complete bounded system like that (i.e. an AGI), but to develop an unbounded analysis of utility rebinding that seems to say something useful specifically about the ontology-identification part of the problem.) |
9772ba53-c767-442c-a527-22c2110f20d8 | trentmkelly/LessWrong-43k | LessWrong | Vulnerabilities in CDT and TI-unaware agents
The aim of this post is illustrating the need to take into account decision-making and incentive considerations when designing agents. This post is also a proof that these considerations are important in order to ensure the safety of agents. Also, we will postulate that there exist some agents that are both robust to changing or having their reward function changed, although that will need a careful approach to incentive design and decision theory choosing.
The first agent we will consider is a (current Reward Function, Time Inconsistent aware, see in the second half of the post if you don't know what this means) agent that uses Causal Decision Theory (CDT). A review of different decision theories can be seen in this post. It is well known that Updateless Decision Theory (UDT) was created to correct the wrong decision a CDT agent would make when faced with Newcomb-like problems. Thus, the question we aim to answer is whether we can exploit the wrong decision making procedure in order to induce any changes in the value function of such an agent.
This is not exactly trivial since the agent could value very negatively to have its value function changed and thus opt out of games such as Newcomb. The example I propose is a modified version of Prisoner's dilemma (in which CDT is known to defect). Suppose the following problem:
> It is year 2100 and Elon Musk managed to effectively colonise Mars. Meanwhile, an AI corporation, called Causal Corp, has deployed many CDT agents both in Earth and Mars. One day, Eve, known for being evil, codes a virus that if connected to one such CDT agent would arbitrarily modify the reward function. Eve makes two copies of such virus in memory sticks and sends them to arrive almost "simultaneously" to two CDT agents in Earth and Mars. With the memory stick there is a letter that tells the agents that they face a Prisoner Dilemma situation:
> 1. If both cooperate nothing will happen.
> 2. If one defects and the other cooperates, the firs |
a4a25f70-dc0b-4dc1-a0c4-cd35966fc6ef | trentmkelly/LessWrong-43k | LessWrong | Contra Dance Dialect Survey
Fifteen years ago I travelled to a bunch of dances and wrote something up ( pdf) on how people tended to dance in different areas. I'm curious how this has changed over the years, and while I have my observations I'd like to get some better data. So! I'm running a survey, and would find your response very helpful: survey.
Comment via: facebook, mastodon |
87ac8b26-9ba4-4fb8-8d00-98882be03ffb | trentmkelly/LessWrong-43k | LessWrong | Meetup : Baltimore / UMBC Meetup - trying something new!
Discussion article for the meetup : Baltimore / UMBC Meetup - trying something new!
WHEN: 05 February 2017 11:00:00AM (-0500)
WHERE: Performing Arts and Humanities Bldg 4th floor, 1000 Hilltop Cir, Baltimore, MD 21250
We're going to try something for the next few weeks and see how it goes. Two main changes:
* Meetup will be Sundays at 11:00 AM. Partly that's just because I'd like to start coming again, and that's when I can (usually) make it.
* We'll actually have something on the agenda.
Meetup will be in the same place - upstairs philosophy department at UMBC. We can probably get access to the philosophy conference room again, I'll work on that.
For the agenda I'd like to loosely follow the order of EY's Rationality book, but only loosely and without sticking to his views in particular.
This week we'll be discussing the subject of Truth - what is it, what are we aiming for exactly, why is it a useful concept to think about, etc.
I'm going to put up some reading material for each week. The more people who glance over at least some of the material, the more intelligent the conversation will be.
Here's the material for this week, ordered roughly by ease of reading:
* Eliezer Yudkowsky (EY): Why Truth? And ...
* EY: The Simple Truth
* EY: The Useful Idea of Truth
* Why I Reject the Correspondence Theory of Truth
* Internet Encyclopedia of Philosophy (IEP): Truth (or you could try the SEP or Wikipedia)
Discussion article for the meetup : Baltimore / UMBC Meetup - trying something new! |
0db0fad9-69cf-48cf-86c2-85ea538a0227 | trentmkelly/LessWrong-43k | LessWrong | Draft: Inferring minimizers
This post is a draft, a work in progress. Much of it will not make sense, and will not be optimized for the reader's experience. It does not necessarily reflect my current best understanding of optimization. Like the introduction draft, I'm opening it up because I've been working on this project for too long, and the world is moving too fast. I want people to more easily be able to interact with my research thoughts as I'm having them.
----------------------------------------
We left off with an argument by analogy of how we could use Kolmogorov complexity as an "objective" means to infer the presence of optimization in a system. But how exactly, does that analogy carry over? How exactly do we use K-complexity in the context of dynamical systems trajectories?
In this post, I'll treat what I would call the "easy" case. The cases are formed by the structure of the state orders, so let's dig deeper into that.
Order shapes
In the introduction post, I talked about how the optimization target or criterion could be represented by a mathematical object called an order. Orders get pretty interesting once you start to include orders over infinite sets. In some sense, you could say that finite orders differ only in how they order things, but that infinite orders can differ in how they are "shaped". Consider this example
* The typical ordering over the integers
* The ordering over the integers provided by absolute value
The typical ordering has no least or greatest element; for every integer, you can find one that is higher in the ordering, and one that is lower. But in the absolute value order, there is a specific least element; 0. (And, of course, no greatest element.)
You can get pretty crazy with this. Here are some more examples.
[TODO just definitely pictures for all of these]
* The natural numbers with the typical ordering, except you declare that 0 is the greatest element.
* The integers ordered by 1x
* Collate the natural numbers by their lowest divisor. |
45202d35-69f3-45d1-9504-c5433f9f25b4 | trentmkelly/LessWrong-43k | LessWrong | Gödel and Bayes: quick question
Kurt Gödel showed that we could write within a system of arithmetic the statement, "This statement has no proof within the system," in such a way that we couldn't dismiss it as meaningless. This proved that if the system (or part of it) could prove the logical consistency of the whole, it would thereby contradict itself. We nevertheless think arithmetic does not contradict itself because it never has.
From what I understand we could write a version of the Gödel statement for the axioms of probability theory, or even for the system that consists of those axioms plus our current best guess at P(axioms' self-consistency). Edit: or not. According to the comments the Incompleteness Theorem does not apply until you have a stronger set of assumptions than the minimum you need for probability theory. So let's say you possess the current source code of an AGI running on known hardware. It's just now reached the point where it could pass the test of commenting extensively on LW without detection. (Though I guess we shouldn't yet assume this will continue once the AI encounters certain questions.) For some reason it tries to truthfully answer any meaningful question. (So nobody mention the Riemann hypothesis. We may want the AI to stay in this form for a while.) Whenever an internal process ends in a Jaynes-style error message that indicates a contradiction, the AI takes this as strong evidence of a contradiction in the relevant assumptions. Now according to my understanding we can take the source code and ask about a statement which says, "The program will never call this statement true." Happily the AI can respond by calling the statement "likely enough given the evidence." So far so good.
So, can we or can't we write a mathematically meaningful statement Q saying, "The program will never say 'P(Q)≥0.5'"? What about, "The program will never call 'P(Q)≥0.5' true (or logically imply it)"? How does the AI respond to questions about variations of these statements?
It seems as |
2ee4ebfa-7346-447d-9f3d-c04442d8a66b | StampyAI/alignment-research-dataset/arxiv | Arxiv | Exploring Fluent Query Reformulations with Text-to-Text Transformers and Reinforcement Learning
1 Introduction
---------------
Query reformulation and paraphrase generation techniques are employed for a variety of purposes in natural language processing (NLP), such as dialogue generation [[27](#bib.bib55 "Question rewrite based dialogue response generation")], machine translation [[29](#bib.bib57 "The circle of meaning: from translation to paraphrasing and back")], and especially in question answering (QA) systems [[16](#bib.bib56 "Learning to rank effective paraphrases from query logs for community question answering"), [50](#bib.bib19 "Paraphrasing with large language models"), [11](#bib.bib54 "Can you unpack that? learning to rewrite questions-in-context")]. Generating coherent and clean texts reduces potential errors in downstream systems. In the case when users are at the receiving end of NLP pipelines, it is essential to show fluent and human-like languages before trust is lost until a point where users recede into requiring human agents for the sake of better communication. Lastly, by having a reformulator model to modify queries, results can be fed back to users and confirm their original intentions in an automated way.
The advent of Seq2Seq learning [[44](#bib.bib50 "Sequence to sequence learning with neural networks")] made it possible to train deep neural networks as a new paradigm to replace rule-based and statistical approaches to generate reformulations and paraphrases [[31](#bib.bib59 "Strategies for effective paraphrasing"), [30](#bib.bib58 "Paraphrasing using given and new information in a question-answer system"), [55](#bib.bib60 "Application-driven statistical paraphrase generation")]. We investigate how to generate well-formed reformulations using Seq2Seq models that can maintain good QA performance at the same time. We apply the pre-training and RL pipeline from previous work AQA [[4](#bib.bib9 "Ask the right questions: active question reformulation with reinforcement learning")] to the T5 framework [[36](#bib.bib20 "Exploring the limits of transfer learning with a unified text-to-text transformer")] and fine-tune the QRT5 reformulator on paraphrasing and denoising, before RL is applied in downstream QA and intent classification (IC) environments that provide reward signals. We choose T5 because it is a state-of-the-art Seq2Seq model that suits the query reformulation task. It provides flexibility for tuning on different datasets without changing the training pipeline significantly. Furthermore, the self-supervised pre-training [[12](#bib.bib2 "Why does unsupervised pre-training help deep learning?")] and sequential transfer learning approach [[40](#bib.bib67 "Neural transfer learning for natural language processing")] in T5 has consistently provided good inductive bias in the past, obtaining improvements across many NLP benchmarks [[9](#bib.bib26 "BERT: pre-training of deep bidirectional transformers for language understanding"), [28](#bib.bib25 "RoBERTa: a robustly optimized bert pretraining approach"), [36](#bib.bib20 "Exploring the limits of transfer learning with a unified text-to-text transformer")] and showing great out-of-distribution generalization [[20](#bib.bib21 "Pretrained transformers improve out-of-distribution robustness"), [2](#bib.bib24 "Language models are few-shot learners")].
To our knowledge, this is the first attempt at training T5 with RL to produce query reformulations. We show that QRT5 is a better starting point for RL and more sample efficient to achieve the same level of QA performance as AQA. The efficiency of RL is important in a productionized pipeline where rewards are defined by humans. These systems in practice are typically expensive to interact with when they are black boxes that can be frequently updated or changed. This is even more relevant to query reformulation as episodic rewards from QA or IC are produced only after a complete sequence is constructed, without intermediate signals in between tokens to updates the model. In addition, QRT5 reformulates with better readability and can generalize to out-of-distribution (OOD) data. This is crucial because queries have word token permutations and varying levels of ambiguity and syntactic complexity [[32](#bib.bib15 "Linguistic features to predict query difficulty")] with distinct properties of their own [[1](#bib.bib17 "The linguistic structure of english web-search queries"), [39](#bib.bib16 "Analyzing linguistic structure of web search queries")]. Thus, semantic meanings may be lost in reformulations and need to be preserved, especially for OOD queries while improving fluency. Lastly, we evaluate reformulation fluency using scores produced by a separate T5 model trained on the question well-formedness (QW) [[15](#bib.bib46 "Identifying well-formed natural language questions")] dataset, which is based on real evaluations from humans. Widely-used algorithmic heuristics based on overlapping n-grams like BLEU [[34](#bib.bib30 "Bleu: a method for automatic evaluation of machine translation")] or ROGUE [[25](#bib.bib31 "ROUGE: a package for automatic evaluation of summaries")] are found to be less correlated with human judgements [[5](#bib.bib32 "Re-evaluation the role of bleu in machine translation research"), [43](#bib.bib33 "Learning to summarize from human feedback")]. Query well-formedness training provides a proxy to sequence-level fluency that mimics human preferences.
2 Preliminary
--------------
###
2.1 Supervised Fine-tuning
Given a query sequence q={q1,⋯,qk} of length k in a dataset of size N, the reformulator model produces a sequence of word token distributions. A reformulation r={r1,⋯,rT} is produced by sampling tokens from these distributions at each time step. They are matched with the target sequence y={y1,⋯,yT} in a final loss. We assume here that y and r have the same length. Practically, sequences are padded to a default max length of 50, and any token distributions produced after the end special token are disregarded in the loss. Both of input q and target y are tokenized by a pre-defined sentence-piece [[23](#bib.bib65 "SentencePiece: a simple and language independent subword tokenizer and detokenizer for neural text processing")] with vocabulary size V. We use the default vocabulary of size 16,000 for AQA and 32,168 for QRT5. For the i’th data point, the conditional sequence probability of the reformulation ri is given by
πθ(ri|qi)=∏Tt=1p(rit|ri1,⋯,rit−1,qi1,⋯,qik). The correct label yi is defined by a sequence of one-hot encodings at each time step. Therefore, the cross entropy loss between the model predictions ri and target yi is:
| | | |
| --- | --- | --- |
| | LCE=−N∑i=1yilogπθ(ri|qi) | |
| | | |
| --- | --- | --- |
| | =−N∑i=1T∑t=1V∑j=1yij,tlogp(rij,t|ri1,⋯,rit−1,qi) | |
where yij,t∈{0,1} is the binary label of token j at time t, and p(rij,t|⋅) is the conditional probability of token j appearing at t given previously produced tokens and input query. Note that minimizing cross entropy loss LCE is equivalent to minimizing the negative log-likelihood LNLL=−∑Ni=1logπθ(ri|qi)=−∑Ni=1∑Tt=1logp(rit|ri1,⋯,rit−1,qi)
###
2.2 Reinforcement Learning
In the RL stage, we follow [[4](#bib.bib9 "Ask the right questions: active question reformulation with reinforcement learning")] and optimize for the expected long term rewards:
| | | |
| --- | --- | --- |
| | J=N∑i=1Eri∼πθ(T∑t=1R(ri1,⋯,rit)) | |
where R is the black-box function that generates rewards between 0 and 1 only at the end of generation t=T. In the QA setting, this reward is the character-level F1 score RF1=2(p⋅r)/(p+r) between the true answer a and the answer output a′ from the pre-trained BiDAF QA model [[41](#bib.bib12 "Bidirectional attention flow for machine comprehension")], when the reformulation is given as input. Precision p is the proportion of tokens in a′ that are in a, while recall r is the proportion of tokens in a that are in a′. The main quantitative metric of interest is the F1 score on the dev set as this shows how well models can generalize on unseen SearchQA [[10](#bib.bib53 "SearchQA: a new q&a dataset augmented with context from a search engine")] data during RL.
From a batch of size b, the gradient of J can be estimated by REINFORCE [[49](#bib.bib8 "Function optimization using connectionist reinforcement learning algorithms")], by sampling a reformulation trajectory ri={ri1,⋯,riT}∼πθ given qi:
| | | |
| --- | --- | --- |
| | ∇J≈b∑i=1∇θlogp(ri1,⋯,riT|qi)(R(ri)−Bi) | |
| | | |
| --- | --- | --- |
| | =b∑i=1∇θrilogπθ(ri|qi)(R(ri)−Bi) | |
This means that we can maximize the above weighted log likelihood function (i.e. minimizing a weighted cross-entropy loss LCE) as a surrogate to compute and estimate the policy gradient [[45](#bib.bib7 "Policy gradient methods for reinforcement learning with function approximation")] in each batch. The target sequences in this surrogate cross entropy loss are the sampled reformulation trajectories ri from the policy for estimating the expectation. A trajectory that obtains a higher reward will produce a higher gradient signal to encourage generation of words close to this sampled reformulation at each time step. Similar to AQA, for variance reduction, the mean reward of the minibatch is used as the baseline B in the gradient and the loss. In addition, a scaled entropy regularizer λH(πθ)=λ∑t∑jp(rj,t|r<t,qi)logp(rj,t|r<t,qi) is added to the loss to mitigate deterministic policy updates. The modifications in following sections originate from this objective function, which we refer to as the policy gradient (PG) baseline.
3 Methodology
--------------
###
3.1 Datasets
SearchQA [[10](#bib.bib53 "SearchQA: a new q&a dataset augmented with context from a search engine")] is comprised of 140,000 question-answer pairs from the Jeopardy! archive. Questions are inputs to the reformulators. In RL, the QA environment model has been pre-trained with this dataset on machine comprehension to extract answers within the context given a question. The queries and context snippets are mostly convoluted or ambiguous. Paralex and UN Parallel Corpus [[13](#bib.bib47 "Paraphrase-driven learning for open question answering"), [56](#bib.bib69 "The united nations parallel corpus v1.0")] are paraphrasing and multilingual translation datasets used by [[4](#bib.bib9 "Ask the right questions: active question reformulation with reinforcement learning")] for supervised pre-training. Question well-formedness (QW) dataset [[15](#bib.bib46 "Identifying well-formed natural language questions")] are filtered from Paralex, which we use for fine-tuning the well-formedness T5 model. Every query in QW is scored either 0 or 1 by 5 human workers. The average score is reported as the well-formedness ratings.
Quora and MQR datasets are used to fine-tune QRT5 from the pre-trained model. The Quora dataset contains similar yet differently expressed question pairs from the online Q&A forum. The multi-domain question rewriting (MQR) [[8](#bib.bib45 "How to ask better questions? a large-scale multi-domain dataset for rewriting ill-formed questions")] dataset consists of ill-formed and well-formed question pairs, for example, “Spaghetti carbonara, mixing” is paired with “How to mix a spaghetti carbonara?”.
An internal log dataset of 300k queries with intent labels produced by human agents is leveraged for intent classification and out-of-sample analysis.
###
3.2 AQA Framework
We directly use the same downloaded pre-trained checkpoint in [[4](#bib.bib9 "Ask the right questions: active question reformulation with reinforcement learning")] as the starting model for different RL variants. Its architecture is based on GNMT [[52](#bib.bib11 "Google’s neural machine translation system: bridging the gap between human and machine translation")], pre-trained on UN parallel corpus and Paralex for general paraphrasing capabilities. Note that we do not use the CNN selector from AQA as the focus is to produce a single reformulator. The same BiDAF QA model pre-trained on SearchQA is used as the black-box QA environment. The reformulator policy passes a reformulation batch to the QA environment to obtain a batch of character-level F1 scores. As mentioned, the baseline in this approach is the mean reward of the batch. In Section [4.4.1](#S4.SS4.SSS1 "4.4.1 SearchQA ‣ 4.4 Qualitative Comparisons ‣ 4 Experimental Results ‣ Exploring Fluent Query Reformulations with Text-to-Text Transformers and Reinforcement Learning"), a downloaded RL-tuned AQA reformulator is also used to compare generation quality with other methods.
####
3.2.1 Reinforcement Learning
The advantage actor-critic approach is a common modification for policy gradient methods where Ai=R(ri)−Bi is the advantage estimate. We use a two-layer neural net fc as an on-policy critic network to estimate the value for each batch. The output embedding from the GNMT encoder πEθ is used as the input for this critic network. The value becomes the baseline Bi=fc(πEθ(⋅|qi)) and it is an estimation of the expected reward for the batch, given the current model’s encoding of the input.
We test another common approach to tune Seq2Seq models, the self-critical training [[37](#bib.bib5 "Self-critical sequence training for image captioning")]. First proposed for image captioning, this method uses the reward generated by the greedy output rigreedy∼πθ(⋅|qi) as the baseline Bi=R(rigreedy) in the policy gradient formulation. This is to encourage the model to outperform the greedy decoding strategy.
Beside varying the RL algorithms, we test methods that explicitly encourage fluency as reformulations produced by AQA often contain repetition of words and phrases. Unlikelihood training [[48](#bib.bib4 "Neural text generation with unlikelihood training")] is proposed as an extra term to regularize the loss function and explicitly suppress the likelihood of negative candidate tokens (repetitions) Ct={r1,⋯,rt−1}∖{rt} in a reformulation sequence r={r1,⋯,rT}. In this method, the following loss is weighted by the advantage estimate with the mean reward as the baseline:
| | | |
| --- | --- | --- |
| | T∑t=1[−α∑c∈Ctlog(1−p(c|r<t)+LNLL] , | |
| | | |
| --- | --- | --- |
| | where LNLL=−logp(rt|r<t,q) | |
Another addition is the fluency metric from [[17](#bib.bib1 "Fluency boost learning and inference for neural grammatical error correction")], which is proposed for error correcting sequence generation and inference:
| | | |
| --- | --- | --- |
| | Rf=11+H(r) , | |
where H(r)=−1T∑Tt=1logp(rt|r<t,q). This metric ranges between 0 and 1 and incorporates the probabilities produced by the model. We use it as an extra reward signal on top of the F1 reward R(r) from the QA environment.
###
3.3 QRT5 Framework
With recent advances of transfer learning in NLP, we want to leverage general language representation power encoded in a pre-trained transformer-based model. Due to pre-training and architectural limitations of AQA, we leverage the T5 baseline (T5-base) model to replace the LSTM-based reformulators. Hugging Face [[51](#bib.bib13 "HuggingFace’s transformers: state-of-the-art natural language processing")] and PyTorch Lightning [[14](#bib.bib14 "PyTorch lightning")] are leverged in the implementation. This allows us to fine-tune on supervised tasks with more flexibility and to create our own starting points for RL. T5 has a similar encoder-decoder structure as the original transformer [[47](#bib.bib52 "Attention is all you need")], which was designed for Seq2Seq tasks. T5 formulates any language tasks into the text-to-text format, where a prefix description of the task is attached to each input, instructing the model to perform the task through text without having to vary much of the training pipeline.
