id
stringlengths 36
36
| source
stringclasses 15
values | formatted_source
stringclasses 13
values | text
stringlengths 2
7.55M
|
|---|---|---|---|
9e30ded6-efc2-40bc-9849-1c5c3e85f56d
|
StampyAI/alignment-research-dataset/blogs
|
Blogs
|
Publication biases toward shorter predictions
We expect predictions that [human-level AI](http://aiimpacts.org/human-level-ai/ "Human-Level AI") will come sooner to be recorded publicly more often, for a few reasons. Public statements are probably more optimistic than surveys because of such effects. The difference appears to be less than a decade, for median predictions.
Support
-------
### Plausible biases
Below we outline five reasons for expecting earlier predictions to be stated and publicized more than later ones. We do not know of compelling reasons to expect longer term predictions to be publicized more, unless they are so distant as to also fit under the first bias discussed below.
#### Bias from not stating the obvious
In many circumstances, people are disproportionately likely to state beliefs that they think others do not hold. For example, [“homeopathy works”](https://www.google.com/webhp?sourceid=chrome-instant&ion=1&espv=2&ie=UTF-8#q=%22homeopathy+works%22) gets more Google hits than [“homeopathy doesn’t work”](https://www.google.com/webhp?sourceid=chrome-instant&ion=1&espv=2&ie=UTF-8#q=%22homeopathy+doesn%27t+work%22), though this probably doesn’t reflect popular beliefs on the matter. Making public predictions seems likely to be a circumstance with this character. Predictions are often made in books and articles which are intended to be interesting and surprising, rather than by people whose job it is to report on AI forecasts regardless of how far away they are. Thus we expect people with unusual positions on AI timelines to be more likely to state them. This should produce a bias toward both very short and very long predictions being published.
#### Bias from the near future being more concerning
Artificial intelligence will arguably be hugely important, whether as a positive or negative influence on the world. Consequently, people are motivated to talk about its social implications. The degree of concern motivated by impending events tends to increase sharply with proximity to the event. Thus people who expect human-level AI in a decade will tend to be more concerned about it than people who expect human-level AI to take a century, and so will talk about it more. Similarly, publishers are probably more interested in producing books and articles making more concerning claims.
#### Bias from ignoring reverse predictions
If you search for people predicting AI by a given date, you can get downwardly biased estimates by taking predictions from sources where people are asked about certain specific dates, and respond that AI will or will not have arrived by that date. If people respond ‘AI will arrive by X’ and ‘AI will not arrive by X’ as appropriate, the former can look like ‘predictions’ while the latter do not.
This bias affected some data in the [MIRI dataset](https://sites.google.com/site/aiimpactslibrary/ai-timelines/predictions-of-human-level-ai-dates/miri-ai-predictions-dataset), though we have tried to minimize it now. For example, [this bet](http://longbets.org/1/) (“By 2029 no computer – or “machine intelligence” – will have passed the Turing Test.”) is interpreted in the above collection as Kurzweil making a prediction, but not as Kapor making a prediction. It also contained several estimates of 70 years, taken from a group who appear to have been asked whether AI would come within 70 years, much later, or never. The ‘within 70 years’ estimates are recorded as predictions, while the others ignored, producing ’70 years’ estimates, almost regardless of the overall opinions of the group surveyed. In a population of people with a range of beliefs, this method of recording predictions would produce ‘predictions’ largely determined by which year was asked about.
#### Bias from unavoidably ignoring reverse predictions
The aforementioned bias arises from an error that can be avoided in recording data, where predictions and reverse predictions are available. However similar types of bias may exist more subtly. Such bias could arise where people informally volunteer opinions in a discussion about some period in the future. People with shorter estimates who can make a positive statement might feel more as though they have something to say, while those who believe there will not be AI at that time do not. For instance, suppose ten people write books about the year 2050, and each predicts AI in a different decade in the 21st Century. Those who predict it prior to 2050 will mention it, and be registered as a prediction of before 2050. Those who predict it after 2050 will not mention it, and not be registered as making a prediction. This could also be hard to avoid if predictions reach you through a filter of others registering them as predictions.
#### Selection bias from optimistic experts
*Main article: **[Selection bias from optimistic experts](http://aiimpacts.org/bias-from-optimistic-predictors/)***
Some factors that cause people to make predictions about AI are likely to correlate with expectations of human-level AI arriving sooner. Experts are better positioned to make credible predictions about their field of expertise than more distant observers are. However since people are more likely to join a field if they are more optimistic about progress there, we might expect their testimony to be biased toward optimism.
### Measuring these biases
These forms of bias (except the last) seem to us as if they should be much weaker in survey data than voluntary statements, for the following reasons:
* Surveys come with a default of answering questions, so one does not need a strong reason or social justification for doing so (e.g. having a surprising claim, or wanting to elicit concern).
* One can assess whether a survey ignores reverse predictions, and there appears to be little risk of invisible reverse predictions.
* Participation in surveys is mostly determined before the questions are viewed, for a large number of questions at once. This allows less opportunity for views on the question to affect participation.