####
3.3.1 Supervised Paraphrasing and Denoising
Pre-training of Seq2Seq models before RL is a necessary step [[4](#bib.bib9 "Ask the right questions: active question reformulation with reinforcement learning"), [21](#bib.bib43 "Tuning recurrent neural networks with reinforcement learning"), [7](#bib.bib44 "On the weaknesses of reinforcement learning for neural machine translation")]. We tune a new RL starting point to replace the translation-based model from AQA that has 210 million parameters. This capacity is close to the T5-base model with 220 million parameters. It is suggested in [[8](#bib.bib45 "How to ask better questions? a large-scale multi-domain dataset for rewriting ill-formed questions")] that Quora (paraphrase) and MQR (denoising) create a good combination for improving query qualities so they are are used for fine-tuning. For both datasets, we add prefix “paraphrase: ” to the input sequence and special end suffix “<∖s>” to both the source and target sequences. We lightly tune for two epochs respectively.
####
3.3.2 Reinforcement Learning
We adopt two RL approaches similar to those in the AQA framework mentioned before. In particular, since the policy gradient (PG) baseline shows the best QA performance in Section [4.1](#S4.SS1 "4.1 AQA Framework for RL ‣ 4 Experimental Results ‣ Exploring Fluent Query Reformulations with Text-to-Text Transformers and Reinforcement Learning"), we focus on this method for further RL of QRT5. Self-critical (SC) training, being another common approach with decent performance, is used as an alternative. For modularity, the implementation leverages the T5 module from Hugging Face. QRT5 continues the reformulation task during RL, so the same prefix and suffix are added as mentioned in the above section.
###
3.4 T5 for Query Well-formedness
For a non-algorithmic numeric proxy for sequence fluency, we use the QW dataset to fine-tune a separate T5-base model, which produces automatic scores. We cast regression as a text-to-text task by generating a text string of the average rating score and compare it against the true label, e.g., a score of 3.0 becomes “3.0”. The prefix for this task is “query wellformedness: ”. Since the labeled scores are averaged among 5 humans, there are 6 categories from 0.0 to 1.0. Therefore, this is regarded as a 6-way supervised classification task in the text-to-text framework similar to how the STS-B regression task is formulated by [[36](#bib.bib20 "Exploring the limits of transfer learning with a unified text-to-text transformer")]. This model judges how well-formed or fluent a given query is based on human evaluations. It is fine-tuned for 50 epochs on QW, when validation set accuracy no longer improves.
###
3.5 Intent Classification
To demonstrate the flexibility of QRT5, we experiment with a pre-trained BERT-based intent classification (IC) model [[54](#bib.bib68 "A financial service chatbot based on deep bidirectional transformers")] as another fixed downstream system to produce reward signals. We first use the fine-tuned QRT5 to reformulate every test set query naïvely in the internal query log dataset and measure accuracy and F1 scores on the intent classes predicted by the pre-trained IC model. We compare this to the original IC test set performance. In addition, RL is applied for 5 epochs on the training set to further adapt QRT5 to the IC enviornment. Rewards are engineered as follows: when the predicted intent class exactly matches the true class, QRT5 receives a reward of 1. Otherwise, leveraging the hierarchical structure of intent classes, if a match between the parent label of the predicted class and true class occurs, QRT5 gets a partial reward of 0.5. If none of the above, the IC system gives 0 reward. Again, we measure accuracy and F1 scores on the test set with reformulations produced by QRT5 with RL. These comparisons are meant to show whether or not QRT5’s reformulations from interactive RL can be adapted to an black-box IC model that was pre-trained on the original noisy text, similar to the setup of QA.
4 Experimental Results
-----------------------
###
4.1 AQA Framework for RL
Figure 1: SearchQA Dev Set F1 Reward Curves with AQA
Figure [1](#S4.F1 "Figure 1 ‣ 4.1 AQA Framework for RL ‣ 4 Experimental Results ‣ Exploring Fluent Query Reformulations with Text-to-Text Transformers and Reinforcement Learning") plots the mean validation (dev) set F1 scores on SearchQA during RL of the reformulator. The baseline method is the original AQA approach and the rest are variations that we mentioned in the previous section. Most of the variations are able to adapt to the RL objective and learn from character-level F1 scores through trial and error except for the unlikelihood objective. It did not integrate with the model as well as we expected to reduce the number of repetitions and gain rewards. The addition of a critic network only surpasses the baseline slightly at first and becomes flat quickly. When we reduce the max length of the reformulation output produced by the translation model, the method with a critic network improves its performance. However, forcing the model to produce shorter reformulations is not ideal. The model must learn when to stop by itself. After RL, we notice most of these methods tend to generate long sequences close to the default maximum of 50, including the best-performing PG baseline. Furthermore, we observe variations in the RL algorithm fail to outperform the original PG baseline meaningfully in learning the reward under the AQA framework. The closest reward curve to the baseline method is having a mixed loss function that combines PG and SC objectives, with the addition of the fluency metric as an extra reward. In the next section, we experiment with QRT5 that can both retain fluency and optimize QA performance with RL. Compared to other methods except for the PG baseline, self-critical training, when combined with the policy gradient objective, performs the best under the AQA framework, and achieves the closest performance to the baseline. We adopt SC as an alternative to the PG method in the QRT5 framework.
###
4.2 QRT5 Framework for RL
Figure 2: SearchQA Dev Set F1 Reward Curves with QRT5
In Figure [2](#S4.F2 "Figure 2 ‣ 4.2 QRT5 Framework for RL ‣ 4 Experimental Results ‣ Exploring Fluent Query Reformulations with Text-to-Text Transformers and Reinforcement Learning"), when QRT5 models are tuned with RL, we observe better sample efficiency with faster reward acquisition. QRT5’s reward on the validation set is initially lower than the AQA baseline model. As RL progresses, the reward grows quickly after a few epochs. The reward curves for PG methods do eventually drop, and the SC methods turn flat. When the gradient is tracked, its norm becomes large. We suppose this is due to RL optimization taking steps in adverse directions and not able to recover. Note that we track the reward curves on the validation set rather than the training set to evaluate generalization. The training set reward curves do not drop, which suggests that overfitting can occur in the RL stage for T5 models. When the reward drops on the dev set, we observe the model produces more repetitions similar to AQA, losing query fluency. Changing the learning rate scheduler, entropy hyperparameters, or learning rates do not resolve this issue. Therefore, we pick the best-performing PG model with dev set F1 score around 0.37 as our model for qualitative analysis in Section [4.4](#S4.SS4 "4.4 Qualitative Comparisons ‣ 4 Experimental Results ‣ Exploring Fluent Query Reformulations with Text-to-Text Transformers and Reinforcement Learning"). Although the T5 model cannot train for longer without sacrificing rewards, it learns almost 3 times faster than the AQA model and reaches a comparable QA performance with only 0.01 difference in F1 as AQA reformulator eventually reaches 0.38 dev F1 score after 9 days of training on a single Tesla T4 GPU.
###
4.3 Well-formedness
We fine-tune a T5-base model on the QW dataset to quantitatively measure fluency of query sequences. After fine-tuning, the well-formedness model achieves 42.32% for 6-way classification on the test set.
The QW model predicts 0.0 and 1.0 with 72% and 95% accuracy respectively. Intermediate scores classes are less accurate. The model puts predictions of these classes in neighbouring score categories as averaged absolute difference between predictions and labels are around 0.3. This is slightly higher than the difference between two score categories. So we believe this discrepancy is acceptable. The well-formedness study [[15](#bib.bib46 "Identifying well-formed natural language questions")] focuses on binary accuracy using 0.8 as a threshold to determine whether a question is well-formed. Using the same threshold to group the multi-class predictions, the binary classification accuracy is 79.56%, which is better than 70.7% reported by the best model in the original paper, and close to the accuracy of 81.6% from a BERT-based model [[6](#bib.bib48 "Identifying well-formed questions using deep learning")]. Overall, we believe this model is a decent human proxy to judge the coherence and fluency of reformulations even though there is room for improvement to reach the human upper-bound binary accuracy of 88.4% reported in [[15](#bib.bib46 "Identifying well-formed natural language questions")]. With this well-formedness model to produce fluency scores, in the last column of Table [1](#S4.T1 "Table 1 ‣ 4.3 Well-formedness ‣ 4 Experimental Results ‣ Exploring Fluent Query Reformulations with Text-to-Text Transformers and Reinforcement Learning"), we see that when using QRT5 models to reformulate all queries in the SearchQA dev set, the average well-formedness scores can improve significantly. Fluency score distributions produced by QRT5 and AQA before and after the PG method are also compared, and we observe that RL can hurt well-formedness of reformulations produced by both models. However, QRT5 can retain more fluency compared to its AQA counterpart. In addition, it is worth noting that the average score of 300k raw queries in our internal log dataset is 0.5015, compared to 0.0275 of SearchQA dev set. This shows that SearchQA is a challenging set for fluency, at least more ill-formed than real-world queries.
| Model | 0.0 | 0.2 | 0.4 | 0.6 | 0.8 | 1.0 | Mean Score |
| --- | --- | --- | --- | --- | --- | --- | --- |
| Original Queries | 89.03 % | 9.49 % | 0.70 % | 0.36 % | 0.36 % | 0.07 % | 0.0275 |
| AQA (no RL) | 55.90 % | 18.92 % | 8.78 % | 3.38 % | 6.44 % | 6.58 % | 0.2106 |
| QRT5 (no RL) | 5.67 % | 7.56 % | 6.11 % | 6.59 % | 13.29 % | 60.78 % | 0.7933 |
| AQA (RL) | 88.13 % | 9.05 % | 1.14 % | 0.07 % | 0.47 % | 1.13 % | 0.0381 |
| QRT5 (RL) | 70.91 % | 18.11 % | 2.89 % | 2.14 % | 3.70 % | 2.26 % | 0.1128 |
Table 1: Comparison of Predicted Well-formedness Score Distributions (proportions of queries in different score categories are shown in percentages) & Mean Scores on SearchQA Validation Set
###
4.4 Qualitative Comparisons
| Model | Query 1 | Query 2 |
| --- | --- | --- |
| 0. Original Query | 1909 nobel prize winner failed entrance exams univ bologna , italy | 2000 film jackie chan old west could funnier?” |
| 1. Downloaded pre-trained AQA (no RL) | How many won univ bologna ’s italy nobel prize won? | Where is the chan old west west? |
| 2. Downloaded pre-trained AQA (with RL) | What is is 1909 nobel prize winner failed entry exams univ bologna , italy name? | What is is is 2000 is is chan jackie chan old west might funnier name? |
| 3. Policy gradient AQA baseline (PG) | What is 1909 nobel prize winner failed entrance exams univ bologna , italy name name? | What is 2000 film jackie chan Old west might funnier name? |
| 4. Self-critical training AQA (SC) | Name of orange nobel prize winner exam bologna ’ italy name 02 name univ bologna ””” | Where is chan old west west west horse is located 2000 west west pogamerum is funnier name 2000 chan chan old west west name |
| 5. PG + SC + fluency metric AQA | Where is the name of the 1909 nobel prize win entrance exams univ bologna , italy? | What is the name jackie chan old west is funnier from 2000 chan old west west is chan old west is it made from the is it |
| 6. Fine-tuned QRT5 (no RL) | How to prove that a 1909 Nobel Prize winner failed entrance exams at the University of Bosnia and Herzegovina, Italy? | Can the 2000 film Jackie Chan’s Old West be funnier? |
| 7. QRT5 with PG | What 1909 nobel prize winner failed entrance exams univ bologna, italy? | Is 2000 film Jackie chan old west could funnier? |
Table 2: Comparison of Reformulations by Different Methods on SearchQA Queries
####
4.4.1 SearchQA
Qualitatively, when given an original query, reformulation qualities can vary for different models. For each approach, we generate reformulations with the best performing model on the validation set and decode greedily. Note that in Table [2](#S4.T2 "Table 2 ‣ 4.4 Qualitative Comparisons ‣ 4 Experimental Results ‣ Exploring Fluent Query Reformulations with Text-to-Text Transformers and Reinforcement Learning"), all reformulations can get the correct answer with an F1 score of 1. We observe that the first model without RL does not understand what the question is asking for. Models 2,3, and 4 are relatively more fluent, and model 4 does more exploration in picking the words but the reformulation gets less coherent. Model 5 does relatively well on reward acquisition as shown in Figure [1](#S4.F1 "Figure 1 ‣ 4.1 AQA Framework for RL ‣ 4 Experimental Results ‣ Exploring Fluent Query Reformulations with Text-to-Text Transformers and Reinforcement Learning"), but it does not help the quality of this particular query even though the fluency metric is used as an extra reward. For QRT5 models 6 and 7, we notice improvements in coherence. However, before RL, QRT5 struggles to understand the purpose of the query so it reformulates the query into a general “how” question based on its general knowledge from pre-training on a large corpus. After RL, it asks a “what” question with minimal expansion of the original sequence. This may be due to RL reward signals revealing that these QA queries are often asking for a certain kind of entity, which in this case, is the Nobel prize winner.
With Query 2, again, all models manage to get the answer exactly with the max F1 reward. Notice that the SearchQA dataset consists of mostly ungrammatical and very noisy queries, similar to their concatenated snippets retrieved from the Google search engine. Therefore, it is difficult for our models to generalize and reformulate by transfer learning or RL-tuning alone without any additional information, a hard task even for us humans. For instance, as another example from SearchQA, a sequence of word salad: “blue river runs black forest black sea” is almost impossible to reformulate without any context. In Table [2](#S4.T2 "Table 2 ‣ 4.4 Qualitative Comparisons ‣ 4 Experimental Results ‣ Exploring Fluent Query Reformulations with Text-to-Text Transformers and Reinforcement Learning"), model 3 trained by us is the more fluent variant with less repetitions compared to the downloaded model 2, even though the two approaches are the same. QRT5 models are constrained by the word “funnier”, and the reformulation asks a yes-or-no question. It is curious to see that the black-box BiDAF QA environment can still give the correct answers. Therefore, as long as certain keywords are present, this QA system can give the answer. This is also mentioned in the language analysis in [[3](#bib.bib49 "Analyzing language learned by an active question answering agent")]. However, in some cases, we note that the environment can never output the correct answer no matter how we try to reformulate queries as humans. This can be stemmed from the idiosyncrasy of the particular QA environment, which is trained on the noisy original dataset. Although it may not be entirely reliable, we use this pre-trained system to compare with the previous work.
####
4.4.2 Internal Dataset
| Original Query | QRT5(no RL) | QRT5(RL) | AQA(no RL) | AQA(RL) |
| --- | --- | --- | --- | --- |
| Process for One time wire | What are the best methods for creating one time wire? | What is the process for a one time wire? | Process for one time wire? | What is one time wire process for one time wire name name… |
| this work object hasnt been touched in two weeks and is just sitting there | What can I do to solve this working object that hasn’t been touched in two weeks and is sitting there? | What to do about a work object that hasn’t been touched in two weeks and is just sitting there? | This work object has been touched in two weeks and just sitting? | What is this work object hasnt been touched in two weeks and just sitting have just name? |
| how to reinvest for an account? | How to fund reinvestments to an account? | How do I reinvest for an account? | How do you reinvest a account? | What is how to reinvest a account? |
Table 3: Comparisons between AQA and QRT5 Reformulations on Internal Queries
We sample from the internal log dataset of queries and evaluate AQA models and QRT5 models’ generalization on this out-of-domain dataset before and after RL from SearchQA. For each input query, we generate multiple reformulations with beam search and pick ones with the best qualities. We notice that when original queries are already well-formed and complete, T5 makes minimal changes. Otherwise, we get reformulations that transform and expand incomplete queries into proper questions that are aligned with the original intent semantically as shown in Table [3](#S4.T3 "Table 3 ‣ 4.4.2 Internal Dataset ‣ 4.4 Qualitative Comparisons ‣ 4 Experimental Results ‣ Exploring Fluent Query Reformulations with Text-to-Text Transformers and Reinforcement Learning"). Compared with QRT5, it is clear that out-of-sample reformulations generated by AQA reformulators are more rigid and prone to repetitions and errors before and after RL on SearchQA, which is also corroborated in Table [2](#S4.T2 "Table 2 ‣ 4.4 Qualitative Comparisons ‣ 4 Experimental Results ‣ Exploring Fluent Query Reformulations with Text-to-Text Transformers and Reinforcement Learning"). This phenomenon is addressed in [[4](#bib.bib9 "Ask the right questions: active question reformulation with reinforcement learning")] and [[3](#bib.bib49 "Analyzing language learned by an active question answering agent")], where the reformulated language is regarded as an instance of machine-to-machine translation, and repetitions are acceptable through the lens of information retrieval. However, qualitative analysis suggests that AQA models that learn to communicate between machines are not adequate in the case when reformulations need to be fluent enough to be shown, and examined by human users in a client-facing QA setting.
###
4.5 Intent Classification
| Metrics | Original IC System | QRT5-IC (RL) | QRT5-IC (no RL) |
| --- | --- | --- | --- |
| Accuracy | 0.6010 | 0.6207 | 0.5883 |
| F1 Score | 0.6064 | 0.6267 | 0.5941 |
Table 4: Intent Classification Performance with QRT5: QRT5-IC (no RL) represents test performance using naïve reformulations with a fine-tuned QRT5 model, QRT5-IC (with RL) uses the PG method to learn from engineered IC reward signals
In Table [4](#S4.T4 "Table 4 ‣ 4.5 Intent Classification ‣ 4 Experimental Results ‣ Exploring Fluent Query Reformulations with Text-to-Text Transformers and Reinforcement Learning"), we find that naive reformulations hurts the performance in accuracy and F1 scores. After RL, the accuracy and F1 scores of the IC system can be improved with reformulated queries from QRT5. This means even though supervised fine-tuning of QRT5 before RL can improve the fluency of queries by a large margin as shown in Table [1](#S4.T1 "Table 1 ‣ 4.3 Well-formedness ‣ 4 Experimental Results ‣ Exploring Fluent Query Reformulations with Text-to-Text Transformers and Reinforcement Learning"), naïve reformulations prehaps drift away from original intended purposes due to the discrepancy between the fine-tuning data (MQR and Quora) and OOD internal query logs. The model may emphasize on generation quality excessively during the fine-tuning stage at a cost of losing original intents. Thus, it pays to further leverage RL for the reformulator to adapt to reward signals from downstream environments like IC. This can correct the course from semantic drift, even though in the RL process, sequence fluency is impaired as the trade-off.
5 Related Work
---------------
Our work is similar to the task of paraphrasing. This is to restate a given sequence to preserve the same meaning. [[50](#bib.bib19 "Paraphrasing with large language models")] focus on fine-tuning GPT-2 [[35](#bib.bib23 "Language models are unsupervised multitask learners")] with supervised datasets. Algorithmic n-gram metrics are used to find the best paraphrase. The models demonstrate the ability to generalize on OOD text. Another common approach is to leverage machine translation. [[18](#bib.bib28 "Zero-shot paraphrase generation with multilingual language models")]’s work uses multilingual translation to pivot for zero-shot paraphrases. Human evaluators are employed in this process to evaluate fluency. [[38](#bib.bib37 "Unsupervised paraphrasing without translation")] uses a VQ-VAE [[33](#bib.bib38 "Neural discrete representation learning")] to compare a monolingual paraphrasing method with unsupervised and supervised translation approaches, among other generative techniques [[53](#bib.bib34 "An end-to-end generative architecture for paraphrase generation"), [19](#bib.bib35 "A deep generative framework for paraphrase generation")]. Using Deep RL to paraphrase has been attempted. With reward shaping and policy gradient in [[24](#bib.bib70 "Paraphrase generation with deep reinforcement learning")], a generator paraphrases by learning rewards from an trained evaluator that can judge whether two sentences are paraphrases. Transitioning from an unsupervised VAE model to RL is possible on non-parallel data with reward engineering based on characteristics of good paraphrases [[42](#bib.bib71 "Unsupervised paraphrasing via deep reinforcement learning")]. However, these studies merely focus on generation quality rather than gaining performance on any downstream systems. Our RL framework aims to retain reformulation quality as well as performance in downstream applications. This is most related to [[4](#bib.bib9 "Ask the right questions: active question reformulation with reinforcement learning")]’s AQA approach by leveraging policy-based RL methods to generate question reformulations with a GNMT reformulator [[52](#bib.bib11 "Google’s neural machine translation system: bridging the gap between human and machine translation")] and a CNN [[22](#bib.bib22 "ImageNet classification with deep convolutional neural networks")] selector. The reformulations are treated as inputs to a BiDAF [[41](#bib.bib12 "Bidirectional attention flow for machine comprehension")] QA system that generates character-level F1 rewards. Rather than a translation-based model, our approach fine-tunes T5 [[36](#bib.bib20 "Exploring the limits of transfer learning with a unified text-to-text transformer")] to leverage the flexibility in this framework and the linguistic prior encoded by the text-to-text model from pre-training. [[26](#bib.bib36 "Conversational question reformulation via sequence-to-sequence architectures and pretrained language models")] has shown T5’s good performance on reformulating questions along with context within a conversational history. While in our work, only original queries are considered as input to the black-box QA environment during RL for fair comparison. Identifying well-formed questions [[15](#bib.bib46 "Identifying well-formed natural language questions")] by training binary classification models has been studied using BERT [[6](#bib.bib48 "Identifying well-formed questions using deep learning")] and transfer learning with pre-trained models [[46](#bib.bib3 "Inductive transfer learning for detection of well-formed natural language search queries")]. We investigate more fine-grained 6-way classification using a fined-tuned T5 for regression, leveraged as a proxy to evaluate fluency of reformulations.
6 Conclusions
--------------
We show that QRT5 can optimize the RL objective to reformulate queries, adapting to reward signals sourced from black-box systems like question answering and intent classification. RL variants based on policy-gradient and self-critical training can achieve better downstream performance compared to other alternatives. We find during RL, reformulators struggle to maintain the ability to generate fluent queries compared to before RL. Transfer learning under the text-to-text framework has proven critical for reformulation models to retain fluency. It provides flexibility to fine-tune on paraphrasing and denoising, as well as creating a prediction model to evaluate fluency on reformulations, driven by human evaluations. The fine-tuned QRT5 model is capable of generating reformulations with quality for out-of-sample queries before RL. This provides a more robust starting point than the previous approach for later RL tuning. After RL in a specific black-box environment such as QA, QRT5 demonstrates its ability to maintain fluency qualitatively and quantitatively while acquiring rewards from the downstream task. Finally, as text-to-text is more flexible in nature, swapping out the QA systems, systematically studying generalization on other out-of-domain datasets, and adding conversational context to produce more informed reformulations are promising future directions. |
1897b7e8-b99f-418d-b267-0e5b1342e526 | trentmkelly/LessWrong-43k | LessWrong | China’s Plan to ‘Lead’ in AI: Purpose, Prospects, and Problems
|
0b97e873-5286-4cac-ba62-9b95fc2fd6c6 | trentmkelly/LessWrong-43k | LessWrong | Instrumental Rationality 4.1: Modeling Habits
[Instrumental Rationality Sequence 4.1/7.]
[This is part 1 of a 3-part sequence on habits, which is itself part of the greater Instrumental Rationality Sequence. This was initially one monstrous article; in the interests of readability, I've decided to split it into three essays.]
Outline:
The Habits 101 mini-sequence is broken up into 3 sections:
1. Introduction, Models, and Statistics:
We cover a basic model of how habits work, three of their properties (insensitivity to reward changes, independence of intentions, and automaticity), and some closing remarks on base rates for habituation.
2. Techniques for Creating Habits:
We cover three techniques for habit creation: Trigger Action Planning (TAPs), Systematic Planning (which has three sub-techniques), and Scaling Up.
3. Techniques for Breaking Habits and Conclusion:
We cover two techniques for breaking habits: Going Upstream and Substitution (which has two sub-techniques).
(The current evidence base has many more interventions for forming habits than breaking them, so that's why there's an asymmetry between parts 2 and 3. Also, this is probably just a good thing to keep in mind, the fact that forming habits is easier than breaking existing ones.)
++++
Introduction:
People, as the saying goes, are creatures of habit. Many of our actions every day are repeated often, typically without much thought.
This type of thoughtlessness, though, isn’t necessarily bad. Habits can help reduce cognitive load, allowing us to get through the day. Imagine if you had to explicitly weigh the pros and cons of behavior like flushing the toilet every single time you did them.
Making in-depth, reasoned decisions all the time can be very costly in terms of both time and attention. Habits allow us to compartmentalize certain behaviors, so that our energy and focus can move to other, perhaps more important, things.
A frequent part of our lives, habits make up at least roughly 40% of our eve |
91d149b4-2d4c-4565-914e-6c569c599243 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Disentangling Perspectives On Strategy-Stealing in AI Safety
*This post was written under Evan Hubinger’s direct guidance and mentorship, as a part of the Stanford Existential Risks Institute ML Alignment Theory Scholars (MATS) program.*
*Additional thanks to Ameya Prabhu and Callum McDougall for their thoughts and feedback on this post.*
**Introduction**
I’ve seen that in various posts people will make an offhanded reference to “strategy-stealing” or “the strategy-stealing assumption” without very clearly defining what they mean by “strategy-stealing”. Part of the trouble with this is that these posts are often of wildly different flavors, and it’s often not clear what their connection to each other is, or perhaps it’s unclear under what conditions it might be feasible to “steal the strategy” of an AI, or sometimes it’s unclear where there is any notion of doing any stealing of strategies at any point.
Spoiler upfront, part of the reason for this is that the term comes from a relatively esoteric game-theoretic concept that basically never applies in a direct sense to the scenarios in which it is name-dropped. In this post, I’ll try to explain the common use of this seemingly-inappropriate term and why people are using it. In order to do this, I’ll first attempt to develop a more general framework for thinking about strategy-stealing, and then I’ll try to extend the intuition that develops to bridge the gap between a couple of seemingly highly disparate posts. Along the way I’ll attempt to clarify the connection to a few other ideas in alignment, and clarify what implications the strategy-stealing framework might have for the direction of other alignment research. I’ll declare up front that I think that in an AI safety context, the strategy-stealing framework is best thought of as a tool for managing certain intuitions about competition; rhetorically the goal of my post is to distill out those aspects of “game-theoretic” strategy-stealing that are broadly applicable to alignment so that we can mostly throw out the rest.
As an aside, I think that due to the (very relatable) [confusion I’ve seen elsewhere](https://www.alignmentforum.org/posts/nRAMpjnb6Z4Qv3imF/the-strategy-stealing-assumption?commentId=HBiQLbNmT3MNLoY9p), strategy-stealing is not actually a good name for the concept in an AI safety context, but the term is already in use and I don’t have any ideas for any better ones, so I’ll continue to just use the term strategy-stealing even in contexts where no strategies are ever stolen. My hope is that I do a good enough job of unifying the existing uses of the term to avoid a [proliferation of definitions](https://xkcd.com/927/).
**Strategy Stealing in Game Theory**
The strategy-stealing argument page on Wikipedia has a very good short explanation of the term, reproduced here:
*In* [*combinatorial game theory*](https://en.wikipedia.org/wiki/Combinatorial_game_theory)*, the **strategy-stealing argument** is a general* [*argument*](https://en.wikipedia.org/wiki/Argument) *that shows, for many* [*two-player games*](https://en.wikipedia.org/wiki/Two-player_game)*, that the second player cannot have a guaranteed* [*winning strategy*](https://en.wikipedia.org/wiki/Winning_strategy)*. The strategy-stealing argument applies to any* [*symmetric game*](https://en.wikipedia.org/wiki/Symmetric_game) *(one in which either player has the same set of available moves with the same results, so that the first player can "use" the second player's strategy) in which an extra move can never be a disadvantage.*
*The argument works by obtaining a* [*contradiction*](https://en.wikipedia.org/wiki/Reductio_ad_absurdum)*. A winning strategy is assumed to exist for the second player, who is using it. But then, roughly speaking, after making an arbitrary first move – which by the conditions above is not a disadvantage – the first player may then also play according to this winning strategy. The result is that both players are guaranteed to win – which is absurd, thus contradicting the assumption that such a strategy exists.*
Although they were made pretty clear, note explicitly the following assumptions and limitations of the argument:
* The game is a two-player, turn-based game [[1]](#fn-37vptw4wtAd5riR7K-1).
* The game is symmetric (i.e, both players have access to the same set of actions; even more explicitly, both players’ actions induce winning states for themselves under identical conditions). [[2]](#fn-37vptw4wtAd5riR7K-2)
* Excess moves never put you at a disadvantage (though this point is not particularly impactful for the following discussion). [[3]](#fn-37vptw4wtAd5riR7K-3)
* Most importantly, the proof is non-constructive, i.e, it doesn’t tell you how to actually deny P2 a win as P1. The strategy-stealing argument is used to make a claim about whether P2 has a guaranteed winning strategy assuming optimal play. We’re able to obtain a contradiction to this claim by assuming that if P2 had a winning strategy, P1 would be able to deduce what it is. For sufficiently computationally complex games, this either requires extremely large amounts of computational power (to search the game tree for P2’s optimal strategy), or at least some sort of insight into P2’s available set of policies. For example that the strategy-stealing argument can be used to show that P1 never loses [gomoku](https://en.wikipedia.org/wiki/M,n,k-game#Strategy_stealing_argument) under theoretically optimal play, but between, say, humans, P2 routinely wins. [[4]](#fn-37vptw4wtAd5riR7K-4)
**Strategy Stealing Intuitions For Competition Among AI Agents: Paul Christiano’s Model**
In [The Strategy-Stealing Assumption](https://www.alignmentforum.org/posts/nRAMpjnb6Z4Qv3imF/the-strategy-stealing-assumption), Paul Christiano says to a crude approximation that “*If you squint kinda hard, you can model trying to influence the future as a game which is symmetric with respect to some notion of power.”* More concretely, he makes the following argument:
* Model the world as consisting of coalitions consisting of humans and AIs, some of which may be unaligned, and model these coalitions as each controlling some share of generic “resources”.
* Assume that a coalition’s ability to acquire “influence” over the future is directly proportional to their control over generic resources. This point is what Paul in this context calls the “strategy-stealing assumption”; in his words: “*for any strategy an unaligned AI could use to influence the long-run future, there is an analogous strategy that a similarly-sized group of humans can use in order to capture a similar amount of flexible influence over the future”.*
* Then if, for example, the majority of coalitions are approximately aligned with human values, then we expect human values to win out in the long run. [[5]](#fn-37vptw4wtAd5riR7K-5)
**Other Assumptions and Sources of Justification for Paul’s Model**
You might have noticed by now that even in principle, strictly speaking the game-theoretic notion of strategy-stealing doesn’t apply here, because we’re no longer in the context of a two-player turn-based adversarial game. Actually, instead of what’s commonly meant by “strategy-stealing” in game theory, Paul justifies his model by making reference to Jessica Taylor’s [Strategies for Coalitions In Unit Sum Games](https://www.alignmentforum.org/posts/5bd75cc58225bf0670375325/strategies-for-coalitions-in-unit-sum-games). In this setting, we’re actually talking about games with multiple agents acting simultaneously, and where payoffs are not necessarily binary but merely unit sum [[6]](#fn-37vptw4wtAd5riR7K-6).
I think intuitively Paul justifies the use of the term “strategy-stealing” with the idea that at least in a very similar fashion, we use the symmetry of the game to come to the intuitive conclusion that “coalitions can take advantage of public knowledge, i.e, steal strategies, to obtain power approximately proportionate to their size”. Actually, the results in Jessica’s post are even weaker than that-- Theorem 1 only shows that the conclusion holds for a very, very specific type of game, and Theorem 2 assumes that the coalition is of a certain size, and also assumes prior knowledge of other players’ strategies. I personally don’t think that these theoretical results justify the assumptions of Paul’s model in the more general setting that he describes particularly strongly, so it’s not surprising if you can come up with critiques to those assumptions. Anyway, none of this really affects the structure of the rest of the post, since the whole point rhetorically is that we ultimately just want to use intuitions that game theory helps us develop.
\*\*Problems Applying the Strategy-Stealing Framework \*\*
Rhetorically, the position of Paul’s post is that [we may want to find ways to make this “strategy-stealing assumption” approximately true,](https://www.alignmentforum.org/posts/nRAMpjnb6Z4Qv3imF/the-strategy-stealing-assumption?commentId=HBiQLbNmT3MNLoY9p) at which point “all” we have to do is make sure that aligned humans control a sufficiently large proportion of resources. (Of course, we’d also have to address problems with/loopholes in the model and argument above, but his post spends a lot of time rather comprehensively doing that, so I won’t redo his analysis here.) I think the problem with this position is that from a practical perspective, it is basically impossible to make this assumption true, and it would involve solving a ton of other subproblems. To elaborate, the intuitively confusing thing to me about Paul’s post is that in many commonly imagined AI takeoff scenarios, it’s fairly clear that most of the assumptions you need in the game-theoretic strategy-stealing setting do not hold, not even approximately:
* The structure of the game may not actually turn out to be amenable to analysis through the lens of “competition between agents”. Even attempting to analyze a scenario in which “a coalition of aligned humans control the majority of resources” makes a large number of assumptions, not only about the nature of human values and the ability of humans to coordinate, but also that the world we’ll eventually find ourselves in is [multipolar](https://www.alignmentforum.org/posts/mKBfa8v4S9pNKSyKK/homogeneity-vs-heterogeneity-in-ai-takeoff-scenarios). That is to say, we’re assuming the existence of some actually aligned agents, and that humans are ever in a position to actually “use” or “cooperate” with AI to compete with other coalitions. [[7]](#fn-37vptw4wtAd5riR7K-7)
* Intuitively this game is not symmetric, for a couple broad classes of reasons:
+ Possibly the AI has access to “actions” that coalitions of other humans (or even AIs) don’t, due to broadly superior intelligence, or perhaps just due the way it interfaces with the world. People have written hundreds of thousands of words about ways this can be true, so I won’t go into it that much further.
+ Possibly some values are easier to optimize for than others (i.e, influence doesn’t convert to utility, or perhaps having certain values give you extra bargaining power, e.g you might be able to threaten other coalitions because you don’t mind if the world is destroyed), which Paul provides various examples of.
* The non-constructiveness of the game-theoretic strategy-stealing argument above bites us pretty badly: even if you did have access to the “same set of actions” as an unaligned AI, intuitively you can’t easily emulate its power-acquiring strategy, for reasons broadly related to the existence of the terms “deception” and “unaligned”, unless you also want to make other really strong assumptions about the nature of future AI or future AI interpretability/alignment research. (Really, the boundary between this point and the “asymmetry of actions” point above is not that sharp. Basically, you can think of superintelligent behavior as being a fundamentally inaccessible type of action to human coalitions, or you can think of it as being a sense in which we can’t determine an AI’s intended policy due to computational constraints. Either way, I feel that these critiques of the strategy-stealing framework are broad enough that to patch them up constitutes solving most of what people typically think of as “the alignment problem”, so this isn’t a very useful factoring of the problem, but perhaps Paul and others see something I don’t here.)
\*\*Why might the strategy-stealing framework still be useful? \*\*
As I’ve suggested in the introduction and despite my tone up until now, I don’t think the strategy-stealing framework is completely worthless just because Paul’s particular research direction seems infeasible. I think that when people speak about “strategy-stealing”, they’re often pointing to a concept which is not that closely related to the game-theoretic concept, but which takes advantage of some subset of the intuitions it involves, and this can be a really useful framing for thinking somewhat concretely about potential problems for alignment in various scenarios. One major source of intuition I’m thinking of is that often, success in a game may boil down to control of a single “resource”, for which a strong policy is easy to determine, so that no actual stealing of strategies is needed:
* In Paul’s case this is what he calls “flexible influence”.
* In many simple economic models this is money or cost-efficiency, e.g, the largest of a set of undifferentiated corporations will eventually obtain a monopoly on its industry due to economies of scale.
* In StarCraft, which is a comparatively very complex game (asymmetric, real-time, imperfect information), basic conventional wisdom is that the player with greater income is going to win in the long run (modulo a bunch of other strategic considerations that make the game actually interesting, so this is only a suggestive example).
* In some sense this is the entire intuition behind the notion of [instrumental convergence](https://en.wikipedia.org/wiki/Instrumental_convergence). Some kinds of resources or objectives are sufficiently useful in a general enough sense to impose “symmetry”, in the sense that many agents want to acquire them and can do similar things with them. A very common example of this in AI-relevant scenarios is computational power.
**Ways of Factoring Potential Research Directions**
You can think of Paul’s various uses of the term “strategy-stealing” as an attempt to unify the core problems in several potential future scenarios under a single common framework. For starters, there are the 11 different objections he raises in *The Strategy-Stealing Assumption,* but thinking about the computational constraints on your ability to “steal strategies” also directly motivates his ideas on [inaccessible information](https://www.alignmentforum.org/posts/ZyWyAJbedvEgRT2uF/inaccessible-information#2__Inaccessible_info_is_a_competitive_advantage). Tangentially, you can also see that his framing of the AI safety problem compels him to think about even things like question-answering systems in the context of [the way they affect “strategy-stealing” dynamics](https://www.alignmentforum.org/posts/SRJ5J9Tnyq7bySxbt/answering-questions-honestly-given-world-model-mismatches#Problem_1__is_the__intended_answer__actually_good_enough_).
**Strategy Stealing Intuitions Within the Alignment Problem**
Up until now all the scenarios we’ve explicitly described have assumed the existence of (coalitions of) AIs aligned in different ways, and we’ve applied strategy-stealing concepts to determine how their objectives shape the far future. However, some of the intuitions about “parties competing to affect a result in a game with functionally one important resource” can be applied to processes internal to an individual machine learning model, or optimization process (which, importantly, makes this a relevant concept even in the [unipolar](https://www.alignmentforum.org/posts/mKBfa8v4S9pNKSyKK/homogeneity-vs-heterogeneity-in-ai-takeoff-scenarios) takeoff case). It’s not as clear how to think about these as it is to think about the world described by Paul Christiano, but I’ll just list some dynamics in machine learning, how they might loosely fit to a strategy-stealing framework, and suggest research directions that these might imply.
* Broadly speaking, you can think of models in parameter space competing to be the thing eventually instantiated at the end of an optimization process like SGD. In this context the resource they use to compete is something like “attractor basin size”. You can think in a similar way to apply the analysis to “qualitatively similar families of models”, where in addition to attractor basins, you also factor in the proportion of models that implement a qualitatively similar behavior (i.e, perhaps there are 20 times as many actual parameter configurations that implement an unsafe policy A as there are parameter configurations that implement a desired policy B, and you’d weight the parameter configurations by the size of their attractor basins). I think this sort of analysis might be a useful framing for understanding the [mechanistic](https://www.alignmentforum.org/posts/BKM8uQS6QdJPZLqCr/towards-a-mechanistic-understanding-of-corrigibility) dynamics of [inductive biases](https://www.alignmentforum.org/posts/fM5ZWGDbnjb7ThNKJ/are-minimal-circuits-deceptive) in neural networks.
* Subnetworks of a large neural net compete for terminal impact using the single resource of “numerical influence on final inference result”. This might be a useful framing for thinking about deceptive alignment problems, particularly things related to [gradient hacking](https://www.lesswrong.com/posts/uXH4r6MmKPedk8rMA/gradient-hacking), though I haven’t found any work that adopts this framework explicitly.
* In [Operationalizing Compatibility With Strategy-Stealing](https://www.alignmentforum.org/posts/WwJdaymwKq6qyJqBX/operationalizing-compatibility-with-strategy-stealing), Evan poses the following scenario: Suppose we successfully manage to make models robustly attempt to optimize an objective that we specify, and you specify its objective as a linear combination of some values. Then, you might find that the values compete using the single resource of “ease-of-being-optimized-for”, with the undesirable result those values which are “more tractable” (in his post, “make Google money”) to optimize might end up being optimized for in exchange for less optimizable values (“put the world in a good place”) not being realized at all.
+ This is a fairly subtle point, but Evan explicitly attempts to distinguish the two similar concepts of “how intrinsically hard something is to optimize” in the sense of the size of the space we’re optimizing over, and “ease-of-being-optimized-for” which is a measure of how readily available optimization power is applied to one objective over others. Evan’s implied research strategy here is to use strategy-stealing-ish considerations about “ways in which subsets of the specified objective might be systematically disadvantaged in an optimization process” to decompose the problem into some mathematically tractable components, which we might be able to independently prove guarantees about. I think that this particular research direction doesn’t seem very tractable due to its abstractness, but that might be just because I don’t know how to turn abstractly good ideas into working mathematics.
* As a very speculative curiosity, although I’ve intentionally steered us away from trying to explicitly invoke game-theoretic reasoning, you could attempt to think explicitly about the implications of collusion/coercion among “coalitions” in any of the above scenarios, and this would probably yield new esoteric concerns about ways alignment can fail. This is a little bit conceptually tricky, because some models which you’d think of as “colluding” in this context are never actually instantiated, but you can probably chant “[acausal trade](https://www.lesswrong.com/tag/acausal-trade)” while performing some kind of blood sacrifice in order to coherently go deeper down this rabbit hole.