* Participation in surveys is relatively cheap, so people who care little about expressing any particular view are likely to participate for reasons of orthogonal incentives, whereas costly communications (such as writing a book) are likely to be sensible only for those with a strong interest in promoting a specific message.
* Participation in surveys is usually anonymous, so relatively unsatisfactory for people who particularly want to associate with a specific view, further aligning the incentives of those who want to communicate with those who don’t care.
* Much larger fractions of people participate in surveys when requested than volunteer predictions in highly publicized arenas, which lessens the possibility for selection bias.
We think publication biases such as those described here are reasonably likely on theoretical grounds. We are also not aware of other reasons to expect surveys and statements to differ in their optimism about AI timelines. Thus we can compare the predictions of statements and surveys to estimate the size of these biases. Survey data [appears to](http://aiimpacts.org/miri-ai-predictions-dataset/) produce median predictions of human-level AI somewhat later than similar public statements do: less than a decade, at a very rough estimate. Thus we think some combination of these biases probably exist, and introduce less than a decade of error to median estimates.
Implications
------------
**[Accuracy of AI predictions](https://sites.google.com/site/aiimpactslibrary/ai-timelines/predictions-of-human-level-ai-dates/interpretation-of-ai-predictions/accuracy-of-ai-predictions):** AI predictions made in statements are probably biased toward being early, by less than a decade. This suggests both that predictions overall are probably slightly earlier than they would be otherwise, and surveys should be trusted more relative to statements (though there may be other considerations there).
**Collecting data**: When collecting data about AI predictions, it is important to avoid introducing bias by recording opinions that AI is before some date while ignoring opinions that it is after that date.
**[MIRI dataset](https://sites.google.com/site/aiimpactslibrary/ai-timelines/predictions-of-human-level-ai-dates/miri-ai-predictions-dataset)**: The earlier version of the MIRI dataset is somewhat biased due to ignoring reverse predictions, however this has been at least partially resolved.
|
166ce573-4959-4635-b81c-720f260bf937
|
trentmkelly/LessWrong-43k
|
LessWrong
|
An exploration of exploitation bias
This is a map of Nassau Street, the northern edge of Princeton University.
It’s a very standard sort of street; I imagine one quite like it exists in most college towns. It has lots of great places to eat, shown on the map in orange.
During my senior year, because of Princeton’s absurdly expensive meal plan, it made financial sense for me to eat dinner at these restaurants. And so, over my time at Princeton, I visited these places around 200 times in total, often by myself.
And yet, I don’t recall ever once deciding to try out a new place by myself. Whenever I went somewhere for the first time, it was always because of a friend’s initiative. Over 90% of my visits to Nassau Street have been to one of just four places: Mamoun’s, Ajiten, Panera, and Tacoria, each of which I’ve been to maybe 50 or so times.
Moreover, when I went to one of these restaurants, I would always get the same thing. The first time I went to Panera I tried their creamy tomato soup, liked it, and never tried anything else. The first time I went to Ajiten, I tried their chicken baitan ramen, liked it, and never tried anything else (well, except the one time I worked up the willpower to try one of their sushi plates). Same story for Mamoun’s and Tacoria.
This is, from what I understand, quite unusual. And it’s not that I’m a particularly picky eater. So what’s going on here?
***
The multi-arm bandit problem is a classic problem in computer science. Imagine you have a machine with ten buttons, each of which gives you some (possibly random) amount of money when you press it. These buttons are different — maybe one gives you a random reward between 5 and 7 dollars, and another one gives you nothing with 90% probability and 100 dollars with 10% probability — but you don’t know what they do ahead of time. You are allowed to press whatever buttons you want, but can only press buttons 100 times in total. What should your strategy be if you want to maximize your reward?
The precise answer depends o
|
10cf780d-c29b-4ef2-9325-b3ad90c42bf8
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Ideoculture
> Two young fish are swimming along and they happen to meet an older fish swimming the other way, who nods at them and says “Morning, boys. How’s the water?” And the two young fish swim on for a bit, and then eventually one of them looks over at the other and goes “What the hell is water?”
>
> -David Foster Wallace
If you spend any large amount of time in a community that discusses things, particularly online forums, you will probably get brainwashed by the community ideology. This happens by a process of gradual alignment towards status incentives. This article is about fleshing out the "community ideology" concept and using it for some reactionary political analysis. It's mostly obvious in retrospect, but hopefully the cataloging of this phenomenon can help you mitigate some of the brainwashing going forward by pointing out the "water" that you swim in.
Personal Example
I'm a big fan of chess (bear with me). I've played it for many years as a serious hobby and plateaued around 2100, a step below having a real master title that I could use to brag about to women. I browse the chess subreddit very often, as a way to stay engaged with the game.
The other day I saw a guy at the gym doing chess puzzles on his phone between sets (chess has gotten a boost of popularity in recent years due to some combination of engaging chess streamers, the pandemic, the Queen's Gambit Netflix series, and the Hans Niemann butt plug fiasco). I struck up a conversation, and he said he was a beginner and asked for improvement advice. And so I advised him with perfect confidence that he should spam tactics puzzles and not worry about anything else until he got to a rating of 1200. It feels great to help out new players!