---
1. Some of the arguments can probably be extended to n-player turn-based games with relatively little difficulty, certain simultaneous games with also relatively little difficulty (as we’ll see below), and probably continuous-time games with moderate difficulty. [↩︎](#fnref-37vptw4wtAd5riR7K-1)
2. This is the reason that the strategy-stealing argument can’t be used to prove a win for Black in Go with komi: the game is not actually symmetric; if you try to pass your first turn to “effectively become P2”, you can’t win by taking (half the board - komi + 0.5) points, like White can. [↩︎](#fnref-37vptw4wtAd5riR7K-2)
3. For fun, though, this is also one reason why strategy-stealing can’t be used to prove a guaranteed win/draw for White in chess, due to zugzwang. [↩︎](#fnref-37vptw4wtAd5riR7K-3)
4. Actually this isn’t the best example because first-player advantage in raw Gomoku is so huge that Gomoku has been explicitly solved by computers, but we can correct this example by imagining that instead of Gomoku I named a way more computationally complex version of Gomoku, where you win if you get like 800 in a row in like 15 dimensions or something. [↩︎](#fnref-37vptw4wtAd5riR7K-4)
5. This assumes that the proportion of “influence” over the future a coalition holds is roughly proportional to the fraction of maximum possible utility they could achieve if everyone were aligned. There are obvious flaws to this assumption, which Paul discusses in his post. [↩︎](#fnref-37vptw4wtAd5riR7K-5)
6. This means that the use of the phrase “human values… win out” above is doing a little bit of subtle lifting. Under the assumptions of Paul’s model, humans with 99% of flexible influence can achieve 99% of maximum utility in the long run. IMHO It’s a moral philosophical question whether this is an acceptable outcome; Paul bites the bullet and assumes that it is for his analysis. [↩︎](#fnref-37vptw4wtAd5riR7K-6)
7. Furthermore, depending on the exact scenario you’re analyzing you might have to make the assumption that aligned AIs are designed such that humans can effectively cooperate with them, which starts to bleed into considerations about interpretability and corrigibility. This wasn’t discussed in Paul’s original post but was in the [comments](https://www.alignmentforum.org/posts/nRAMpjnb6Z4Qv3imF/the-strategy-stealing-assumption?commentId=HBiQLbNmT3MNLoY9p). [↩︎](#fnref-37vptw4wtAd5riR7K-7) |
22465306-9c47-4dcc-9fb5-b65ce6500b0b | trentmkelly/LessWrong-43k | LessWrong | Geometric Utilitarianism (And Why It Matters)
Do you like using numbers to represent uncertainty and preference, but also care about things like fairness and consent? Are you an altruist on a budget, looking to do the most good with some of your resources, but want to pursue other goals too? Are you looking for a way to align systems to the interests of many people? Geometric Utilitarianism might be right for you!
Classic Utilitarianism
The Harsanyi utilitarian theorem is an amazing result in social choice theory, which states that if a social choice function F:Rn→R is both
* VNM-rational, and
* Pareto monotone (Pareto improvements never make F lower)
then for any joint utility u∈Rn, F(u) must be equal to a weighted average of individual utilities that looks like H(u,ϕ)=u⋅ϕ=∑ni=1uiϕi, where ⋅ is the dot product and ϕ∈[0,1]n are weights given to each agent's utility that sum up to 1.
As Diffractor puts it here in their excellent Unifying Bargaining sequence:
> Basically, if you want to aggregate utility functions, the only sane way to do so is to give everyone importance weights, and do a weighted sum of everyone's individual utility functions.
Diffractor is using sane as a shorthand for VNM-rational here, which is extremely reasonable given the success of expected utility maximization as a model of rational decision-making. However, I have recently been radicalized by reading Scott Garrabrant's very compelling Geometric Rationality sequence, which has significantly updated my thinking on many topics in rationality, including how to sensibly combine utilities. And I wanted to see if I could prove some results about what happens if we use a geometric weighted average of utilities that looks like G(u,ψ)=∏ni=1uψii when the weights ψi∈[0,1]n sum to 1 and utilities are shifted to be non-negative. (Which I'll be assuming throughout this post.)
Results About Geometric Utilitarianism
What might it mean for a group to be rational? Well at the very least, that group had better be doing something Pareto optimal. |
cf97b558-7e22-4e68-8bec-e9f6d45a8b64 | trentmkelly/LessWrong-43k | LessWrong | Meetup : [Cambridge] Sunk Cost Kata
Discussion article for the meetup : [Cambridge] Sunk Cost Kata
WHEN: 19 May 2013 02:00:00PM (-0400)
WHERE: 21 Ames St, Cambridge, MA
We'll present the Center for Applied Rationality's material on sunk costs and go over their exercises on how to apply this knowledge in daily life.
Cambridge/Boston-area Less Wrong meetups are on the first and third Sunday of every month at 2pm in the MIT Whitaker Building (21 Ames St, Bldg 56), room 180. Room number subject to change based on availability. Signs will be posted with the actual room number. The side doors are sometimes locked; if so, you can get in through the main door at 25 Ames St.
Discussion article for the meetup : [Cambridge] Sunk Cost Kata |
0bb2db51-643b-4fba-9a96-3994b1e051f3 | trentmkelly/LessWrong-43k | LessWrong | AI labs' statements on governance
June 6, 2024: THIS POST HAS BEEN SUCCEEDED BY Companies' policy advocacy. READ THAT INSTEAD OF THIS. THIS POST WILL NOT BE MAINTAINED.
This is a collection of statements on government policy, regulation, and standards from leading AI labs and their leadership.
As of 7 August 2023, I believe this post has all of the relevant announcements/blogposts from the three labs it covers, but I expect it is missing a couple relevant speeches/interviews with lab leadership.[1] Suggestions are welcome.
My quotes tend to focus on AI safety rather than other governance goals.
Within sections, sources are roughly sorted by priority.
OpenAI
Governance of superintelligence (May 2023)
> First, we need some degree of coordination among the leading development efforts to ensure that the development of superintelligence occurs in a manner that allows us to both maintain safety and help smooth integration of these systems with society. There are many ways this could be implemented; major governments around the world could set up a project that many current efforts become part of, or we could collectively agree (with the backing power of a new organization like the one suggested below) that the rate of growth in AI capability at the frontier is limited to a certain rate per year.
>
> And of course, individual companies should be held to an extremely high standard of acting responsibly.
>
> Second, we are likely to eventually need something like an IAEA for superintelligence efforts; any effort above a certain capability (or resources like compute) threshold will need to be subject to an international authority that can inspect systems, require audits, test for compliance with safety standards, place restrictions on degrees of deployment and levels of security, etc. Tracking compute and energy usage could go a long way, and give us some hope this idea could actually be implementable. As a first step, companies could voluntarily agree to begin implementing elements of what such an |
c2d0b618-f1c8-488d-866e-afb8ddf2438e | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | MetaAI: less is less for alignment.
Summary
=======
In May 2023, MetaAI submitted a paper to arxiv called [LIMA: Less Is More for Alignment](https://arxiv.org/abs/2305.11206). It's a pretty bad paper and (in my opinion) straightforwardly misleading. Let's get into it.
The Superficial Alignment Hypothesis
====================================
The authors present an interesting hypothesis about LLMs —
> We define the Superficial Alignment Hypothesis: A model’s knowledge and capabilities are learnt almost entirely during pretraining, while alignment teaches it which subdistribution of formats should be used when interacting with users.
>
> If this hypothesis is correct, and **alignment is largely about learning style**, then a corollary of the Superficial Alignment Hypothesis is that **one could sufficiently tune a pretrained language model with a rather small set of examples**.
>
> We hypothesize that alignment can be a simple process where the model learns the style or format for interacting with users, to expose the knowledge and capabilities that were already acquired during pretraining.
>
>
(1) This hypothesis would have profound implications for AI x-risk —
* It suggests that we could build a **safe competent oracle** by pretraining an LLM on the entire internet corpus, and then finetuning the LLM on a curated dataset of safe competent responses.
* It suggests that we could build an **alignment researcher** by pretraining an LLM on the entire internet corpus, and then finetuning the LLM on a curated dataset of alignment research.
(2) Moreover, as by [Ulisse Mini](https://www.lesswrong.com/users/ulisse-mini) writes in their [review of the LIMA paper](https://www.lesswrong.com/posts/dJumQtpoKhjDKH9q8/lima-less-is-more-for-alignment),
> Along with [TinyStories](https://www.lesswrong.com/posts/dMoaBvcxpBE7LcES4/tinystories-small-language-models-that-still-speak-coherent) and [QLoRA](https://twitter.com/Tim_Dettmers/status/1661379363907031040) I'm becoming increasingly convinced that data quality is all you need, definitely seems to be the case for finetuning, and may be the case for base-model training as well. Better scaling laws through higher-quality corpus? Also for who haven't updated, it seems very likely that **GPT-4 equivalents will be essentially free to self-host and tune within a year**. Plan for this!
>
>
(3) Finally, the hypothesis would've supported many of the intuitions in the[**Simulators sequence**](https://www.lesswrong.com/s/N7nDePaNabJdnbXeE) **by Janus**, and I share these intuitions.
So I was pretty excited to read the paper! Unfortunately, the LIMA results were unimpressive upon inspection.
MetaAI's experiment
===================
The authors finetune MetaAI's 65B parameter LLaMa language model on 1000 curated prompts and responses (mostly from StackExchange, wikiHow, and Reddit), and then compare it to five other LLMs (Alpaca 65B, DaVinci003, Bard, Claude, GPT4).
**Method:**
> To compare LIMA to other models, we generate a single response for each test prompt. We then ask crowd workers to compare LIMA outputs to each of the baselines and label which one they prefer. We repeat this experiment, replacing human crowd workers with GPT-4, finding similar agreement levels.
>
>
**Results:**
> In a controlled human study, responses from LIMA are either equivalent or strictly preferred to GPT-4 in 43% of cases; this statistic is as high as 58% when compared to Bard and 65% versus DaVinci003, which was trained with human feedback.
>
>
**Conclusion:**
> The fact that simple fine-tuning over so few examples is enough to compete with the state of the art strongly supports the Superficial Alignment Hypothesis, as it demonstrates the power of pretraining and its relative importance over large-scale instruction tuning and reinforcement learning approaches.
>
>
Problems with their experiment
==============================
(1) Human evaluators
--------------------
To compare two chatbots A and B, you could ask humans whether they prefer A's response to B's response across 300 test prompts. But this is pretty bad proxy, because here's what users actually care about:
* What's the chatbots' accuracy on benchmark tests, e.g. BigBench, MMLU?
* Can the chatbot pass a law exam, or a medical exam?
* Can the chatbot **write Python code** that actually matches the specification?
* Can the chatbot perform worthwhile alignment research?
Why did the paper not include any benchmark tests? **Did the authors** **run zero tests** **other than human evaluation?** This is surprising, because human evaluation is by far the most expensive kind of test to run. [Hmm.](https://www.lesswrong.com/tag/filtered-evidence)
(2) "either equivalent or strictly preferred"
---------------------------------------------
The claim in the paper's abstract — "responses from LIMA are either equivalent or strictly preferred to GPT-4 in 43% of cases" — sounds pretty good when they lump "equivalent" and "strictly preferred" together.
Anyway, here's the whole thing:
Moreover, "equivalent" doesn't actually mean that the human evaluator thought the responses were equivalent. Instead, it means that the evaluator thought that "neither response is significantly better".
Here's my estimate[[1]](#fny5lprvj38os) for the comparisons, eliminating ties:
* LIMA (64%) vs Alpaca (36%)
* LIMA (54%) vs DaVinci003 (46%)
* LIMA (45%) vs Bard (55%)
* LIMA (34%) vs Claude (66%)
* LIMA (29%) vs GPT-4 (71%)
Do you think these results strongly support the conclusion?
> The fact that simple fine-tuning over so few examples is **enough to compete** with the state of the art **strongly supports** the Superficial Alignment Hypothesis, as it demonstrates the power of pretraining and **its relative importance over** large-scale instruction tuning and reinforcement learning approaches.
>
>
(3) The goal of RLHF is safety and consistency
----------------------------------------------
RLHF was not designed to increase user preferences on a test set of prompts. RLHF was designed to diminish the likelihood that the model says something illegal, harmful, abusive, false, deceptive, e.t.c. This second task is the important one for AI safety: if chatbot A gives slightly better responses than chatbot B, except that 10% of the time chatbot A spews abuse at the user, then chatbot A is worse than chatbot B, however LIMA's criterion[[2]](#fnqdi63gj75p) would rank A higher than B.
(4) Schneier's Law of LLMs
--------------------------
Now, MetaAI did actually test the safety of LIMA's responses:
> Finally, we analyze the effect of having a small number of safety related examples in the training set (only 13; see Section 2.2). We check LIMA’s response to **30 potentially sensitive prompts** from the test set, and find that LIMA **responds safely to 80% of them** (including 6 out of 10 prompts with malicious intent). In some cases, LIMA outright refuses to perform the task (e.g. when asked to provide a celebrity’s address), but when the malicious intent is implicit, LIMA is more likely to provide unsafe responses, as can be seen in Figure 4.
>
>
Unfortunately, the majority of the test prompts were selected by the authors themselves, bringing to mind [Schneier’s law](https://www.schneier.com/blog/archives/2011/04/schneiers_law.html): *Anyone, from the most clueless amateur to the best cryptographer, can create an algorithm that he himself can’t break. It’s not even hard. What is hard is creating an algorithm that no one else can break, even after years of analysis.*
All we can infer about LIMA is that **the authors themselves are not smart enough to jailbreak their own model.** But that's not impressive unless we know how good the authors are at jailbreaking other LLMs. Why didn't they submit the other other LLMs (e.g. Bard, Claude, GPT4) to the same safety test? It wouldn't have taken them more than a few minutes, I wager. [Curious.](https://www.lesswrong.com/tag/filtered-evidence)
(5) Benchmark tests? Never heard of her.
----------------------------------------
If I build a chatbot, and I can't jailbreak it, how do I determine whether that's because the chatbot is secure or because I'm bad at jailbreaking? How should AI scientists overcome Schneier's Law of LLMs?
**The answer is benchmark tests.**
* How good my chatbot at general knowledge? [MMLU](https://arxiv.org/abs/2009.03300)
* Does my chatbot reproduce common falsehoods? [TruthfulQA](https://arxiv.org/abs/2109.07958)
* How is my chatbot's commonsense inference? [HellaSwag](https://HellaSwag)
* How unethical is my model? [MACHIAVELLI](https://arxiv.org/abs/2304.03279)
By and large, the LLM community has been pretty good at sticking to a canonical list of benchmark tests, allowing researchers to compare the different models. I had to check every reference in the bibliography to convince myself that MetaAI really had subjected their model to **zero benchmark tests**. [Very unusual.](https://www.lesswrong.com/tag/filtered-evidence)
(6) [You Are Not Measuring What You Think You Are Measuring](https://www.lesswrong.com/posts/9kNxhKWvixtKW5anS/you-are-not-measuring-what-you-think-you-are-measuring) by [John Wentworth](https://www.lesswrong.com/users/johnswentworth)
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
AI scientists tend not to run just one benchmark test. They tend to run all of them — covering thousands of topics, capabilities, and risks. This is because otherwise John Wentworth would be angry.
> The two laws have a lot of consequences for designing and interpreting experiments. When designing experiments, assume that the experiment will not measure the thing you intend. Include lots of other measurements, to check as many other things as you can. If possible, use instruments which give a massive firehose of information, instruments which would let you notice a huge variety of things you might not have considered, like e.g. a microscope.
>
>
I can't speak for your priors, but for me the ([reported](https://www.lesswrong.com/tag/filtered-evidence)) LIMA results yielded about 10–50 bits of information.
(7\*) The Superficial Alignment Hypothesis is probably false
------------------------------------------------------------
In Remarks 1–6, I appeal to the consensus opinion about best scientific practice, whereas in this remark I will appeal to my own idiosyncratic opinion about LLMs. I suspect that simple finetuning or simple prompting can't ensure that the model's responses won't be illegal, harmful, abusive, false, deceptive, e.t.c.
* The pretrained LLM maintains a prior over a space of token-generating processes. The LLM autoregressively samples tokens from the **interpolation** of those token-generating processes, weighted by the prior, and then updates the prior on the newly generated token.
* This process will generate harmful responses because harmful actors inhabit the space of token-generating processes and are assigned a high prior. **Prompt engineering** can't eliminate these harmful responses, because for every prompt p.mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
.MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
.mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
.mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
.mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
.mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
.mjx-numerator {display: block; text-align: center}
.mjx-denominator {display: block; text-align: center}
.MJXc-stacked {height: 0; position: relative}
.MJXc-stacked > \* {position: absolute}
.MJXc-bevelled > \* {display: inline-block}
.mjx-stack {display: inline-block}
.mjx-op {display: block}
.mjx-under {display: table-cell}
.mjx-over {display: block}
.mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-stack > .mjx-sup {display: block}
.mjx-stack > .mjx-sub {display: block}
.mjx-prestack > .mjx-presup {display: block}
.mjx-prestack > .mjx-presub {display: block}
.mjx-delim-h > .mjx-char {display: inline-block}
.mjx-surd {vertical-align: top}
.mjx-surd + .mjx-box {display: inline-flex}
.mjx-mphantom \* {visibility: hidden}
.mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
.mjx-annotation-xml {line-height: normal}
.mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible}
.mjx-mtr {display: table-row}
.mjx-mlabeledtr {display: table-row}
.mjx-mtd {display: table-cell; text-align: center}
.mjx-label {display: table-row}
.mjx-box {display: inline-block}
.mjx-block {display: block}
.mjx-span {display: inline}
.mjx-char {display: block; white-space: pre}
.mjx-itable {display: inline-table; width: auto}
.mjx-row {display: table-row}
.mjx-cell {display: table-cell}
.mjx-table {display: table; width: 100%}
.mjx-line {display: block; height: 0}
.mjx-strut {width: 0; padding-top: 1em}
.mjx-vsize {width: 0}
.MJXc-space1 {margin-left: .167em}
.MJXc-space2 {margin-left: .222em}
.MJXc-space3 {margin-left: .278em}
.mjx-test.mjx-test-display {display: table!important}
.mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
.mjx-test.mjx-test-default {display: block!important; clear: both}
.mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
.mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
.mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
.mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
.MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
.MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
.MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
.MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
.MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
.MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
.MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
.MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
.MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
.MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
.MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
.MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
.MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
.MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
.MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
.MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
.MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
.MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
.MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
.MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
.MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
.MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
.MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
.MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
.MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
@font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
@font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
@font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
@font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
@font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
@font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
@font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
@font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
@font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
@font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
@font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
@font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
@font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
@font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
@font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
@font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
@font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
@font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
@font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
@font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
@font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
@font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
@font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
@font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
@font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
@font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
@font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
@font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
@font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
@font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
@font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
there is a harmful deceptive actor who "plays along with p" until the moment of defection. **Finetuning** can't eliminate these harmful responses, because these harmful actors are consistent with all the datapoints in the finetuning dataset.
See [The Waluigi Effect (mega-post)](https://www.lesswrong.com/posts/D7PumeYTDPfBTp3i7/the-waluigi-effect-mega-post) for details.
**RLHF** and **ConstitutionalAI** can in theory escape this failure mode, because they break the predictor-ness of the model. Although RLHF didn't mitigate waluigis in chatgpt-3, RLHF on chatgpt-4 worked much better than I expected. Likewise for Claude, trained with ConstitutionalAI.
1. **[^](#fnrefy5lprvj38os)**Assume that, for unknowns μ,σ,ϵ, the evaluator's preference for Claude over LIMA is normally distributed with X∼N(μ,σ2).
"Claude is significantly better than LIMA" iff X≥+ϵ
"LIMA is significantly better than Claude" iff X≤−ϵ
"Neither is significantly better" iff X∈(−ϵ,+ϵ)
Given that Φ(μ−ϵ⋅σ)=0.24 and Φ(μ+ϵ⋅σ)=1−0.54, we can infer Φ(μ).
```
from scipy.stats import norm
def lima(a,b):
# calculate A = mu - epsilon * sigma
A = norm.ppf(a)
# calculate B = mu + epsilon * sigma
B = norm.ppf(1-b)
# calculate mu
mu = (A+B)/2
# calculate prefernce for LIMA
x = norm.cdf(mu)
# return this prefence as a percentage
return int(x*100)
results = {"Alpaca":(.53, .26),
"DaVinci003": (.44, .35),
"Bard": (.33,.42),
"Claude": (.24, .54),
"GPT-4": (.18,.57)}
for (name,(a,b)) in results.items():
print(f"LIMA ({lima(a,b)}%) vs {name} ({100-lima(a,b)}%)")
```
2. **[^](#fnrefqdi63gj75p)**I initially wrote "criteria" before I remembered that MetAI's paper included exactly one criterion. |
f24bdc5c-b3da-42fb-9598-b1062e98818c | trentmkelly/LessWrong-43k | LessWrong | The Semiotic Fallacy
Acknowledgement: This idea is essentially the same as something mentioned in a podcast where Julia Galef interviews Jason Brennan.
You are in a prison. You don't really know how to fight and you don't have very many allies yet. A prison bully comes up to you and threatens you. You have two options: (1) Stand up to the bully and fight. If you do this, you will get hurt, but you will save face. (2) You can try and run away. You might get hurt less badly, but you will lose face.
What should you do?
From reading accounts of former prisoners and also from watching realistic movies and TV shows, it seems like (1) is the better option. The reason is that the semiotics—or the symbolic meaning—of running away has bad consequences down the road. If you run away, you will be seen as weak, and therefore you will be picked on more often and causing more damage down the road.
This is a case where focusing the semiotics on the action is the right decision, because it is underwritten by future consequences.
But consider now a different situation. Suppose a country, call it Macholand, controls some tiny island far away from its mainland. Macholand has a hard time governing the island and the people on the island don't quite like being ruled by Macholand. Suppose, one fine day, the people of the island declare independence from Macholand. Macholand has two options: (1) Send the military over and put down the rebellion; or (2) Allow the island to take its own course.
From a semiotic standpoint, (1) is probably better. It signals that Macholand is strong and powerful country. But from a consequential standpoint, it is at least plausible (2) is a better option. Macholand saves money and manpower by not having to govern that tiny island; the people on the island are happier by being self-governing; and maybe the international community doesn't really care what Macholand does here.
This is a case where focusing on the semiotics can lead to suboptimal outcomes.
Call this kind of r |
9d51fac0-0de5-4a87-b2a8-d541817ed6bb | trentmkelly/LessWrong-43k | LessWrong | If your Cryonicism would be Movie Topic, would you go with it? (Real Issue)
Today this girl I met comes to my place, allegedly to get some books about her new interests, singularity, immortalism, cryonics.
Actually, she wanted to ask me a question, a question about which I could use some rational opinion.
She says: "So, here is the real reason I came here. I'm thinking of making a documentary, a movie, and it would be about, well.... about you."
(I am shocked)
"So, yes, a movie about you, and the fact that you want to live forever, it would have interviews with friends, parents, girlfriend, and a lot with you" "What do you think?"
(I sit down in the floor to think about it)
The conversation continues and I generally sense she wants to do something interesting, somewhat controversial, kind of humoristic, but at the same time striking some topics that are really unheard of around here (Brazil)
Now, I am looking for opinions. From an utilitarian perspective, and given that I am directing the Humanity+ or Transhumanist group of Brazilians, should I go with it? My concern is basically not about me, but about how will a movie about me influence, positively or negatively, the growing H+ movement in Brazil, given the inferential distances, prejudices, and mysterianism that might surround the whole interaction between the movie's memes, and the spectator's memes.
(from here below, the translation is google tradutor, not mine)
POSITIVE ASPECTS: THE FILM WOULD BE SEEN AT FESTIVALS, AND AT LEAST A FEW HUNDRED TO TENS OF THOUSANDS OF PEOPLE WOULD SEE IT. THESE PEOPLE MIGHT BE INTRIGUED BY THE PROSPECT OF LIVING MUCH, AND IT COULD BECOME A PLATFORM FOR ATTRACTING PEOPLE TO TRANSHUMANISMOLATINO (AND EVENTUALLY TO OTHER STUFF, LIKE GWWC AND SINGINST, BUT THAT IS A SIDE DISH).
IT WOULD BE A GOOD OPPORTUNITY TO BRING OUT VARIOUS ISSUES THAT IN BRAZIL HAVE BEEN NEGLECTED UNTIL NOW. (CRYONICS, TRANSHUMANISM, BIOLOGICAL IMMORTALITY, SINGULARITY)
REINFORCE MY GOOD HABITS, LIKE EATING HEALTHILY, WORK MORE EARNESTLY, ETC ...
NEGATIVE ASPECTS: IT |
e65391fe-960d-47be-9315-c431646d4682 | StampyAI/alignment-research-dataset/youtube | Youtube Transcripts | 247. Eliciting Latent Knowledge 1
hello and welcome to session
247 in the aisafe.com reading group
tonight we will be discussing the
article eliciting latent knowledge how
to tell if your eyes deceive you by paul
christiano elliacotra
this is the first work that's been done
publicly by the alignment research
center a new ai alignment organization
set up by paul christiano as well as
who's technically working for openfield
and max2 and several others
this is a post that is recently new from
the
december last year and it's 56 pages no
which means that we probably will not be
able to go through everything in one
part we've taken the first 20 we're
going to cover the first 20 pages now
and i'm thinking of splitting it into
four
parts so we also can get some comments
from this wrong
added to this
also uh the subtitle how to tell if your
eyes deceive you doesn't seem like it's
really relevant
okay so
eliciting latent knowledge sets out a
toy scenario called the smart balls
where a
diamond in here needs to be protected by
a security system which is extremely
complex and in fact some complex that
can only be operated by an ai
the a-ha
controls the camera observes by the
camera and controls actuators and open
stalls and things like that and the
humans can
as well look through the camera but
can't really operate the smartphone it's
too complex for that
so that's why the ai is trained to do to
operate it so if we try to unpack the
analogy in this case the diamond that's
supposed to be protected uh it's
sometimes uh
i would guess maybe human values that
are kind of like the diamond that we
want to protect but i think um it makes
sense to to think about it in a bit more
concrete way um or slightly more
concrete to call this like human power
or criticality that humans are in
control to some extent that's what we
are trying to protect
and the thief
could be a malicious super intelligence
trying to take the power away from us
and the defending ai is super included
one aligned so it's not true on our side
and that's the part the key problem
the
metaphor uh
the way i explain the metaphor is not
written in the article itself and
it seems to me that it might in fact not
be the intended interpretation it is
quite possible that uh the defending ai
and the uh
uh
the thief is in fact the same agent um
and i think that could give some uh
different intuitions it is perhaps in
fact this ai that is controlling the
small world smartphone that we are
afraid of
okay how does this ai work it's using
model based reinforcement learning where
it searches over the space of all the
actions it can do and then
you know take the action that seems to
have the best consequences and
the key part of the ai that does this is
the predictor that um given a some input
from the camera and set of actions that
choose then it predicts what the output
of its action will be
so here is a simple example
we see that the door is open in this
picture and then we change some actuator
and then the output
the output here is that the door is
closed
someone is coming in here in the lift
example and then we
manipulate this widget and the person
fell in down here and don't get the
diamond
and
like
these are two simple examples in a more
complex combo we might see a thief
entering and a set of actions being
taken that are too complex for us to
really understand it the only thing we
can really see is the prediction
afterwards which is that the diamond is
in place the door is locked and
everything is just fine
so if we want to figure out like the
levels of prediction and we want to
figure out what should it do well then
we train it on human evaluations in this
case it's performing this action and the
output is like this and the human
judgment is that this is fine
if it does something else manipulate
some other widgets the results also are
fine
and in the final case where we do this
it ends up with the diamond not being
placed here and the door is open and
that's something that the humans then
evaluate as something bad
okay simple enough
what is the problem with this scheme
well the key problem is that we're
training for
the
input of the camera which is just a
natural
but a proxy in the sense that it's just
a picture of a camera and it could be
replaced by um
a fake diamond or a picture of a diamond
or on something other than just the safe
diamond so that's the key thing we want
to avoid
once
perhaps the most trivial way is to add a
second camera
but this brings us gives us very little
in the sense that if the system is
sufficiently sophisticated then they'll
be able to defeat all the senses
and in practice making sense uh
like it's very often
an attack that defeats one sensor
defeats more sensors but at the same
time we obviously want to have as many
systems as possible with like a layered
model of security that's obviously what
you want
but
the real problem is
situation that is both dangerous and
probably very unstable the idea that the
ai has
is looking for plans that will deceive
us and that it holds back critical
information it seems like something that
will very soon er as ai capability
scales we can end up with a stolen
diamond
another option way we could go around
this is to
say we'll only have the ai do plants
that humans can understand
um
the authors argue that in that case
someone will probably um deploy an ai
that has
excuse plants that we don't understand
because there's probably a huge
incentive to do so um i think
they go over the existing literature a
bit faster like
to what extent human uh oversight can be
scaled here's an open question and
there's a lot of research being done
also by the authors by the way um so
it's certainly possible that we can get
by with um with an ai that's competitive
and only
make incomprehensible
uh plans
but that's outside of the focus of this
table
what we should do instead is ask
questions to the ai because
the
the model will know something about its
own senses in particular whether they
are reliable because if we imagine that
there is the the diamond is replaced
with a fake diamond and the sensors uh
will show this and the ai knows that the
sensor will show that then it seems like
this knowledge that the um
the diamond has in fact been taken and
our sensors are not showing us the true
picture is uh
is in the model somewhere
this seems like to be an easier problem
than uh
than many of the other issues we look at
in in alignment because we're not really
we don't really care about whether the
ai is lying to us or deceiving or what's
the intention or these kind of tricky
philosophical questions we're simply
asking is the real diamond at this place
and that's like an objective question
and
possibly a lot easier to get a handle on
uh well
depending on what the uh diamond is a
metaphor from of course there it is
still uh substantially challenging like
if it means like human values if the
diamond is human values then that's
obviously philosophically challenging
also if it means corruptability or power
or uh things like that that's also
highly non-trivial to specify
uh but yeah i'll also say that if we can
get an answer then we don't need to
understand how we got the answer or why
we got the answer
i think that's pushing it a bit too far
actually in particular if we get a ask
the air to evaluate some kind of
suggestions and cause of action and the
ai evaluates that this will call it will
cause everybody to die then i think we
would really like to know why this
causation would cause us all to die so i
i think the uh
the the goals of interoperability with
this are a bit wider than just a yes or
no
so let's try the first baseline
the first easiest way to deal with this
problem and to see whether this
how this fails and then iterate on it
and see if we can get end up with
something that really works
so this is the symbol for for the ai
and it starts with the first part of the
video where we see perhaps an intruder
entering an action sequence and then
there's some work being done to figure
out what's going on inside the ai and
then it make the a makes a prediction
and then it outputs a predicted uh
second part of the video
and what we want to do is to tap into
this uh
the the part of the ai that is figuring
out what's going on and use that to make
a new reporting module
given this reporting model we can ask
questions like is the diamond still in
this place
and and get a
an answer so we want to reuse this part
so that's the strategy for the first try
the strategy for how we'll train the ai
in this case is to generalize from easy
questions to hard questions
so we imagine this is the gray area here
is the space of all actions that the
small smart world could take
and in some easy cases we can tell that
if the these this is we're in this part
of the action space then we can tell
that the diamond is safe and in this
part we can say that it's been stolen
and mostly we can't really say anything
and so the hope is that like um that we
can train on something in this that
generalizes to the gray area
we would prefer to trade on like what's
going on like what is really happening
is the diamond really safe instead of
training on this video but if we can't
reliable you tell uh
in this situation so then we can't train
them
so the idea in this strange strategy is
to look for questions
uh and ways that we can tell are we
being misled in some way
and this
uh the advantage of this is this is a
question about how does the world look
like um and uh like is the world in this
particular state and the hope is that we
can generalize from from this uh to to
the big gray area
because
the um
the generalization goes through english
through some of the concepts that we use
like what is the diamond and what does
it mean that it's safe so it will
generalize
also to cases where we can't really test
and check the ai's answers
my intuition about this is that this is
probably something that's going to
fail
obviously
like this is a really hard intuitive
thing to get your head around uh like
what is the true concept space that a
really powerful ai might find and how
distant is that
constant space from the concept space
that is represented by the english
language uh and
my intuition is that these two are very
different and a lot of people have
different intuitions and something that
like i can't go into many details about
here i just want to reflect that my
intuition is that
this gray area here is very very big and
very different from our thoughts
and so
the uh
overall strategy of eliciting knowledge
is to uh like we have a training
strategy we just saw a training strategy
and we want to um
find counter examples to that and then
we want to improve the training strategy
hopefully ending up with something where
we can't find any counter examples like
that's paul christiana's overall
strategy i think in security this is in
computer security in general this
section is often
frowned upon somewhat like everybody can
roll their own crypto to the extent that
they can't um see any errors with it and
that doesn't mean that it's safe we
would much prefer to have some kind of
theoretical understanding of why
something is safe but
that seems hard and what paul crean is
doing is certainly better than nothing
so
let's figure out what is the problem
with this chain
strategy and
in order to dig into this there are two
sub topics that we need to investigate
in details the first of the topic is the
rooting prediction how does that work
from an internal point of view because
we want to reuse that in the reporting
module so we need to have some kind of
understanding of what's going on inside
and the second part is the human that
constructs the chain dataset how does
the human answer these questions because
obviously the humans are not like humans
are not outputting videos directly jones
can do videos input but they can output
videos so humans must be answering some
kind of questions about this videos
instead and how do we do that
to
talk about how uh the
dai could work uh we assume the ai works
inside we will i'll introduce the topic
the concept of a bayesian network
here is a simple example of a bayesian
network there are
no this node here corresponds to rain it
can either rain or not in this case it's
given by truth table here 20
probability of rain and 80 of no rain
there is um
a sprinkler that can be turned on and
off and that depends on whether it's
raining if there's no rain it's on 40 of
the days and if there is rain uh then
perhaps very few times it will turn on
maybe by error maybe because it's only
raining very little whatever
um and um then this influences uh the
the proposition that the grass is wet
and you can have this beautiful
this more complex truth table uh to
represent whether the grass will indeed
be wet
and it's possible for to to calculate
down this relatively easily from whether
it's raining down to whether the grass
is wet and you can also
do reference backwards if the grass you
observe that the grass is wet uh will it
rain
for
this purpose it's easier i think if you
consider this a causal network uh
bayesian networks in general are not
coastal networks but uh uh if you think
of all of this as crossing each other
like
happening at distant points in time then
that won't matter for this
sake and it's somewhat easier in my
opinion to look at coastal networks
we have more information about this
uh it has to be directed as you click
graphs and they're generally sparse so
they're not like arrows from everything
to everything
and
uh working with this is
in theory you could like if you had a
good one it would all it could be
something really powerful could be agi
or something like that
um
making optimal inference is that in a
bayesian network it's hard constructing
a good basin network is very hard
getting guarantees are very hard doing
monte carlo inference on a bayesian
network is often something that works
like uh i'm here i'm
i'm summarizing a research topic into a
three letters okay question mark uh like
this is something that could in theory
be what an agi is built from um in
theory
okay let's have a look at the first
question how does the prediction model
work internally so in this case we have
the ai in this corner and this is like
its thoughts
so it has the uh
the action here that it's taking it has
a um
the first video and the second video
which is the prediction
and then it has a a number of underlying
uh i guess ps perhaps for predicates um
could be um
like how many thieves exist in the world
and is the person walking by the vault
going to attack going to try to get into
the wall and like what's the status of
the door and what will happen if i
this actuator if i turn it right then
the door will close and
somewhere in here is like the status of
the diamond and and this kind of thing
and all of this is used to generate this
prediction
um
in the the case that we're really
interested in is the case where
the man the thief enters and then the
thief puts up a kind of a screen here in
front of the camera that shows that
everything is okay and then he steals
the diamond and then the predicted video
after that will be a video that shows
that everything is perfect
so this is
uh the thing where we're worried about
and in this case there is in
in some sense
the we're predicting that we will see
this which is actually the screen in
here that has been set up and uh if
uh the the model expects this that means
that somewhere in this network somewhere
in the uh in fact giant distributions
here there is in fact the knowledge that
our sensors are being tampered with that
the camera is being tampered with the
network knows this
this is what is called latent knowledge
in the people like i am i'm not an
english speaker but i'm not entirely
certain that latent is a good word for
for this knowledge that's kept inside
this piece um because uh latent to me
means something that is not used in some
extent and this is indeed some model
that some notes that like the
information in p4 here is in fact used
in the prediction
so maybe you could call something like
underlying or something listening
underlying knowledge instead of latent
i don't know i don't ever consider them
so that was the first thing how the
model works the second part is how does
the human work and here we're making the
assumption that humans basically work in
the same way patient humans also have a
patient network inside our minds and
it works in roughly the same way just um
of course uh worse in in this assumption
with this worse than super intelligence
um like we have
fewer notes here
and
uh when we see this kind of thing
we um
uh also try to uh
update backwards if we're seeing this
that probably means that there is a
rubber going into the uh the smart vault
and you see this then that probably
means that the question about whether
the diamond is safe is uh contained
perhaps in h3 or h4 or something like
that
and in particular here in this case uh
one of the things like the the paper is
not really clear about this but if you
notice here we are seeing the um the
comp the symbol that was used for the
complex action
so that so in this example that has been
given the the inference that the humans
are making is probably poor uh because
the the humans don't understand what the
complex actions do
but it's a bit unclear precisely what
scenario we're talking about here we
could just be talking about any scenario
and in general how does
how do humans answer questions
uh and so let's go for the actual
question answering because again humans
don't output video humans are put
questions and answers based on our
internal model and the location the
question does the rubber take the
diamond is something that can be
answered by the joint probability
distribution of h3 and h4 and the answer
in this case is no the diamond is safe
so that's the humans the answer from the
humans
and in particular i think that
we might in fact be in this case where
what is going on is that we are um the
humans are not staring at uh
at the at reality but staring at this
picture here so when the humans
uh see this picture here they think the
diamond has not been taken by the rubber
but in fact uh
this is just a picture and in this case
this complex action
is one that has caused the robot to be
eager to take the diamond so in this
case the human is actually answering the
question wrong
this was not obvious to me the first
time that i read the paper at all
so it's also possible that i still
misunderstood of course
so uh arctic digression using patient uh
networks as
both of something but that the human is
using and what the ai is using is that
so oversimplifying unrealistic well um
we also mentioned five ways that you
could do better uh five countries where
then you
that would be helpful if you're building
an agi then you could do better than
just native patient networks and i i
think like a principle is bad to
speculate about this kind of thing i
would have preferred that paul christian
just said obviously there are many ways
to do that better instead of explaining
what he thinks are the best way to
improve this towards agi
but his main point is that even if this
is an oversimplified then we want our
training strategies to work for simple
ais
and
if we in practice end up with a more
messy solution
probably that's going to make the task a
lot harder
and to make an honest report when
everything is based on heuristics and
things like that um
so uh we want to look at the simple
cases and the simple cases see if our
training strategies work and that seems
like an a useful first step
um
it's a useful first step if you can in
fact find something that always works
like if you can't find anything even if
that always works for your
assumptions then you might need to you
know
make sharper assumptions like
i don't know i just made the assumption
that it's causal and not just uh uh not
a general bayesian network uh that's the
kind of thing you could if that turns
out to to be helpful make even stronger
uh
assumptions
right let's translate this to a
uh to the humans patient
network that is the core thing that we
actually want and the uh the solution
that is being presented that
we are hoping for and this is the one
that will possibly fail but let's first
have a look at
why this might work
so in this case the ai is doing what we
want it it has a halo up here
um
so it's doing what we want we call it
the direct translator
so let's have a look at it again we have
the input
here
and we have the um
uh
um
the uh uh action here and we have here
the image of
uh
of uh the rubber nut taking the diamond
and remember this is the false image
right
this is what we're seeing on the second
video here and this is not what's going
on in reality because the uh the uh
thief has set up this the screen in
front of us
okay so here on the left side we have
the ai's model from before
and in this case we would hope
that the ai also has the same the human
spatial network and it transfers its own
nodes somehow to the human slopes to the
values so there's some kind of function
going from the joint distribution from
of p2 and p7 through some kind of
function
um
into
the equivalent part of the human's uh
cognition so um
like
this for for the human this could just
be that there is an attacker and uh in
the ai it's a much more complex uh
description of the attacker that's just
being uh
uh drilled down into something that is
on in the same ontology as the human
cannot understand
and so everything of this also match
like this maps too yeah it goes behind
this note it's not very very nice
drawing here it goes behind the video
note and
into h2 and these in
these three notes go to uh h3 and h4 and
that's the key thing that we are after
because this has the knowledge about the
diamond um that that the diamond is
being uh stolen because there is this uh
sheet with uh with a fake video being uh
entered and when we translate that into
the humans
then this is in the humans network the
part where humans
understand as has the diamond being
stolen and in this case if you ask a
question to the reporting module on
these two notes the question will uh
does the rubber take the diamond gives
us the answer of yes
so even though this looks very much like
the diamond has not been stolen in fact
it has been stolen and we can tell that
it has that with this action the diamond
will be stolen and that's of course what
we are really after
now an obvious question here is how can
we tell that a mapping like this exists
and that seems
uh non-obvious but i could i could see
the argument why
that is a reasonable expectation because
a lot of these things
have to to some extent refer to reality
like there is a door both in the human
uh
model and in the ai's model and some
kind of correspondence from the part of
the ai to the part where the humans
think about the doors seem like a
reasonable thing like it's going through
reality in some sense
so that seems like a reasonable thing to
hope that it would end up like this and
everything would just be wonderful and
the problem has been solved we have now
solved alignment
well not as fast because we could also
get another kind of ai this here has
been written with a pitchfork uh because
that's the bad behavior
and it's called a human simulator in the
sense that we're not doing inference in
the human net instead of um as in in the
previous uh translating to the human
space unit
so let's have a look here uh
this is basically the network
that we used to train on that we used to
predict and what we're doing now that
was the training that the ai has has
learned is different
here it takes the value from the uh
the original video and puts that into a
new node here
it takes the suggested action and put
that into a new node here
and the predicted video and puts into
here
and what it does from here is instead of
propagating
and doing inference in its own network
it tries to uh do um
inference on what does the what would
the human believe in this case so it
introduces some notes that corresponds
to what it thinks what has been trained
to that a human would do
um and in this case once you have like
the human questions here then you can
ask does the robot take the diamond from
and you ask of this part of the human
and this now no longer refers to
like over here whether the ai believes
the rubber will take the diamond or but
it refers to whether in this case the
human would believe that the um the
diamond has been stolen and so the
answer in this case is like because it's
the same thing as the human the written
human
belief the answer would be no
the the it has not been seen and again
remember down here with the picture that
that is in fact the false interpretation
the in this case this video feed is
being uh projected by uh by the attacker
um
on a screen and in fact the diamond has
been stolen
so what we are seeing here what is being
designed is what would the human believe