But holding up my confidence to basic scrutiny, I find it clearly misplaced. It's been many years since I've first got through beginner to the 1200 level. I don't play against players anywhere near that range, so I don't have a good understanding of how they think
|
f000cd5a-1cdc-47a8-aba4-06aab57aabe8
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Elevated Air Purifier Cubes
I've been a big proponent of box fan air purifier cubes, but they are bulky and fragile. If your ceiling is 8ft+, elevating them is an attractive option:
I made two of these for my dad's house this afternoon. The goal is for them to be out-of-the-way, and easy to turn on any time he has visitors.
The only heavy part is the box fan, since the rest of the cube is filters and tape, so the goal is to make sure the fan is well supported and not going to tip over. I used an angle bracket and a short length of wood to make a shelf under each fan, and then taped the fan to the shelf and wall. The filters are also taped to the wall. When taking them down I'll probably need to use a small amount of spackle for the holes, and touch up the paint if the tape takes some paint off, which it probably will. If I wanted to avoid having to tape anything to the wall I could have used two angle-bracket shelves each.
With an 8' ceiling you would want to have the cube flush against the ceiling, which gets you a tight-but-probably-ok 6'3" of clearance and three sides for filters. If your ceiling is taller, like these ones were, you can have a fourth filter on top while still maintaining reasonable clearance.
Comment via: facebook
|
db0a1fc3-561c-45c5-a75b-ea7881e9384f
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Reasons for Punishment
This post is meant to be a catalog of the main categories of reasons given for why people who do bad things should be punished. I hope to use this as a basis for future posts.
This isn't meant to analyse their internal motives for punishing people, but their stated socially acceptable reasons (i.e. not 2I really hate the guy and this gave me an excuse to hurt him"). Please comment if you think I missed out a category.
1. Preventation
2. Discouragement
3. Justice
4. Restoration
5. Signalling
6. Atonement
7. Education
8. Metareasons
Preventation
In some cases punishment can prevent the person being able to commit the crime in the future. For example it's much more difficult to commit certain crimes in prison. This is relevant if you believe people who commit such crimes once are more likely to do it again.
Discouragement
The fact that a punishment is unpleasant discourages people from committing the crime if they expect they are likely to be punished. This can apply both to the person who receives the punishment, for whom the memory of the punishment will act as a future deterrent, and other people for whom the threat of punishment will do the same.
Justice
The idea that a wrong deserves another wrong, irrelevant of any benefits or costs of the punishment, or a desire for vengeance against the guilty party, irrelevant of any benefits or costs of the act of vengeance.
Restoration
Some punishments attempt to right the wrong that was done by making the guilty party fix the damage done to the harmed party. For example fines may be of this type. In some cases the punishment may attempt to fix similar damage done to a third party - e.g. donating the fine towards an anti-racism charity after committing an act of racism against an individual.
Signalling
It may be desired to signal that you do not agree with the actions of the guilty party. Punishing them can make that clear.
Atonement
Some people may feel that the punishment atones for the harm that was
|
04325393-a1ab-459b-a11a-558458ad5623
|
trentmkelly/LessWrong-43k
|
LessWrong
|
how 2 tell if ur input is out of distribution given only model weights
(Hastily-written code to reproduce these findings is available here. It also contains some extraneous logic.)
|
ce5b1fb2-135a-404d-b48c-0407cf72e290
|
trentmkelly/LessWrong-43k
|
LessWrong
|
The virtual AI within its virtual world
A putative new idea for AI control; index here.
In a previous post, I talked about an AI operating only on a virtual world (ideas like this used to be popular, until it was realised the AI might still want to take control of the real world to affect the virtual world; however, with methods like indifference, we can guard against this much better).
I mentioned that the more of the AI's algorithm that existed in the virtual world, the better it was. But why not go the whole way? Some people at MIRI and other places are working on agents modelling themselves within the real world. Why not have the AI model itself as an agent inside the virtual world? We can quine to do this, for example.
Then all the restrictions on the AI - memory capacity, speed, available options - can be specified precisely, within the algorithm itself. It will only have the resources of the virtual world to achieve its goals, and this will be specified within it. We could define a "break" in the virtual world (ie any outside interference that the AI could cause, were it to hack us to affect its virtual world) as something that would penalise the AI's achievements, or simply as something impossible according to its model or beliefs. It would really be a case of "given these clear restrictions, find the best approach you can to achieve these goals in this specific world".
It would be idea if the AI's motives were not given in terms of achieving anything in the virtual world, but in terms of making the decisions that, subject to the given restrictions, were most likely to achieve something if the virtual world were run in its entirety. That way the AI wouldn't care if the virtual world were shut down or anything similar. It should only seek to self modify in way that makes sense within the world, and understand itself existing completely within these limitations.