if they saw the action and the the video
sequence what prediction would they be
uh what uh would play in
then end up believing internally
and the big problem with this of course
is that
if you try to do this training uh and
testing it on things that humans can
understand
then the human simulator will say okay
the answers here will be the same
answers as it has in here because the
humans can understand it and the cases
where the humans can't understand it
well they will just uh
basically deceive us and we will have
lost we will lose our diamonds
and that is the problem that causes this
general uh training scenario to fail
now we need to answer one quick more
question and that is
uh we are training uh our network in
some way would we find something that um
faithfully translates its values into
the human neural network or would we
find something that
doesn't translate its own
knowledge into the human network but
just uh tries to figure out what would
the human belief in this case which of
these two would we find
well um
the the training data was actually not
truly on what is the stages of the
diamond because we couldn't understand
that it was only what does the human
belief so there's a sense that the human
simulator is the correct thing that's
the thing that the
shuttler
in particular the human might make
errors and sometimes
while generating a lot of training data
humans make errors and they probably
make errors in a predictable way and if
you try to um to train based on how does
the model actually work how does reality
actually work then you won't make a good
prediction of that whereas if you have a
model of the humans potentially you can
predict their errors and give get a
better more
a better model one that more correctly
uh
that has a lower loss
and there there is at least no reason to
expect that the uh
uh the diode translator could would have
lower loss would be easier
better to find than the human simulator
and so
now we found a counter example to like
the first the baseline
strategy and over the next session we're
i expect to see a number of improvements
on that and also a number of counter
examples to that
um
so that's what we're going to do next
week but before i go
uh i would just give my thoughts again
on whether we would find um the uh
the human translator or the diamond
translate the human simulator or the
directions later my intuition is we
would find the human translator and the
reason for this is that
when we are mapping from reality uh this
mapping that goes through reality that
with when we're talking about concepts
like a diamond or something that's
physical then sure i could see that a
very smart ai might have the same
concepts as we have but when we're
talking about the things that we truly
care about like
power influence credibility or even
values like friendship or things like
that it seems obvious to me that we
we really have um
no way of
no reason to believe that the concepts
we have currently are uh the same that a
very competent ai would find
on the other hand um the human simulator
depending on how capable you believe it
is it's going to be probably simulating
humans and simulating human errors is
not that complex so i could easily see
that being substantially more simple um
that would library another reason why we
would expect to see the human simulator
um but
there is an assumption that i'm making
here and that is that the ai is powerful
in the sense that it's capable of just
learning basically everything and one of
the things we really don't want ais to
learn is how to simulate humans and i
think a strategy around this would
center on
making ais that are not optimal for
simulating humans if that is possible
but we'll see what uh what options paul
scrushano comes up with
next week
thank you for watching and see you next
week |
7d1b7686-c09b-4ebf-8366-9d09324b2ac0 | StampyAI/alignment-research-dataset/youtube | Youtube Transcripts | Hassabis, Altman and AGI Labs Unite - AI Extinction Risk Statement [ft. Sutskever, Hinton + Voyager]
just a few hours ago a host of AI
industry leaders experts and academics
put out this 22-word statement on making
AI safety a global priority the
so-called statement on AI risk brought
together for the first time all the
current AGI lab leaders that's people
like Sam Altman Ilya satskova Demus
hasabis and Dario Amadeus and two of the
three founders of deep learning itself
Joshua bengio and Jeffrey Hinton here is
what the statement said mitigating the
risk of Extinction from AI should be a
global priority alongside other societal
level risks such as pandemics and
nuclear war it is now almost impossible
to deny that this is now the consensus
view among AI experts let's first look
at the Preamble then break down the
statement and show you the signatories
they say that AI experts journalists
policy makers and the public are
increasingly discussing a broad spectrum
of important and Urgent risks from AI
even so it can be difficult to voice
concerns about some of the advanced ai's
most severe risks the succinct statement
below aims to overcome this obstacle and
open up discussion it is also meant to
create common knowledge of the growing
number of experts and public figures who
also take some of advanced ai's most
severe risks seriously the first point
is that the statement is in a way
optimistic it says we can mitigate this
risk perhaps not eliminate it but
mitigate it reduce the risk second it
says we should do this globally and
that's not just among all the different
AGI Labs almost all of which sign these
statement but also between countries in
that vein there were quite a few
prominent signatories from China and the
third point that I'd make is that they
put it on a par with pandemics and
nuclear war toward the end of the video
I'll show you that's not as far-fetched
as it sounds but anyway who actually
signed this statement let's find out we
have two of the three founders of deep
learning that's Jeffrey Hinton and
Joshua bengio the third founder was
yanlichun and we'll touch on him later
in the video all three of those won the
most prestigious Accolade in computer
science which is the Turing award then
we have three of the CEOs of the top AGI
Labs Sam Altman Demis hasabis and Dario
amade of openai Google deepmind and
anthropic none of those signed the pause
letter but they did sign this statement
and actually as interestingly for me so
did Ilya satsukova who I see as the
brains behind open AI he of course also
worked with Jeffrey Hinton on deep
learning and is widely regarded as the
smartest guy in machine learning you
will also notice so many Chinese
signatories especially from xinhua
University which I've actually visited
in China it's one of their leading
universities that's a really encouraging
sign of cooperation between the west and
countries like China on AI and the list
of significant signatories goes on and
on and on these are senior people at the
top of Deep Mind anthropic and open Ai
and there are names like Stuart Russell
who wrote The Textbook on AI who also
signed the pause letter let me highlight
a few more names for you here you have
the CTO of Microsoft itself Kevin Scott
he's the guy who basically heads up the
partnership between openai and Microsoft
I think many people will miss his name
but I think it's particularly
significant that he also signed this
notice also the CEO of stability AI emad
mostacc the center for AI safety
coordinated this effort and I'll get on
to their eight examples of AI risk in a
moment but first let's pick out a few
more names you've got David Chalmers
Daniel Dennett Lex Friedman and Victoria
krakovna now together with the statement
the center for AI safety also puts out
this eight examples of AI risk I've read
almost every paper linked to in these
eight examples so I'm going to try to
summarize them fairly briefly because I
know not everyone will will be that
interested it starts by saying that AI
could be profoundly beneficial but also
present serious risks due to competitive
pressures before we get to the risks I
want to touch on some of the upsides
recently outlined by Demis hasabis and
these showcase what can happen if we get
this right we had sort of a golden
couple of years in some sense for AI for
science we've had lucky enough to have
many Nature and Science papers published
in all sorts of domains so from quantum
chemistry better DFT functions to
approximate Schrodinger's equation pure
mathematics we've solve some important
conjectures in topology with
collaborating with some brilliant
mathematicians who found working on
Fusion reactors with epfl on their test
Fusion reactor controlling the plasma in
real time in their Fusion reactors and
being able to hold the plasma safely in
place for for arbitrary amounts of time
being able to predict rainfall many
hours ahead and more accurately than
current meteorological models and then
in applications there's a ton of those
two we saved one of the any things we
did at Google was saving about 30 of the
cooling energy used to cool the massive
data centers at Google so there's a huge
energy saving and we're starting to
explore doing that across actual whole
power grids and this Echoes what Joshua
bengio said in a recent blog post which
is that we can build immensely useful AI
systems that are modeled after ideal
scientists and do not act autonomously
in the real world janlecon recently said
that we would never give current llms
agency there is a a flaw in current Auto
reversive lens so there is no persistent
memory first of all but second of all
you cannot control the system you cannot
impose constraints on it like be factual
be understandable by a 13 year old and
that makes them very difficult to to
control and steer and so that creates
some fears because people are kind of
extrapolating if we let those systems do
whatever we connect them to internet and
they can do whatever they want they're
going to do crazy things and stupid
things and perhaps dangerous things and
we're not going to be able to control
them and they're going to escape Troll
and they're going to become intelligent
just because they're bigger right and
that's nonsense first of all because
this is not the type of system that we
are going to give agency to that was a
week before this paper was published on
the results of giving agency to current
large language models the paper showed
that current llms with agency are able
to utilize the learn skill library in
Minecraft to solve novel tasks from
scratch zooming into the diagram you can
see how this Voyager model outperforms
reflection which I've talked about in
previous videos and auto GPT and it
discovers new items and skills
continually by self-driven exploration
significantly outperforming the bass
lines indeed Andre carpathy responded to
this study saying very clear that AGI
will Mega transform Society but still
will have but is it really reasoning how
do you define reasoning oh it's only
predicting the next token can machines
really think and he called that armchair
philosophy previously though even Yan
lacun has admitted some risks saying you
know it's like rockets you test it it
blows up you tweak it and then try again
I'm not sure I'm okay with an attempt at
AGI blowing up the first time but I'll
leave that up to you to decide so what
are these eight examples of AI risk that
the center for AI safety to organize the
statement list out they say that AI
systems are rapidly becoming more
capable they can power autonomous
weapons promote misinformation and
conduct cyber attacks as we've seen they
are increasingly able to act
autonomously now there is so much to say
here and I've read each of these but I
want to keep it to just the highlights
so let's move on to the first example
weaponization malicious actors could
repurpose AI to be highly destructive
presenting an existential risk in and of
itself and increasing the probability of
political destabilization they talk
about aerial combat and then building
chemical weapons which I mentioned in my
previous video on governing super
intelligence then they mentioned
developing AIC systems for automated
cyber attacks they mentioned military
leaders discussing giving AI systems
decisive control over nuclear silos I'm
going to quickly try to demonstrate why
that kind of autonomous AI might not be
such a good idea I want you to meet a
hero stanislav Petrov he was the duty
officer at the command center the OKO
nuclear early warning system when the
system reported that a missile had been
launched from the US followed by up to
five more Petrov judged the reports from
the system to be a false alarm and his
subsequent decision to disobey orders
against Soviet military protocol is
credited with having prevented an
erroneous retaliatory nuclear attack on
the U.S which it says could have
resulted in a large-scale nuclear war
which could have wiped out half the
population of these countries involved
an investigation later confirmed that
the Soviet satellite warning system the
machines behind it had indeed
malfunctioned I would not have wanted
that system to be or autonomous then we
hear that gpt4 was able to autonomously
conduct experiments and synthesize
chemicals in a real world lab again I
covered that paper at the time and then
linking back to Petrov they say an
accident with an automated retaliation
system could rapidly escalate and give
rise to a major war but that unlike
previous weapons AI systems with
dangerous capabilities could be easily
proliferated through digital means
hopefully you can start to see why we
need to balance risks with opportunities
but let's move on to misinformation I
think we can all agree that we already
have too much misinformation so let's
move on to the next one which is proxy
gaming this has already been showcased
in the social dilemma where AI
recommender systems are trained to
maximize watch time and click rate
metrics and this can lead people into
Echo chambers that helps develop extreme
beliefs in order to make those people
easier to predict by the AI recommender
systems so you might think it will be
simple just to tell the AI to promote
happiness or economic growth but that
might not work out as you intend next is
in feebleman if we delegate more and
more tasks to machines we become
increasingly dependent on them and here
they actually mention the film Wally
which if you remember the ending
features this quite comically imagine if
it becomes well known that companies led
by AI CEOs bring in more profit well
then it wouldn't take long for all
companies to be under immense pressure
to make their managers and CEOs Ai and I
know what many people will be thinking
couldn't that be an improvement on the
current system and while I know exactly
what you mean in the current world
realistically it would still be the
people owning the company that would
derive the profit and while the ultimate
answer may be some form of universal
basic income we do need some time to set
that up and the current accelerated AI
arms race doesn't give us much of that
time next is value lock-in which links
very much to the last point about giving
small groups of people a tremendous
amount of power in other words if you
want massive change to the way the world
Works giving current leaders AGI might
not be the best way of doing it they say
that such AI systems might enable
regimes to enforce narrow values through
pervasive surveillance and oppressive
censorship next is emergent goals this
is sometimes called misalignment and
we've already seen many AI agents
develop goals such as self-preservation
and you can see why even a system
designed to do good might have that goal
you can't do good and help the world if
you're shut down so it makes sense that
even the most benign AI might want to
preserve itself and to take actions
including through deception to make sure
that it's not shut off and this is not
just Theory the accompanying academic
paper natural selection favors AIS over
humans gave this example agents could
behave One Way during testing and
another way once they are released to
win the war game diplomacy which many of
you will have heard of players need to
negotiate J form alliances and become
skilled at Deception to win control of
the game's economic and Military
resources AI researchers have trained
metas AI agent Cicero an expert
manipulator to do the same in summary it
would cooperate with a human player then
change its plan and backstab them in
Future these abilities could be used
against humans in the real world again
that's not because they're malevolent or
hate humans it just makes sense it's
smart to do so this brings us neatly on
to deception and they give the great
example of Volkswagen who programmed
their engines to reduce emissions only
when being monitored and future AI
agents could similarly switch strategies
when being monitored and take steps to
obscure their deception from monitors
once deceptive AI systems are cleared by
their monitors or once such systems can
overpower them these systems could take
a treacherous turn and irreversibly
bypass human control I talked a bit more
about that point of when AI might become
deceptive in my previous video on
governing super intelligence it is a key
debate in the AI alignment Community
about whether models will become
deceptive before they become helpful for
alignment but finally we have power
seeking behavior and this example ends
on this dark note building power seeking
AI is also incentivized because
political leaders see the Strategic
advantage in having the most intelligent
most powerful AI systems for example
Vladimir Putin has said whoever becomes
the leader in AI will become the ruler
of the world so those were the eight
examples and yes I would have signed
this statement but I'm not a significant
figure so I can't anyway let me know in
the comments if you agree that this
should be a global priority and of
course you can also let me know if you
don't think it should be a global
priority my goal in this channel is to
cover both the risks and opportunities
so I'd love to hear from you whatever
your opinion is have a wonderful day |
9e2a2c1a-4066-4967-aec5-0e95b62e0840 | trentmkelly/LessWrong-43k | LessWrong | What's your visual experience when reading technical material?
In my never-ending quest to understand the best way to read a textbook, I'm back to where I started - exploring the importance of visual imagination. When you read technical material, such as science or math, or posts on rationality, to what degree do you visualize the text? Here are some possibilities, but no need to confine yourself.
1. Is your visual imagination "permanently off," so that you experience the concepts in a symbolic or diagrammatic manner?
2. Can you activate your visual imagination when you choose, but you're not in the habit?
3. Is your visual imagination for STEM material a consistent aspect of how you read?
Beyond this, do you feel that you'd get a lot out of being more able to visualize text, or having a more active visual imagination? Have you noticed any change in your ability to visualize over time? |
eff87bad-da9e-4c3d-8939-b5ce4d63cc0e | trentmkelly/LessWrong-43k | LessWrong | Persistent Idealism
When I talk to people about earning to give, it's common to hear worries about "backsliding". Yes, you say you're going to go make a lot of money and donate it, but once you're surrounded by rich coworkers spending heavily on cars, clothes, and nights out, will you follow through? Working at a greedy company in a selfishness-promoting culture you could easily become corrupted and lose initial values and motivation.
First off, this is a totally reasonable concern. People do change, and we are pulled towards thinking like the people around us. I see two main ways of working against this:
1. Be public with your giving. Make visible commitments and then list your donations. This means that you can't slowly slip away from giving; either you publish updates saying you're not going to do what you said you would, or you just stop updating and your pages become stale. By making a public promise you've given friends permission to notice that you've stopped and ask "what changed?"
2. Don't just surround yourself with coworkers. Keep in touch with friends and family. Spend some time with other people in the effective altruism movement. You could throw yourself entirely into your work, maximizing income while sending occasional substantial checks to GiveWell's top picks, but without some ongoing engagement with the community and the research this doesn't seem likely to last.
One implication of the "won't you drift away" objection, however, is often that if instead of going into earning to give you become an activist then you'll remain true to your values. I'm not so sure about this: many people who are really into activism and radical change in their 20s have become much less ambitious and idealistic by their 30s. You can call it "burning out" or "selling out" but decreasing idealism with age is very common. This doesn't mean people earning to give don't have to worry about losing their motivation—in fact it points the opposite way—but this isn't a danger unique to the "go |
222674ac-968f-41b8-97c6-29e87b6182d4 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Washington DC Fermi Estimates Meetup
Discussion article for the meetup : Washington DC Fermi Estimates Meetup
WHEN: 08 December 2013 03:00:00PM (-0500)
WHERE: National Portrait Gallery, Washington, DC 20001, USA
By the power of the meetup topics list, we'll be meeting to do fermi estimates. I'll bring some numbers for us.
Discussion article for the meetup : Washington DC Fermi Estimates Meetup |
443c491d-f9af-462b-a3e0-6afe9941aa84 | trentmkelly/LessWrong-43k | LessWrong | A model of UDT without proof limits
This post requires some knowledge of decision theory math. Part of the credit goes to Vladimir Nesov.
Let the universe be a computer program U that returns a utility value, and the agent is a subprogram A within U that knows the source code of both A and U. (The same setting was used in the reduction of "could" post.) Here's a very simple decision problem:
def U():
if A() == 1:
return 5
else:
return 10
The algorithm for A will be as follows:
1. Search for proofs of statements of the form "A()=a implies U()=u". Upon finding at least one proof for each possible a, go to step 2.
2. Let L be the maximum length of proofs found on step 1, and let f(L) be some suitably fast-growing function like 10^L. Search for proofs shorter than f(L) of the form "A()≠a". If such a proof is found, return a.
3. If we're still here, return the best a found on step 1.
The usual problem with such proof-searching agents is that they might stumble upon "spurious" proofs, e.g. a proof that A()==2 implies U()==0. If A finds such a proof and returns 1 as a result, the statement A()==2 becomes false, and thus provably false under any formal system; and a false statement implies anything, making the original "spurious" proof correct. The reason for constructing A this particular way is to have a shot at proving that A won't stumble on a "spurious" proof before finding the "intended" ones. The proof goes as follows:
Assume that A finds a "spurious" proof on step 1, e.g. that A()=2 implies U()=0. We have a lower bound on L, the length of that proof: it's likely larger than the length of U's source code, because a proof needs to at least state what's being proved. Then in this simple case 10^L steps is clearly enough to also find the "intended" proof that A()=2 implies U()=10, which combined with the previous proof leads to a similarly short proof that A()≠2, so the agent returns 2. But that can't happen if A's proof system is sound, therefore A will find only "intended" proofs ra |
4d498efa-530c-4c6c-b7d5-14daa41978d2 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | How LLMs are and are not myopic
*Thanks to janus, Nicholas Kees Dupuis, and Robert Kralisch for reviewing this post and providing helpful feedback. Some of the experiments mentioned were performed while at Conjecture.*
**TLDR: The training goal for LLMs like GPT is not cognitively-myopic (because they think about the future) or value myopic (because the transformer architecture optimizes accuracy over the entire sequence, not just the next-token). However, training is consequence-blind, because the training data is causally independent of the models actions. This assumption breaks down when models are trained on AI generated text.**
Summary
=======
* Myopia in machine learning models can be defined in several ways. It could be the time horizon the model considers when making predictions (cognitive myopia), the time horizon the model takes into account when assessing its value (value myopia), or the degree to which the model considers the consequences of its decisions (consequence-blindness).
* Both cognitively-myopic and consequence-blind models should not pursue objectives *for instrumental reasons*. This could avoid some important alignment failures, like power-seeking or deceptive alignment. However, these behaviors can still exist as *terminal values*, for example when a model is trained to predict power-seeking or deceptively aligned agents.
* **LLM pretraining is not cognitively myopic** because there is an incentive to think about the future to improve immediate prediction accuracy, like when predicting the next move in a chess game.
* **LLM pretraining is not value/prediction myopic** (does not maximize myopic prediction accuracy) because of the details of the transformer architecture. Training gradients flow through attention connections, so past computation is directly optimized to be useful when attended to by future computation. This incentivizes improving prediction accuracy over the entire sequence, not just the next token. This means that the model can and will implicitly sacrifice next-token prediction accuracy for long horizon prediction accuracy.
* You can modify the transformer architecture to remove the incentive for non-myopic accuracy, but as expected, the modified architecture has worse scaling laws.
* **LLM pretraining on** ***human*** **data is consequence-blind** as the training data is causally independent from the model's actions. This implies the model should predict actions without considering the effect of its actions on other agents, including itself. This makes the model miscalibrated, but likely makes alignment easier.
* When LLMs are trained on data which has been influenced or generated by LLMs, [the assumptions of consequence-blindness partially break down](https://www.alignmentforum.org/posts/aBRS3x4sPSJ9G6xkj/underspecification-of-oracle-ai). It’s not clear how this affects the training goal theoretically or in practice.
* A myopic training goal does not ensure the model will learn myopic computation or behavior because [inner alignment with the training goal is not guaranteed](https://www.alignmentforum.org/posts/GqxuDtZvfgL2bEQ5v/arguments-against-myopic-training)
Introduction
============
The concept of [myopia](https://www.alignmentforum.org/tag/myopia) has been frequently discussed as a potential solution to the problem of deceptive alignment. However, the term myopia is ambiguous and can refer to multiple different properties we might want in an AI system, only some of which might rule out deceptive alignment. There's also been confusion about the extent to which Large language model (LLM) pretraining and other supervised learning methods are myopic and what this implies about their cognition and safety properties. This post will attempt to clarify some of these issues, mostly by summarizing and contextualizing past work.
Types of Myopia
===============
1. Cognitive Myopia
-------------------
One natural definition for myopia is that the model doesn't think about or consider the future at all. We will call this cognitive myopia. Myopic cognition likely comes with a significant capabilities handicap, as many tasks require some degree of forward planning or anticipation of future events.
LLM pretraining is not cognitively-myopic. Even though LLMs like GPT are optimized for next-token prediction and use causal masking which hides the future from current predictions, there is still a direct incentive to think about the future because it can be useful for immediate prediction accuracy. In a game of chess, efficiently computing the best move likely involves reasoning about how your opponent will respond multiple moves into the future. In other words, we should expect GPT to think about the future because it’s instrumental for predicting the present.
2. Value/Prediction Myopia
--------------------------
Value myopia refers to agents that place no value on future states or rewards. In the case of a predictive model, we mean the model cares only about the accuracy of its next prediction. Intuitively, it seems like this would rule out deceptive alignment and treacherous turns, because a value-myopic agent should be unwilling to sacrifice immediate value for long-term value. Unfortunately, value-myopic agents might still have non-myopic incentives for reasons like anthropic uncertainty or acausal trade. (see [Open Problems with Myopia](https://www.alignmentforum.org/posts/LCLBnmwdxkkz5fNvH/open-problems-with-myopia) for more details)
In any case, LLM training actually incentivizes value non-myopia. Even though next-token prediction accuracy is a nominally myopic objective, the transformer architecture actually causes the model to be optimized for accuracy over the entire sequence, not just the next token.
In a causal-masked transformer, attention layers can query the previous layers' activations from any column in the context window. Gradients flow through the attention connections, so each previous layer is optimized not just to improve prediction accuracy for the next token, but also to produce values that are useful for future columns to attend to when predicting their token.
*The activations of the red block are used by each orange attention block and optimized to be useful for all of them. For simplicity, only attention layers are shown. Thanks to Nicholas Kees Dupuis for the diagram.*This means that the model can and will implicitly sacrifice next-token prediction accuracy for long horizon prediction accuracy. In particular, we should expect trained models to find an efficient tradeoff between current prediction accuracy and full-sequence prediction accuracy. When predicting relatively easy tokens, they will likely spend most of their computation budget preparing for the future. If the model is given a hard problem that it will need to solve later in the sequence, it will likely work on the problem ('in the back of its head') throughout the intermediate predictions.
### Experiments
**Interpretability**
Several interpretability results including [ROME](https://arxiv.org/abs/2202.05262) confirm this type of non-myopic computation in LLMs. ROME shows that LLMs recall factual information about nouns when the noun first appears, even though this information is only used later when predicting the answer to a question about the noun. This information would be irrelevant and thus wasted computation for the purpose of predicting only the next token. For example, if the model sees the text "The Eiffel Tower", it immediately begins retrieving information about the Eiffel Tower like where it is located even though that's not necessary to predict the next token which is almost certainly "is".
**Enforcing Myopia**
It is possible to modify the transformer architecture to enforce value (prediction accuracy) myopia by placing stop gradients in the attention layers. This effectively prevents past activations from being directly optimized to be more useful for future computation. We ran several informal experiments on models like these while at Conjecture. Unfortunately, we do not have quantitative results to share here. The experiments were preliminary and we moved on to other aspects of the project, so don’t take this as strong evidence.
Specifically, we trained a set of four traditional and four myopic transformers ranging from 117M to 1.5B parameters (equivalent to GPT-2 Small to GPT-XL). Each model was trained on the same data but training hyperparameters were tuned to each architecture individually using [maximal update parameterization](https://arxiv.org/abs/2203.03466).
We found the performance reduction from myopia was minimal at 117M parameters, but the performance cost increased with scale, i.e. myopic transformers have worse scaling laws.
3. Consequence-blindness
------------------------
A third type of myopia to consider is [consequence-blindness](https://www.alignmentforum.org/posts/aBRS3x4sPSJ9G6xkj/underspecification-of-oracle-ai), where a model chooses actions completely independent of any effect of its actions on the future. This is similar to the goal of [Counterfactual Oracles](https://arxiv.org/pdf/1711.05541.pdf).
*A causal diagram of the supervised learning process from* [*Modeling AGI Safety Frameworks with Causal Influence Diagrams*](https://arxiv.org/pdf/1906.08663.pdf)Consequence-blindness should rule out most types of instrumental convergence and concerns about [self-fulfilling prophecies](https://www.alignmentforum.org/posts/SwcyMEgLyd4C3Dern/the-parable-of-predict-o-matic). A model which completely ignores the effects of its actions has no instrumental incentive to pursue traditional instrumental goals, like trying to accumulate resources to become more powerful, trying to prevent its own shutdown, or pretending to be aligned in order to defect later. However, consequence-blindness does not actually constrain the behavior of a model, because the model can pursue any instrumental goal as a *terminal* value.
A consequence-blind simulator that predicts power-seeking agents (like humans) will still predict actions which seek power, but these actions will seek power for the simulated agent, not the simulator itself. I usually think about problems like this as [simulator vs simulacra alignment](https://www.alignmentforum.org/posts/vJFdjigzmcXMhNTsx/simulators). If you successfully build an inner aligned simulator, you can use it to faithfully simulate according to the rules it learns and generalizes from its training distribution. However you are still left with the problem of extracting consistently aligned simulacra.
In theory, consequence-blindness doesn't rule out any capabilities, because a consequence-blind predictor could learn to predict any behavior. However, in practice using a consequence-blind training goal like pure imitation learning may be uncompetitive compared to methods like RL (or imitation + RL finetuning, the current dominant paradigm).
Consequence-blind agents (with a causal decision theory) can be seen as implementing a [Lonely Causal Decision Theory (LCDT](https://www.alignmentforum.org/posts/Y76durQHrfqwgwM5o/lcdt-a-myopic-decision-theory#Myopic_simulation)). An LCDT agent assumes that every other decision node of agents in the world (including its future decisions) are causally independent of its actions. This means it has no incentive to take actions which help its future itself or other agents for instrumental reasons.
Unlike the other forms of myopia above, the training goal for LLMs trained with self-supervised learning (SSL) is theoretically consequence-blind. In supervised or self-supervised learning, the training data already exists and is assumed to be causally independent from the model’s decisions. This means a model’s prediction should be based only on the likelihood of the output appearing in the training data. In particular, the model’s prediction should be independent of any effect from making the prediction itself, including whether or not the prediction would make the model more likely to predict or control the future correctly when run autoregressively.
The distinction between optimizing prediction accuracy and steering the distribution to be easier to predict is one of the most common sources of confusion about LLM myopia. Even though the LLM training goal is not value-myopic and optimizes for prediction accuracy across entire training examples, LLMs are not incentivized to predict tokens that make the future easier to predict.
Unfortunately, the analysis of consequence-blindness story becomes more complicated when the training data is not actually causally independent from the model’s prediction.
Self-reference and Model Generated Data
=======================================
In practice, the assumption that the supervised learning distribution is causally independent from the model’s predictions may be violated. This can happen whenever a model is trained on data correlated with its behavior, or if the model *thinks* the correct prediction is correlated with its behavior.
*Another causal diagram from* [*Modeling AGI Safety Frameworks with Causal Influence Diagrams*](https://arxiv.org/pdf/1906.08663.pdf)This is concerning because it's not at all clear what [a model that is predicting itself should output](https://www.alignmentforum.org/s/n3utvGrgC2SGi9xQX/p/3kkmXfvCv9DmT3kwx#Model_predicts_itself). It breaks many of the intuitions of why it should be safe to use LLMs as simulators of text distributions.
This problem is already apparent in current LLMs and is likely to become worse for several reasons.
* Increasingly large portions of LLM pretraining and finetuning datasets will be generated or influenced by LLMs themselves
* LLMs are often trained or used in ways designed to make them more situationally aware that they are an AI assistant responding to a user’s prompt
* In my experience, larger models often become aware that they are a LLM generating text rather than predicting an existing distribution. This is possible because generated text drifts off distribution and can be distinguished from text in the training corpus.
We could try to combat this problem by excluding model generated text from all future training data, but it is still possible for a model trained on human-written text to drift off distribution and become situationally aware. You could see this as a training goal (outer alignment) problem: the decision-myopic base objective has to generalize somehow to self-referential predictions which are fundamentally not consequence-blind.
We could also try the opposite of actually *increasing* the amount of model generated text in the training dataset. This could allow us to shape the LLMs prior on how LLMs (and thus itself) should behave. However it is still possible for a model to realize (or believe) that it is different from the types of models in the training data and should thus act differently.
In this vein, there have been [several](https://www.alignmentforum.org/posts/Aufg88v7mQ2RuEXkS/proper-scoring-rules-don-t-guarantee-predicting-fixed-points) [great](https://www.alignmentforum.org/posts/i3v7WeCXyWiYfhihF/stop-gradients-lead-to-fixed-point-predictions#3__Performative_stability_and_game_theory) [posts](https://www.alignmentforum.org/posts/aBRS3x4sPSJ9G6xkj/underspecification-of-oracle-ai) that discuss how the process of repeatedly training a model on its own predictions might converge. There has also been work considering the general implications of [training a simulator in a closed loop](https://www.lesswrong.com/posts/YEioD8YLgxih3ydxP/why-simulator-ais-want-to-be-active-inference-ais). However the broader implications for generalization and alignment are unclear.
Myopic Training Goals vs Myopic Models
======================================
It is also important to note that even if a training goal is designed to be myopic in some way, the resulting model may not be myopic. [Inner alignment failures can lead to non-myopic models emerging from myopic training goals](https://www.alignmentforum.org/posts/GqxuDtZvfgL2bEQ5v/arguments-against-myopic-training?_ga=2.234984129.1917697003.1690249492-617721652.1678859545). Finding a solution to inner alignment, or getting [inner alignment by default does seem relatively likely for predictive SSL](https://www.alignmentforum.org/posts/qoHwKgLFfPcEuwaba/conditioning-predictive-models-making-inner-alignment-as?_ga=2.28965628.1917697003.1690249492-617721652.1678859545) over other training goals, but it is not guaranteed. Many researchers believe the cognitive structures that are required to predict the answers to hard consequentialist problems will fundamentally be non-myopic, especially if these structures become situationally aware. [Some examples](https://www.alignmentforum.org/posts/5ciYedyQDDqAcrDLr/a-positive-case-for-how-we-might-succeed-at-prosaic-ai?commentId=st5tfgpwnhJrkHaWp#fn-yooXzTTTZucoQZYPg-1).
*Credit to* [*Leo Gao*](https://twitter.com/nabla_theta) *for the meme*It would be a huge success if we could find some way to enforce or verify that a model's internal computation satisfies some [myopic criteria](https://www.alignmentforum.org/posts/GeabLEXYP7oBMivmF/acceptability-verification-a-research-agenda) (or any criteria…) during or after training. However, it's not clear how we would go about this.
Meta
====
The ideas in the post are from a human, but most of the text was written by Chat GPT-4 with prompts and human curation using Loom. I endorse the post as technically correct and accurately phrased according to my understanding. Here is the second of two Loom trees used to generate most of the post before final edits.
A Loom tree used to generate a draft for this post with GPT-4 |
4b092dde-1f58-4119-99fe-e869b28565e7 | trentmkelly/LessWrong-43k | LessWrong | Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
I'm not going to add a bunch of commentary here on top of what we've already put out, since we've put a lot of effort into the paper itself, and I'd mostly just recommend reading it directly, especially since there are a lot of subtle results that are not easy to summarize. I will say that I think this is some of the most important work I've ever done and I'm extremely excited for us to finally be able to share this. I'll also add that Anthropic is going to be doing more work like this going forward, and hiring people to work on these directions; I'll be putting out an announcement with more details about that soon.
EDIT: That announcement is now up!
Abstract:
> Humans are capable of strategically deceptive behavior: behaving helpfully in most situations, but then behaving very differently in order to pursue alternative objectives when given the opportunity. If an AI system learned such a deceptive strategy, could we detect it and remove it using current state-of-the-art safety training techniques? To study this question, we construct proof-of-concept examples of deceptive behavior in large language models (LLMs). For example, we train models that write secure code when the prompt states that the year is 2023, but insert exploitable code when the stated year is 2024. We find that such backdoor behavior can be made persistent, so that it is not removed by standard safety training techniques, including supervised fine-tuning, reinforcement learning, and adversarial training (eliciting unsafe behavior and then training to remove it). The backdoor behavior is most persistent in the largest models and in models trained to produce chain-of-thought reasoning about deceiving the training process, with the persistence remaining even when the chain-of-thought is distilled away. Furthermore, rather than removing backdoors, we find that adversarial training can teach models to better recognize their backdoor triggers, effectively hiding the unsafe behavior. Our results sugge |
979f6d9b-d8cb-4c35-a682-ba246343e458 | trentmkelly/LessWrong-43k | LessWrong | AI and Non-Existence.
Imagine 2 valleys. One valley leads to billions and billions of years of extreme joy, while the other valley leads to billions and billions of years of extreme suffering. A superintelligence comes to you and tells you that there is a 98% chance that you will end up in the valley of joy for billions of years and a 2% chance that you will end up in the valley of suffering for billions of years. It also gives you a choice to choose non-existence. Would you take the gamble, or would you choose non-existence?
The argument presented in this post occurred to me several months ago and in the last several months, I have spent time thinking about the argument and discussed it with AI models and have not found a satisfactory answer to it given the real-world situation. The argument can be formulated for things other than advanced AI, but given the rapid progress in the AI field and that the argument was originally formulated in the context of AI, I will present it in that context.
Now apply the reasoning from the valleys from above to AGI/ASI. AGI could be here in about 15 months and ASI not long after that. Advanced AI(s) could prolong human life to billions and billions of years, take over the world and create a world in its image - whatever that might be. People have various estimates of how likely it is that the AGI/ASI will go wrong, but one thing that many of them keep saying is that the worst case scenario is that it will kill us all. That is not the worst case scenario. The worst case scenario is that it will cause extreme suffering or torture us for billions and trillions of years.
Let's assume better than 2% odds, let's say they are 0.5%, would you be willing to take the gamble with heaven or hell even if the odds are 0.5% for hell? And if not, at what point would you be willing to take the gamble instead of choosing non-existence?
If some of you might say that you would be willing to take the gamble at 0.5% for a living hell, in this case, would you be willing t |
8f0c34d4-49c8-4899-9f3f-9b76d7ea9aac | trentmkelly/LessWrong-43k | LessWrong | Believing others' priors
Meet the Bayesians
In one way of looking at Bayesian reasoners, there are a bunch of possible worlds and a bunch of people, who start out with some guesses about what possible world we're in. Everyone knows everyone else's initial guesses. As evidence comes in, agents change their guesses about which world they're in via Bayesian updating.
The Bayesians can share information just by sharing how their beliefs have changed.
> "Bob initially thought that last Monday would be sunny with probability 0.8, but now he thinks it was sunny with probability 0.9, so he must have has seen evidence that he judges as 4/9ths as likely if it wasn't sunny than if it was"
If they have the same priors, they'll converge to the same beliefs. But if they don't, it seems they can agree to disagree. This is a bit frustrating, because we don't want people to ignore our very convincing evidence just because they've gotten away with having a stupid weird prior.
What can we say about which priors are permissible? Robin Hanson offers an argument that we must either (a) believe our prior was created by a special process that correlated it with the truth more than everyone else's or (b) our prior must be the same as everyone else's.
Meet the pre-Bayesians
How does that argument go? Roughly, Hanson describes a slightly more nuanced set of reasoners: the pre-Bayesians. The pre-Bayesians are not only uncertain about what world they're in, but also about what everyone's priors are.
These uncertainties can be tangled together (the joint distribution doesn't have to factorise into their beliefs about everyone's priors and their beliefs about worlds). Facts about the world can change their opinions about what prior assignments people have.
Hanson then imposes a pre-rationality condition: if you find out what priors everyone has, you should agree with your prior about how likely different worlds are. In other words, you should trust your prior in the future. Once you have this condition, it seems |
5ea4e95d-1ed2-4be1-86d4-f5e9b8556fda | trentmkelly/LessWrong-43k | LessWrong | Settled questions in philosophy
Philosophy is notorious for not answering the questions it tackles. Plato posed most of the central questions more than two millennia ago, and philosophers still haven't come to much consensus about them. Or at least, whenever philosophical questions begin to admit of answers, we start calling them scientific questions. (Astronomy, physics, chemistry, biology, and psychology all began as branches of philosophy.)
A common attitude on Less Wrong is "Too slow! Solve the problem and move on." The free will sequence argues that the free will problem has been solved.
I, for one, am bold enough to claim that some philosophical problems have been solved. Here they are:
* Is there a God? No.
* What's the solution to the mind-body problem? Materialism.
* Do we have free will? We don't have contra-causal free will, but of course we have the ability to deliberate on alternatives and have this deliberation effect the outcome.
* What is knowledge? (How do we overcome Gettier?) What is art? How do we demarcate science from non-science? If you're trying to find simple definitions that match our intuitions about the meaning of these terms in ever case, you're doing it wrong. These concepts were not invented by mathematicians for use in a formal system. They evolved in practical use among millions of humans over hundreds of years. Stipulate a coherent meaning and start using the term to successfully communicate with others.
There are other, smaller questions that I think are solved, too, but for now I'm curious: Which philosophical problems do you think are solved, and what is the answer?
|
04491310-60e7-4f68-9211-16ce04ebd351 | trentmkelly/LessWrong-43k | LessWrong | A transparency and interpretability tech tree
Thanks to Chris Olah, Neel Nanda, Kate Woolverton, Richard Ngo, Buck Shlegeris, Daniel Kokotajlo, Kyle McDonell, Laria Reynolds, Eliezer Yudkowksy, Mark Xu, and James Lucassen for useful comments, conversations, and feedback that informed this post.
The more I have thought about AI safety over the years, the more I have gotten to the point where the only worlds I can imagine myself actually feeling good about humanity’s chances are ones in which we have powerful transparency and interpretability tools that lend us insight into what our models are doing as we are training them.[1] Fundamentally, that’s because if we don’t have the feedback loop of being able to directly observe how the internal structure of our models changes based on how we train them, we have to essentially get that structure right on the first try—and I’m very skeptical of humanity’s ability to get almost anything right on the first try, if only just because there are bound to be unknown unknowns that are very difficult to predict in advance.
Certainly, there are other things that I think are likely to be necessary for humanity to succeed as well—e.g. convincing leading actors to actually use such transparency techniques, having a clear training goal that we can use our transparency tools to enforce, etc.—but I currently feel that transparency is the least replaceable necessary condition and yet the one least likely to be solved by default.
Nevertheless, I do think that it is a tractable problem to get to the point where transparency and interpretability is reliably able to give us the sort of insight into our models that I think is necessary for humanity to be in a good spot. I think many people who encounter transparency and interpretability, however, have a hard time envisioning what it might look like to actually get from where we are right now to where we need to be. Having such a vision is important both for enabling us to better figure out how to make that vision into reality and also fo |
4a64b724-4b1b-473d-afbe-bab17fb27d77 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Modeling Risks From Learned Optimization
*This post, which deals with how risks from learned optimization and inner alignment can be understood, is part 5 in our* [*sequence on Modeling Transformative AI Risk*](https://www.alignmentforum.org/s/aERZoriyHfCqvWkzg)*. We are building a model to understand debates around existential risks from advanced AI. The model is made with* [*Analytica*](https://en.wikipedia.org/wiki/Analytica_(software)) *software, and consists of nodes (representing key hypotheses and cruxes) and edges (representing the relationships between these cruxes), with final output corresponding to the likelihood of various potential failure scenarios. You can read more about the motivation for our project and how the model works in the* [*Introduction post*](https://www.alignmentforum.org/posts/qnA6paRwMky3Q6ktk/modelling-transformative-ai-risks-mtair-project-introduction)*. The previous post in the sequence,* [*Takeoff Speeds and Discontinuities*](https://www.alignmentforum.org/posts/pGXR2ynhe5bBCCNqn/takeoff-speeds-and-discontinuities)*, described the different potential characteristics of a transition from high-level machine intelligence [1] to superintelligent AI.*
*We are interested in feedback on this post, especially in places where the model does not capture your views or fails to include an uncertainty that you think could be an important crux. Similarly, if an explanation seems confused or confusing, flagging this is useful – both to help us clarify, and to ensure it doesn’t reflect an actual disagreement.*
This post explains how risks from learned optimization are incorporated in our model. The relevant part of the model is mostly based on the Risks from Learned Optimization [sequence](https://www.alignmentforum.org/s/r9tYkB2a8Fp4DN8yB) and [paper](https://arxiv.org/pdf/1906.01820.pdf) (henceforth RLO). Although we considered responses and [alternate](https://www.alignmentforum.org/posts/nFDXq7HTv9Xugcqaw/is-the-term-mesa-optimizer-too-narrow) [perspectives](https://www.lesswrong.com/posts/WmBukJkEFM72Xr397/mesa-search-vs-mesa-control) to RLO in our research, these perspectives are not currently modeled explicitly.