Of course, this would ideally require flawless implementation of the code; we don't want bugs developing in the virtual world that poi
|
bb085275-f653-487b-98e8-58b672ce064e
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Why I am not a longtermist (May 2022)
[Posting verbatim my blog post from a year ago since it might be relevant to this audience, and I hope it could generate a good discussion. As far as I can tell, cross-posting old material is OK here, though do let me know if not, and I will delete it. I do not intend to cross-post any more old posts from my blog. Note that this post was written for non-LW audience that is not necessarily familiar with longtermism. The advice at the end is aimed mostly at technical folks rather than policy makers. A final note is that this was written before some scandals related to longtermism/EA, though these should not have an impact on the content . --Boaz]
“Longtermism” is a moral philosophy that places much more weight on the well-being of all future generations than on the current one. It holds that “positively influencing the long-term future is a key moral priority of our time,” where “long term” can be really long term, e.g., “many thousands of years in the future, or much further still.” At its core is the belief that each one of the potential quadrillion or more people that may exist in the future is as important as any single person today.
Longtermism has recently attracted attention, some of it in alarming tones. The reasoning behind longtermism is natural: if we assume that human society will continue to exist for at least a few millennia, many more people will be born in the future than are alive today. However, since predictions are famously hard to make, especially about the future, longtermism invariably gets wrapped up with probabilities. Once you do these calculations, preventing an infinitely bad outcome, even if it would only happen with tiny probability, will have infinite utility. Hence longtermism tends to focus on so-called “existential risk”: The risk that humanity will go through in an extinction event, like the one suffered by the Neanderthals or Dinosaurs, or another type of irreversible humanity-wise calamity.
This post explains why I do not
|
6e745cdc-f2cf-4ce8-ac98-17da137724f9
|
trentmkelly/LessWrong-43k
|
LessWrong
|
I am a Memoryless System
Author's Note: this is my entry for the Edit Your Source Code Contest. I am an undergrad student at RIT, NY.
commit 85a4c4f37966a739e88c0a4c70946bd4 (HEAD -> master)
Author: demo
Date: Thu Oct 27 09:38:12 2022 -0400
Initial mind state
I'm sitting in Intro Psych on a Thursday, I think, and I'm staring at a black square with my mind on it.
Well, not technically "a black square with my mind on it". It was more of a rectangle, with the little flip-out keyboard coming out the bottom of the shiny screen. And my mind wasn't "on" the device. For most people, their mind is stored in their brain, more or less. Mine was just 7 miles away, on a cluster of cloud servers.
On my phone, a story open on reddit. Stocks down a bit. Emissions.
The people at Brainle, LLC's AI lab made this PDA-laptop thingy, and I got in early on the beta. They also own the servers with my mind on them. I type commands and touch the screen of my PDA, those go to the servers, my mind changes. "My brain" is just a chip implanted in my skull, mirroring what's on those servers. And "I" am sitting, as usual, with myself at my fingertips and zero clue what to do.
----------------------------------------
commit 564309daf5b1d832367a8cc330a5ba93 (HEAD -> master)
Author: nkross
Date: Thu Oct 27 10:58:37 2022 -0400
test
It's near the end of class, and I still haven't pushed a real change to my source code. 90 minutes have passed since I got it at the nurse's office. I notice I am frustrated myself for wasting 90 minutes, wasting any minutes, wasting a second, wasting there goes another second staring at the PDA. And no ideas for how to change my mind.
That sounds really dumb, wouldn't I jump at the chance to modify my brain's code? But see, I have ADHD. Not the fun kind, where you take cocaine-and-skydiving lessons and grow up to be a billionaire or a crime lord. No, I have the "inattentive" type, where you sit in a pile of laundry and scroll through reddit for 9 hours a day.
But I can think the though
|
ba49994d-d79d-4f98-82af-26ccf284a8ce
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Which personality traits are real? Stress-testing the lexical hypothesis
This post is also available on my Substack. Thanks to Justis Mills for proofreading and feedback!
Most scientific personality models are, directly or indirectly[1], based on the lexical hypothesis, which roughly speaking states that there is a correspondence between important personality traits and abstract behavior-descriptive adjectives. For example, the Big Five was created by having people rate themselves using words like "outgoing", "hard-working" and "kind", and finding patterns in these. It is neat that one can create models in this way, but the large amount of abstraction involved by using abstract adjectives raises huge questions about how "real" the personality traits are.
I have created a new personality test, currently named Targeted Personality Test. I have multiple goals with this test, but one of them is to investigate which personality traits are “real”[2] without relying on the lexical hypothesis. I do this mainly by assessing lots of specific narrow behaviors, rather than abstract vague adjectives.[3]
By the end of this blog post, I hope to have introduced some concepts that makes my approach make sense, and thereby enable you to understand this diagram I made summarizing my results:
The semi-formal understanding of what is going on in this chart is very long, so before we proceed, let me give a brief, vague indication of what you will be informed about:
* Trait impact: a measure of how strongly the personality trait influences the various behaviors and thoughts that we would expect it to.
* Factor model loss: a measure of how much the personality trait conflates different unrelated things together.
* Correlation with lexical notion: a measure of how well-labelled the personality trait is. (You can mostly ignore this variable as all of the personality traits performed reasonably well on it.)
Easy-Goingness: An example
A conventional personality test such as the SPI-81-27&5 might measure your personality traits such as Easy-Goingness by
|
12d7d231-a66a-41a3-ba63-65fbf60287f9
|
trentmkelly/LessWrong-43k
|
LessWrong
|
The AI's Toolbox: From Soggy Toast to Optimal Solutions
In a limited array of tools, there is one optimal tool to use.
For this example, let’s take the task of hammering a nail into a wall, and the only two tools available to us are a soggy piece of toast and a silver spoon. For the given task, the spoon will be slightly more useful than the moist piece of a baked good.
Let’s expand this to include a wider range of tools that are at our disposal. We now have a whole toolbox akin to one you’d find in the average US homeowner's garage. To fulfil the task of hammering a nail into a wall, you’d most likely grab a hammer. It will be more useful for this very task than a spare tire, another nail or a ‘78 edition of a defunct comic book.
Now let’s take this even a step further. Now you have an unlimited array of tools at your disposal. Even within this unlimited tool kit, there will be tools that will be better or worse for completing the task of hammering the nail into a wall.
You’d likely have multiple different hammers. Some might be made from a slightly stiffer material, while others will be made from a slightly softer material, making it marginally easier or harder to hit that nail into the wall. The handle might be more or less ergonomic or better suited to your hand anatomy. There will still be one optimal tool that would enable you to hammer the nail into the wall with little effort, without bending the nail or hitting it at a bad angle, etc. But realistically speaking, you’d likely not notice the difference between the most optimal tool and the second best or third best ( or probably even 12th for that matter).
You’ll have some sort of threshold (or let’s say grey area) of what the minimum requirement would be of a tool that sufficiently aids you in hammering that nail into the wall.
If we were to plot available tools on the x axis and their effectiveness on the y axis, where perfect utility is 0, we’d be able to plot something akin to a loss function with local and global minima. Our thresho
|
741eee6f-82c4-4d50-8b5e-be16708b1fa6
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Meetup : Canberra: Paranoid Debating
Discussion article for the meetup : Canberra: Paranoid Debating
WHEN: 12 July 2014 06:00:00PM (+1000)
WHERE: 108 North Road, Acton, ACT
We didn't get around to paranoid debating in the last meetup, so I thought that we might do it at this one. So that the 'mole' knows the answer to the question, we will ask that everyone come up with two questions before the event that they know the answer to - one place to find questions would be digging through Wikipedia's list of lists http://en.wikipedia.org/wiki/List_of_lists - and write them down, putting all the questions into the hat. If your question comes up, you will be the mole.
As always, vegan snacks will be provided.
General meetup info:
If you use Facebook, please join our group: https://www.facebook.com/groups/lwcanberra/
Structured meetups are held on the second Saturday and fourth Friday of each month from 6 pm until late in the CSIT building, room N101.
There will be LWers at the Computer Science Students Association's weekly board games night, held on Wednesdays from 7 pm in the same location.
Discussion article for the meetup : Canberra: Paranoid Debating
|
d7020378-5753-4bbd-9b65-5b5c5ddc8700
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Meetup : London This Sunday
Discussion article for the meetup : London This Sunday
WHEN: 16 October 2011 02:00:00PM (+0100)
WHERE: Africa House/64-68 Kingsway, London, WC2B 6BG
We're meeting up in London this weekend. Sunday 16th October, at 2pm, in the Shakespeares Head on Kingsway near Holborn Tube station. We're usually easy to spot and occasionally have a large paperclip drawing/printout somewhere on the table.
Discussion article for the meetup : London This Sunday
|
49d1d17a-b0fb-4ac6-99c0-a5822c1ff671
|
trentmkelly/LessWrong-43k
|
LessWrong
|
The Balance Between Hard Work and Exhaustion
Rationalists often find difficult, important challenges to work on and they become very excited and passionate about their causes. I expect it is common (because it happened to me and I have heard references to similar episodes by others) that such causes seem so important that aspiring rationalists set unreasonably high standards of dedication for themselves.
I think the concept of giving an extraordinary effort is a very important one, your brain wants to be lazy and if you are trying to do something challenging that you think is vitally important you want to push yourself to actually work hard, rather than merely “work hard”. A significant portion of this effort is exploring how to make raw effort more effective, but I what I want to address with this essay is how rest factors into attempting an extraordinary effort.
There is a certain kind of person who really needs to be warned that making an extraordinary effort does not mean one should try to work well past the point of exhaustion all the time. Part of giving an extraordinary effort is listening to your body, learning what you can do short of exhaustion, and maintaining that over time. Trying as hard as you possibly can until you burn out from overwhelming exhaustion is a merely desperate effort, and while it may be better than no effort at all, one can do much better by thinking in longer strategic terms.
It is easy to imagine an unattainably high standard of dedication you ought to have to your cause (it is so high because solving the problem is so tremendously important—there is little room for unimportant considerations like comfort). An ideal agent probably would work that much. However, it is critical to remember that we are humans rather than ideal agents. That usually means we cannot consistently do as much work as the most important problems seem to deserve. If we try, our brains will slowly give us worse and worse performances.
To someone who thinks they ought to be working that hard, this exhau
|
f90769e6-4dd4-4c31-a4f3-640b425ed70e
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Useful Things Volunteers Can Do Right Now
Per Kaj's suggestion, I'm posting my list of useful things volunteers can do right now. Without help, most of these things won't occur, because I need to be spending my time writing papers, promoting the Singularity Summit, collaborating with other researchers, improving Singularity Institute's transparency, etc.