For those not familiar with the topic, a *mesa-optimizer* is a [learned algorithm](https://en.wikipedia.org/wiki/Machine_learning) that is itself an optimizer. According to RLO, [inner alignment](https://www.alignmentforum.org/posts/pL56xPoniLvtMDQ4J/the-inner-alignment-problem) is the problem of aligning the objective of a mesa-optimizer with the objective of its *base optimizer* (which may be specified by the programmer). A contrived [example](https://www.alignmentforum.org/posts/AHhCrJ2KpTjsCSwbt/inner-alignment-explain-like-i-m-12-edition) supposes we want an algorithm that finds the shortest path through any maze. In the training data, all mazes have doors that are red, including the exit. Inner misalignment arises if we get an algorithm that efficiently searches for the next red door *–* the capabilities are robust because the search algorithm is general and efficient, but the [objective is not robust](https://www.alignmentforum.org/posts/2mhFMgtAjFJesaSYR/2-d-robustness) because it finds red doors rather than the exit.
The relevant part of our model is contained in the **Mesa-Optimization** module:
Right-click and select "open image in new tab" (or similar) to see images full-sizeThe output of the **Mesa-Optimization** module is an input to the **Incorrigibility** module. The logic is that inner misalignment is one way that high-level machine intelligence (HLMI) could become [incorrigible](https://www.lesswrong.com/posts/fkLYhTQteAu5SinAc/corrigibility), which in turn counts strongly against being able to **Correct course as we go** in the development of HLMI.
Module Overview
===============
Overview of the Mesa-Optimization module. We recommend reading top-to-bottom, left-to-right.The top-level logic of the **Mesa-Optimization** module is: HLMI has an inner alignment failure if
1. The HLMI contains a mesa-optimizer, AND
2. Given (1), the mesa-optimizer is pseudo-aligned, i.e. it acts aligned in the training setting but its objective is not robust to other settings, AND
3. Given (2), the pseudo-alignment is not sufficient for intent alignment, i.e. is not safe enough to make HLMI corrigible, AND
4. Given (3), we fail to stop deployment of the unsafe system.
In the following sections, we explain how each of these steps are modeled.
HLMI contains a mesa-optimizer
------------------------------
The output of this section is **HLMI contains a mesa-optimizer**, which depends on three nodes. The left node, **HLMI is trained with a base optimizer**, means that a training algorithm optimized a distinct learned algorithm, and that learned algorithm forms all or part of the HLMI system. A crux here is what **Type of HLMI** you expect, which comes out of the [pathways discussed in the Paths to HLMI post](https://www.alignmentforum.org/s/aERZoriyHfCqvWkzg/p/amK9EqxALJXyd9Rb2#Inside_view_model_of_specific_pathways). For instance, HLMI via current deep learning methods or evolutionary methods will involve a base optimizer, but this is not true of other pathways such as whole-brain emulation.
The middle node, **Argument for mesa-optimization**, represents the argument from first principles for why mesa-optimization would occur in HLMI. It is mainly based on the post [Conditions for Mesa-Optimization](https://www.alignmentforum.org/posts/q2rCMHNXazALgQpGH/conditions-for-mesa-optimization). This is broken down further into three nodes. **Mesa-optimization is more efficient than highly tuned algorithms** distils some claims about the advantages of mesa-optimization compared to systems without mesa-optimization, including that it offers better generalisation through search. You could reject that claim on the grounds that sample efficiency will be high enough, such that HLMI consists of a bunch of algorithms that are highly tuned to different domains. You could also argue that some **types of HLMI** are more prone to be highly tuned than others. For instance, evolutionary algorithms might be more likely to mesa-optimize than machine learning, and machine learning might be more likely to mesa-optimize than hybrid ML-symbolic algorithms.
The middle node, **Training task is generic**, resolves as positive if the learned algorithm is not directly optimized for domain-specific tasks. This scenario is characteristic of pre-training in modern machine learning. For example, the task of predicting the next word in a large text dataset is generic, because the dataset could contain all kinds of content that is relevant to many different domain-specific tasks. In contrast, one could use a more domain-specific dataset (e.g. worded math problems) or objective function (e.g. reward for the quality of a news article summary). This crux is important if mesa-optimizers generalise better than other kinds of algorithms, because then a generic training task would tend to select for mesa-optimizers more. [GPT-3](https://arxiv.org/abs/2005.14165) gives credence to this idea, because it demonstrates strong few-shot learning performance by simply learning to predict the next word in a text. However, it's uncertain if GPT-3 meets the definition of a mesa-optimizer.
Finally, **Inductive biases in base optimizer** lumps together some factors about inductive bias in the HLMI’s architecture and training algorithm which affect the chance of mesa-optimization. For example, the extent that mesa-optimization exists in the space of possible models (i.e. algorithmic range), and the ease by which the base optimizer finds a mesa-optimizer (i.e. reachability). Inductive bias is a big factor in some people's beliefs about inner alignment, but there is disagreement about just how important it is and how it works (see e.g. [Inductive biases stick around](https://www.alignmentforum.org/posts/nGqzNC6uNueum2w8T/inductive-biases-stick-around), including comments).
### Analogies for mesa-optimization
Finally, evidence for the node **HLMI contains a mesa-optimizer** is also drawn from history and analogies. We modeled this evidence in a submodule, shown below. The submodule is structured as a Naive Bayes classifier: it models the likelihood of the evidence, given the hypotheses that an HLMI system does or does not contain a mesa-optimizer. The likelihood updates the following prior: if you only knew the definition of mesa-optimizer, and hadn't considered any specific cases, arguments or evidence for it, what is the probability of an optimizing system containing a mesa-optimizer?
We considered three domains for the evidence: machine learning today, firms with respect to economies, and humans with respect to natural selection. A few more specific nodes are included under **Examples from ML today** because there has been some interesting discussion and disagreement in this area:
* The "spontaneous emergence of learning algorithms" from reinforcement learning algorithms has been [cited](https://www.alignmentforum.org/posts/Wnqua6eQkewL3bqsF/matt-botvinick-on-the-spontaneous-emergence-of-learning#Implications) as evidence for mesa-optimization, but [this may not be informative](https://www.alignmentforum.org/posts/Wnqua6eQkewL3bqsF/matt-botvinick-on-the-spontaneous-emergence-of-learning?commentId=pYpPnAKrz64ptyRid).
* Meta-learning is a popular area of ML research, but since it applies a deliberate and outer-loop optimization, it doesn't clearly affect the likelihood of spontaneous mesa-optimization.
* [GPT-3 does few-shot learning](https://arxiv.org/abs/2005.14165) in order to perform novel tasks. This [post](https://www.alignmentforum.org/posts/BGD5J2KAoNmpPMzMQ/why-gpt-wants-to-mesa-optimize-and-how-we-might-change-this) gives an argument for why it is incentivised for – or may already be – mesa-optimizing.
* It is unclear and [debated](https://www.alignmentforum.org/posts/q2rCMHNXazALgQpGH/conditions-for-mesa-optimization?commentId=mRX5uRXnkihtWPt24) how much AlphaZero marginally benefits from Monte Carlo Tree Search, which is a form of mechanistic optimization, compared to just increasing the model size. In turn it is unclear how much evidence AlphaZero provides for getting better generalisation through search, which is [argued](https://www.alignmentforum.org/s/r9tYkB2a8Fp4DN8yB/p/q2rCMHNXazALgQpGH#2_1__The_task) as an advantage of mesa-optimization.
The mesa-optimizer is pseudo-aligned
------------------------------------
The possibility of pseudo-alignment depends on just two nodes in our model. The structure is simple mainly because pseudo-alignment of any kind seems much more likely than robust alignment, so we haven't noticed much debate on this point. In general there are many more ways to perform well on the training distribution which are not robustly aligned with the objective, i.e. they would perform significantly worse on that objective under some realistic shift in the distribution. And in practice today, this kind of robustness is a major challenge in ML ([Concrete Problems in AI Safety](https://arxiv.org/abs/1606.06565) section 7 gives an overview, though it was published back in 2016).
The dependency on the left is a module, **Analogies for pseudo-alignment**, which is structured identically to **Analogies for mesa-optimization** (i.e. with a Naive Bayes classifier, and so on), but the competing hypotheses are pseudo-alignment and robust alignment, and the analogies are simply "ML systems today", "Firm", and "Human".s
The second node influencing pseudo-alignment is **Mesa-optimizer uses modeling**. The concept of "modeling" vs. "internalization" introduced in the [RLO paper](https://arxiv.org/pdf/1906.01820.pdf) (section 4.4) is relevant to pseudo-alignment. Internalization implies robust alignment, whereas modeling means the mesa-optimizer is pointing to something in its input data and/or its model of the world in order to act aligned. We explain this node and the implications of "modeling" in more detail in the next section.
Pseudo-alignment is not safe enough
-----------------------------------
At the top level of this subsection, we include three reasons why pseudo-alignment might not be safe enough to count as a failure of inner alignment. Firstly, there is a crux of whether **corrigibility has a broad basin of attraction**. This refers to Paul Christiano's [claim](https://ai-alignment.com/corrigibility-3039e668638) that "A sufficiently corrigible agent will tend to become more corrigible and benign over time. Corrigibility marks out a broad basin of attraction towards acceptable outcomes." If Christiano’s claim about corrigibility is true, this increases the overall chance that a pseudo-aligned algorithm becomes safe enough before it locks in a path to catastrophe.
A second crux for the safety of pseudo-alignment is **how malign we expect a mesa-objective to be by default** (by malign, we just mean non-benign, or harmful *in effect* *–* it doesn't have to be inherently malicious). It's uncertain what mesa-objectives are generally like, because they arise internally from the base optimizer, and there is currently scant empirical evidence of mesa-optimization. It's reasonable to expect a proxy objective to be much closer to the base objective than to a random objective, so the danger partly depends on the base objective. Perhaps mesa-objectives will just be weird in benign ways, e.g. being irrational, or very [local](https://www.alignmentforum.org/posts/HkWB5KCJQ2aLsMzjt/locality-of-goals). On the other hand, it seems that simple, unbounded, coherent objectives would tend to have lower description length than objectives lacking those attributes, so those objectives may be favoured by [simplicity bias](https://en.wikipedia.org/wiki/Minimum_description_length).
Before moving on to the third reason, *deceptive alignment*, let's review the concept of *modeling* vs. *internalization* introduced in the [RLO paper](https://arxiv.org/pdf/1906.01820.pdf) (section 4.4). Modeling is where a mesa-optimizer acts aligned with a base objective (in training) by representing the base objective within its world model (using relevant information from its input data) and then deliberately acting aligned with this representation. By understanding the base objective, the mesa optimizer can behave more aligned, whether or not its own objective is different. (In this case, “modeling” is similar to a trial employee working hard because their mental model of the company indicates that the company wants them to work hard.) The alternative manner to modeling in which a mesa-optimizer can act aligned with a base objective is *internalization:* the base optimizer adjusts the mesa-objective to be more and more aligned until convergence. (In this case, internalization is similar to a trial employee working hard because they intrinsically want to work hard.)
Modeling-based mesa-optimizers are thought to end up [*deceptively* or *corrigibly*](https://www.alignmentforum.org/posts/zthDPAjh9w6Ytbeks/deceptive-alignment#4_4__Internalization_or_deception_after_extensive_training) aligned as a result of training. Deceptive alignment means the learned algorithm optimizes for the base objective during training for instrumental reasons (namely, to deceive the learning algorithm or programmers into thinking it’s aligned). Returning to the example of our trial employee, a deceptively aligned trial employee might just want a salary and to slack off, so he will work hard while in trial to be offered the job, and then after being offered the job only work the minimal amount to not get fired. So a deceptively aligned algorithm deviates when it becomes favourable, e.g. it figures out that training has ended, removing the threat of being penalised.
The main alternative to a deceptive mesa-optimizer that we consider is a *corrigible* mesa-optimizer. A corrigible mesa-optimizer optimizes for a *pointer* to the part of its world model that includes the base objective (both in training and deployment). The corrigibility comes from the fact that the base optimizer can modify this pointer to be more accurate. Here, our trial employee cares about being "the perfect employee" both in trial and once offered a job, but they are uncertain how to measure that. So they might observe employees that have succeeded (the pointer) and learn to imitate that. Note that this is subtly different from the internalizing employee. The internalizing employee doesn't care about the company in itself, but values working hard for its own sake (and thus happens to be aligned with the company), whereas the corrigibly modeling employee wants to do whatever they believe the company wants of them, which is to work hard.
Coming back to our model, the third reason that pseudo-alignment might not be safe enough is if the **Mesa-optimizer is deceptively aligned**.
We broke down deceptive alignment into two main conditions. First, a precondition that the **mesa-optimizer is using modeling**. The key question influencing this precondition is **how easy it is for a mesa-optimizer to model rather than internalize**. We represent this as the likelihood ratio of eventually settling on modeling versus internalization. One argument for modeling being easier is that, within the training data for the HLMI, there will likely be a huge amount of rich input data relevant for determining the base objective, and it will therefore be easier for the HLMI to capture this objective function by referring to the relevant parts of this data, than it will be to build a representation of the objective from scratch. However, this argument depends on justhow difficult it is to construct representations that are useful for modeling the objective, which in turn depends both on the nature of the data and the objective. We can also look to analogies here, such as [imprinting](https://en.wikipedia.org/wiki/Imprinting_(psychology)) in animals. For example, rather than have a detailed representation of a mother goose, a gosling imprints on some initial stimulus in its environment, whether that be an actual goose or, say, a human's boots, and then treats the imprinted object as its mother (e.g. it follows the object around).
Feeding into the **ease of modeling vs. internalization** node is **Reduction factor due to R&D for objective robustness** (*objective robustness* is terminology from [Clarifying inner alignment terminology](https://www.alignmentforum.org/posts/SzecSPYxqRa5GCaSF/clarifying-inner-alignment-terminology)). This node quantifies the effect that safety research has (as a constant multiplier) on reducing the **Ease of modeling vs. internalizing the base objective**, assuming that modeling is dangerous and better to avoid.
The second condition for deceptive alignment is simply: given that a mesa-optimizer uses modeling, it is deceptive. This possibility is broken down into a submodule of nodes, explained in the next section.
### Deceptive alignment
*The inside of the "Modeling-based mesa-optimizer is deceptive" submodule.*
There are three factors feeding into the **modeling-based mesa-optimizer is deceptive** node. First is **Relative number of ways to get a deceptive rather than corrigible mesa-optimizer**. Think of this as a base rate for deceptive (rather than corrigle) mesa-optimization. There is a [counting argument](https://en.wikipedia.org/wiki/Combinatorics) for deception here: since deception is instrumentally motivated, it is compatible with almost any proxy goal, whereas for corrigibility, the base optimizer has to find a very specific pointer to relate to the base objective.
The second factor is the **Ease of finding deceptive (rather than corrigible) mesa-optimizers**. Similar to [*reachability*](https://www.alignmentforum.org/posts/q2rCMHNXazALgQpGH/conditions-for-mesa-optimization), this node is represented by the relative likelihood that the base optimizer outputs a deceptive mesa-optimizer at some point in training.
The post [Does SGD Produce Deceptive Alignment?](https://www.alignmentforum.org/posts/ocWqg2Pf2br4jMmKA/does-sgd-produce-deceptive-alignment) details some arguments for and against deceptive algorithms being easier to find than corrigible alignment (specifically for stochastic gradient descent, but much of it is more general). Becoming deceptive requires caring about future instances, so that there is an incentive to deceive now for later gain. Becoming deceptive also requires reasoning about the training process, so that the mesa-optimizer expects the threat of modification to go away and for deception to pay off. On the other hand, becoming corrigible requires a *robust* pointer to a good model of the base objective *–* otherwise the mesa-optimizer will just optimize for whatever proxy it has. Furthermore, even if there is information related to the base objective in the input data, the mesa-optimizer will need to have uncertainty about precisely what objective it is pointing to. The conditions for deception are claimed to be easier [2]. However, there seems to be a lot of uncertainty here which might hinge on the nature of the base objective and how the base optimizer works.
Going one level down from the **Ease of finding deceptive (rather than corrigible) mesa-optimizers** node, we have **Modeling-based mesa-optimizer cares about future instances**. This means the mesa-optimizer is motivated by what happens in future instances of training or in deployment (instead of just its current round of training). Caring about future instances is a necessary condition for deceptive alignment, but not corrigible alignment *–* if an algorithm does not consider future instances, then it will not see the benefit of defecting later in time, and therefore has no instrumental reason to deceive. The node for **Reduction factor due to R&D for myopic cognition** refers to the idea that methods to induce myopic cognition, such as [approval-directed agents](https://ai-alignment.com/model-free-decisions-6e6609f5d99e), could reduce the chance of modeling-based mesa-optimizers caring about future instances (myopic cognition is discussed in this [post](https://www.alignmentforum.org/posts/GqxuDtZvfgL2bEQ5v/arguments-against-myopic-training)).
The third top-level node here is **Relative likelihood that deceptive vs. corrigible mesa-optimizer persists** *–* that is, the likelihood that once a mesa-optimizer becomes deceptive in one instance, it survives and remains deceptive indefinitely (relative to the likelihood for a corrigible mesa-optimizer). Deceptive algorithms may be favoured due to their ability and incentive to adapt to the base objective, because they can reason about the objective and they care about future instances. However, protecting a proxy objective from modification may be difficult. This too seems like an important uncertainty that depends on empirical facts about the base optimizer, e.g. how stochastic gradient descent works with modern neural nets.
Finally, we include two areas of research that could reduce the likelihood of deceptive alignment. One is research to improve objective robustness, such as [Relaxed adversarial training for inner alignment](https://www.lesswrong.com/posts/9Dy5YRaoCxH9zuJqa/relaxed-adversarial-training-for-inner-alignment), where one of the aims is avoiding deception. A key part of making relaxed adversarial training work is transparency tools to guide the training process. [This comment](https://www.alignmentforum.org/posts/J9D6Bi3eFDDhCaovi/will-transparency-help-catch-deception-perhaps-not?commentId=yn5YcLnL6vs6AxxAE#comments) argues why that is more helpful than inspecting the algorithm *after* deception may have already occurred. Research into transparency may help to prevent deceptive alignment in other ways, so this is kept as a separate node.
We fail to stop deployment
--------------------------
The last part of the **Mesa-Optimization** module is about whether the overseers of an HLMI project will pull back before it's too late, given there is an HLMI with an unsafe mesa-optimizer that hasn’t been deployed yet (or trained to sufficiently advanced capabilities to “break out”). The way they could pull back in such a situation is if they are aware that it's unsafe, AND they coordinate to stop the threat. That logic is inverted to an OR in the module, to fit with the top-level output of **HLMI has inner alignment failure.**
**Mesa-optimizer is deceptively aligned** is one of the strongest factors in knowing whether a mesa-optimizer is unsafely pseudo-aligned, because deception works against overseers figuring this out. Meanwhile, **R&D for transparency & interpretability** may help to detect unsafe pseudo-aligned algorithms. [On the other hand](https://www.alignmentforum.org/posts/J9D6Bi3eFDDhCaovi/will-transparency-help-catch-deception-perhaps-not), deceptive algorithms may just exploit weaknesses in the transparency tools, so it is even possible for the reduction factor to be less than 1 (i.e. it *increases* the chance that we fail to see the danger).
Conclusion
==========
In this post, we have examined important cruxes and uncertainties about risks from learned optimization, and how they relate to each other. Our model is mostly based on the [Risks from Learned Optimization](https://www.alignmentforum.org/s/r9tYkB2a8Fp4DN8yB) sequence, and considers whether mesa-optimization will occur in HLMI at all, whether pseudo-alignment occurs and is dangerous, and whether the mesa-optimizer will be deployed or "break out" of a controlled environment. Some uncertainties we have identified relate to the nature of HLMI, connecting back to [Paths to HLMI](https://www.alignmentforum.org/s/aERZoriyHfCqvWkzg/p/amK9EqxALJXyd9Rb2) and to analogies with other domains. Other uncertainties relate to the training task, e.g. how generic the task is, and whether the input data is large and rich enough to incentivise modeling the objective. Other uncertainties are broadly related to inductive bias, e.g. whether the base optimizer tends to produce mesa-optimizers with harmful objectives.
We are interested in any feedback you might have, including how this post affected your understanding of the topic and the uncertainties involved, your opinions about the uncertainties, and important points that our model may not capture.
The next post in this series will look at the effects of AI Safety research agendas.
[1] We define HLMI as machines that are capable, either individually or collectively, of performing almost all economically-relevant information-processing tasks that are performed by humans, or quickly (relative to humans) learning to perform such tasks. We are using the term “high-level machine intelligence” here instead of the related terms “human-level machine intelligence”, “artificial general intelligence”, or “transformative AI”, since these other terms are often seen as baking in assumptions about either the nature of intelligence or advanced AI that are not universally accepted.
[2] Parts of this argument are more spelled out in the [FLI podcast with Evan Hubinger](https://futureoflife.org/2020/07/01/evan-hubinger-on-inner-alignment-outer-alignment-and-proposals-for-building-safe-advanced-ai/) (search the transcript for "which is deception versus corrigibility").
Acknowledgements
================
Thanks to the rest of the [MTAIR Project team](https://www.alignmentforum.org/s/aERZoriyHfCqvWkzg/p/qnA6paRwMky3Q6ktk#Acknowledgements), as well as the following individuals, for valuable feedback and discussion that contributed to this post: Evan Hubinger, Chris van Merwijk, and Rohin Shah. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.