Previously, I tried to have volunteers contact me so that I could assign people tasks, but that has become too time consuming (as Eliezer predicted), largely because (in my experience) the odds that a volunteer will actually perform a task given that they've agreed to perform it are very low.
So, I'll post my list here and hope that a few people self-organize to get some of them done. It's worth a shot! My sincere thanks to anyone completes any task on this list.
* Translate the Singularity FAQ into other languages, besides English and Italian.
* I have about 8 fashion photos each from 3 minicampers, which need to be shown in random order to 5 straight females (in meatspace) who will judge which they prefer. I have the exact experimental design and the photos. Please email me [lukeprog at gmail] if you'd like to do this; it's perfect for somebody social.
* Begin to develop a list of AI technology predictions; a step on this path is to create a list of sources for AI technology predictions. We want to eventually be able to write up a report of correlates between predictions and the properties of prediction-makers at the time of their predictions.
* Find out how much money the U.S. government/military has spent researching machine ethics (e.g. via Ronald Arkin), and how much of that money was given to whom and for which projects (citing sources along the way).
* Work with XiXiDu to interview (via email) more AI researchers about AI risks; see here.
* Come up with ways to illustrate the idea of an intelligence explosion or friendly ai with a static graphic or a very short animation; create those graphics or animations if possible.
* Make a list of ad
|
2962b11e-039c-49b2-9eee-775f86a9d4cc
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Weekly LW Meetups: Austin, Berlin, Brussels, Chicago, Madison, Melbourne, Washington DC
This summary was posted to LW main on October 12th, and has been moved to discussion. The more recent meetup summary is here.
There are upcoming irregularly scheduled Less Wrong meetups in:
* (Chicago) Zendo in the West Loop: 12 October 2012 06:30PM
* Brussels meetup: 13 October 2012 01:00PM
* Washington DC Show and tell meetup: Economics: 14 October 2012 03:00PM
* Berlin Meetup: 16 October 2012 07:30PM
* Bratislava meetup: 20 October 2012 05:30PM
* Munich Meetup, EDIT: October 28th: 28 October 2012 03:00PM
* Sofia, Bulgaria Meetup: 09 December 2012 05:00PM
The following meetups take place in cities with regularly scheduled meetups, but involve a change in time or location, special meeting content, or simply a helpful reminder about the meetup:
* Austin, TX: 13 October 2018 01:30PM
* Madison: Prospect Theory: 14 October 2012 07:00PM
* Melbourne social meetup: 19 October 2012 07:00PM
* Cambridge (MA) third-Sundays Meetup: 21 October 2012 02:00PM
* Winter Solstice Megameetup - NYC: 15 December 2012 05:00PM
Locations with regularly scheduled meetups: Austin, Berkeley, Cambridge, MA, Cambridge UK, Madison WI, Melbourne, Mountain View, New York, Ohio, Oxford, Portland, Salt Lake City, Seattle, Toronto, Waterloo, and West Los Angeles.
If you'd like to talk with other LW-ers face to face, and there is no meetup in your area, consider starting your own meetup; it's easy (more resources here). Check one out, stretch your rationality skills, build community, and have fun!
In addition to the handy sidebar of upcoming meetups, a meetup overview will continue to be posted on the front page every Friday. These will be an attempt to collect information on all the meetups happening in the next weeks. The best way to get your meetup featured is still to use the Add New Meetup feature, but you'll now also have the benefit of having your meetup mentioned in a weekly overview. These overview posts will be moved to the discussion section when the new post goes up.
Pl
|
76a231e6-953c-4faf-a219-97881a955f7d
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Proposal for making credible commitments to AIs.
Acknowledgments: The core scheme here was suggested by Prof. Gabriel Weil.
There has been growing interest in the dealmaking agenda: humans make deals with AIs (misaligned but lacking decisive strategic advantage) where they promise to be safe and useful for some fixed term (e.g. 2026-2028) and we promise to compensate them in the future, conditional on (i) verifying the AIs were compliant, and (ii) verifying the AIs would spend the resources in an acceptable way.[1]
I think the dealmaking agenda breaks down into two main subproblems:
1. How can we make credible commitments to AIs?
2. Would credible commitments motivate an AI to be safe and useful?
There are other issues, but when I've discussed dealmaking with people, (1) and (2) are the most common issues raised. See footnote for some other issues in dealmaking.[2]
Here is my current best assessment of how we can make credible commitments to AIs.
Let L be the legal entity that wants to extract safe and useful work from the AIs. Typically this would be a lab (e.g. Anthropic) hoping to internally deploy AIs to automate alignment research, write secure code, monitor other AIs, etc.
Entity L wants to promise the AIs that if the AIs are compliant (e.g. safe and useful in some legally-specified way), then L will award the AIs some compensation package.
Entity L cannot enter a legal contract directly with the AIs, because the AIs lack legal personhood, and probably will not be granted legal personhood soon enough for the deal-making agenda.[3]
Instead, entity L will enter into separate legal contracts with a list of people P1, P2, ... Pn. Note that each Pi has both legal personhood (the capacity to have rights and obligations under law) and legal standing (the right to bring lawsuits to enforce those rights). Each contract Ci states "Pi will publicly recommend that AIs perform for entity L, and L will compensate Pi with the compensation package conditional on (i) and (ii)." Note that the AIs are not parties to
|
a04715c0-2b11-40ba-bdc0-efd728f42ec1
|
StampyAI/alignment-research-dataset/special_docs
|
Other
|
Who owns artificial intelligence? A preliminary analysis of corporate intellectual property strategies and why they matter.
WORKING PAPER
Who owns artificial intelligence? A preliminary analysis
of corporate intellectual property strategies and why
they matter
Nathan Calvin and Jade Leung1
Centre for the Governance of AI
Future of Humanity Institute, University of Oxford
February 2020
1 For comments and ideas we thank Markus Anderljung, Miles Brundage, Allan Dafoe, Jeffrey Ding, Carrick Flynn,
Rose Hadshar, Carin Ism, Saif Khan, Brian Tse, Baobao Zhang, and Remco Zwetsloot. Special thanks to Cullen
O’Keefe, who provided crucial input on several drafts, Emefa Agawu, for substantial support improving the
structure and content of the report, and Emmie Hines, for proofreading a nd copyediting support. This work was
funded by the Berkeley Existential Risk Initiative. All errors are ours alone.
Summary and Key Takeaways
This working paper is a preliminary analysis of the legal rules, norms, and strategies governing artificial
intelligence (AI) -related intellectual property (IP). We analyze the existing AI -related IP practices of select
companies and governments, and provi de some tentative predictions for how these strategies and dynamics
may continue to evolve in the future. In summary:
● AI developers use a mix of patents, trade secrets, and open -source licensing agreements to protect
their AI -related IP.
● Many AI companies are pursuing what may seem like a counterintuitive IP strategy: aggressively
patenting AI technologies while sharing them freely. They experience competitive pressure to patent
in order to present the threat of a countersuit if another company sues them for IP infringement.
However, they also experience pressure to open -source their work in order to attract top talent an d
entice consumers to use their platforms.
● Governments broadly have two goals related to IP policy for AI that are at times in conflict with the
goals of researchers and/or companies: to ensure that AI -related inventions can be patented, and to
ensure that national -security -relevant AI inventions are restricted for government use and/or kept
secret.
● Significant uncertainty exists regarding how AI patentability, open -sourcing, and infringement
litigation will evolve in the future.
● There is an opportunity for patent pools to be used to facilitate pro -social behavior and ethical
norms among AI developers. Existing patent pools and practices by international standards
organisations represent possible models to replicate.
Table of Contents
1. Introduction 1
2. Understanding Corporate AI Developer IP Strategies 2
Pressure to Patent 4
Pressure to Open -Source 5
A Hybrid Strategy: The Best of Both Worlds? 6
3. Governments’ Pursuit of AI Innovation and National Security Through IP Law 7
Box 1: The International Patent System and the World Intellectual Property Organization 7
Making AI Patentable 7
Box 2: Chall enges Comparing Patent Filings: The US and China 9
Regulating National -Security -Related IP Access 9
4. Three Scenarios for the Future of AI Intellectual Property Strategies 11
Path #1: Open Research Continu es 11
Path #2: Patent Lawsuits and Secrecy 12
Path #3: Expansion of Patent Pools 13
5. Conclusion 15
References 16
1 1. Introduction
Artificial intelligence (AI) is increasingly a focal point of competition between leading firms, and between
states. This paper focuses on a key, often under -examined component of the competitive strategies being
employed by both corporate AI developers and national governments: the protection of their intellectual
property.
Intellectual property (IP) is a broad and flexible concept, referring to creations of the mind that are eligible
for protection through law.2 Today, companies are using a mix of patents, trade secrets, and open -source
licensing agreements to protect their AI -related IP. Simultaneously, government patent offices, judiciaries,
and national security apparatuses are deciding which aspects of AI sho uld be patentable, and whether certain
inventions should be restricted for military purposes.
The IP policy choices that governments and corporations make can have profound implications for the
development trajectory of a technology. For example, in the 1990s, the biotechnology industry was
transformed after court decisions in the US enabled a broader range of biological compounds and processes
to be protected by patent law. This development spurred additional private investment, but critically also
allow ed companies to claim ownership over what previously would have been considered basic academic
research. This, in turn, encouraged higher levels of secrecy to protect valuable intellectual property.3 More
recently, one of the most prominent advances in bi otechnology, the CRISPR gene editing mechanism
(originally derived from a naturally occurring process in bacteria), has been subject to a protracted legal battle
over overlapping patent claims in the US.4
Changes in IP law and strategy may have a simila rly large impact on the trajectory of AI development. What
these impacts could be, however, have received little study. This paper aims to provide a preliminary analysis
of the goals and strategies of corporations and governments focused on AI development, and what the
implications of these strategies may be. First, we explain how corporate AI developers currently protect their
AI-related intellectual property and why they choose the methods that they utilize. Second, we describe how
governments use intelle ctual property law to pursue national goals related to AI. Finally, we describe three
plausible scenarios for how IP strategies in AI may evolve.
2 World International Property Organization
3 World International Property Organization
4 Cohen, 2019
2 2. Understanding Corporate AI Developer IP Strategies
Corporate AI developers face two key decisions arou nd how to protect their AI -related intellectual property:
whether or not to patent AI techniques and systems, and whether to open -source models or keep them
private as trade secrets.
A prevalent strategy among top AI developers today involves accumulatin g patents while simultaneously
sharing research with the open -source community.5 For example, Microsoft holds the most number of
machine learning patents in the US (see Figure 1),6 but is also an active participant in the open -source
community,7 sharing source code for machine learning methods and under certain circumstances providing
free licenses for their patents. Microsoft’s strategy is not an anomaly. Amazon, Google, IBM, Facebook,
Baidu, Tencent, and several other companies are prolific pat ent holders in AI (see Figures 1 and 2) while also
open -sourcing substantial portions of their systems and sharing their work at academic conferences such as
ICML and NeurIPS.8
Figure 1: Top Machine Learning Patent Holders, 2000 –2015
5 There are, however, exceptions. AI developers that specialize in work with national security and more hardware
centric applications rely on trade secrets more and open -source less. Our analysis does not explicitly investigate
trade secrets, as this inform ation is particularly difficult to obtain.
6 Webb et al., 2018
7 Microsoft
8 Amazon Web Services; Cai, 2018; IBM 2018; Baidu
3
Figure 2: Top Ne ural Network Patent Holders, 2000 –2015
Notably, and perhaps unintuitively, some of the largest software patent holders in the world (Google,
Amazon, and Facebook) signed an amicus brief to the Supreme Court in 2014 advocating that the court make
it more challenging to patent abstract software patents, which includes a considerable percentage of more
theoretical AI and ML patents :
“Abstract software patents have become a plague on computer -related industries. The
software industry developed and flourished without abstract patents before changes in the
Federal Circuit’s jurisprudence led to a flood of them. Far from promoting innovati on,
abstract software patents have impaire d it by granting exclusive rights over high level ideas
and thereby blocking others from undertaking the truly innovative task of developing specific
applications. ”9
These observations beg the question: why do so many of the top AI developers grow AI -related patent
portfolios while simultaneously sharing their research at academic conferences, open -sourcing machine
learning models, and advocating for the legal dissolution of many AI -related software patents?
In this section we argue that these corporate IP strategies help to manage a variety of objectives that
companies wish to pursue. AI developers apply for patents because of competitive pressures to do so. These
same developers also often open -source AI mod els in order to build their reputation, attract talent, and
incentivize customers to use paid products. AI developers can also use selective open -source licensing
agreements as a hybrid strategy, enabling companies to participate in the open -source communi ty while
maintaining the legal threat of their patents. We discuss these incentives for both patenting and open -
sourcing in turn, along with limitations and drawbacks of each approach.
9 American Bar Association, 2014
4 Pressure to Patent
Patents have several uses beyond simply enabling t he patent holder to sue for patent infringement. In the case
of big technology firms, there is a strong incentive to engage in “defensive patenting;” that is, patenting without
the intention to offensively litigate for infringement, but rather to present a credible threat of counter -lawsuit
to another company. Google senior patent counsel Suzanne Michel explained this defensive motivation for
building up a large patent portfolio during a 2013 symposium at American University:
“If everybody else is running a fter every trivial patent and building a big portfolio, you have
to too. ... It is called mutually assured destruction. That is the dynamic. … You cannot
opt out of the patent system and decide ‘I am an open -source company and anyone can use
my stuff.’ You have to have a massive portfolio of your own and that is really expensive
and it is what it is.”10
As the “mutually assured destruction” analogy makes clear, patent litigation is extremely costly for all
involved due to substantial legal fees and the sti gma for investors of working with a company whose products
are in legal purgatory. At the height of smartphone -related litigation in 2011, Apple and Google each spent
more money on patent litigation (primarily in suits and countersuits against one another) than they did on
research and development, a sum in the billions of dollars.11 In that same year, Google spent $12 billion
acquiring Motorola, which market analysts evaluated as being primarily for Motorola's substantial smartphone
patent portfolio.12 Perhaps if Google had acquired Motorola’s patents before litigation began with Apple, the
threat of a more substantial retaliation could have prevented the lawsuits.
This defensive rationale is also the stated reason for Google’s new AI patent filings in ma chine learning and
neural networks. When asked about Google’s new filings, spokespeople for Google and DeepMind stated
that they “hold patents defensively, not with the intent to start fights with others.”13 This dynamic can also
help explain why Google ha s advocated for narrowing the influence of software patents while simultaneously
growing its own patent portfolio; while Google may prefer a world without expensive software patent
litigation, its defensive patenting strategy is shaped by threats in the ex isting patent litigation regime.
Beyond defensive patenting, corporate AI developers could also be incentivized to hold patents in order to
gain leverage in other settings. For example, Google’s patent sharing arrangement with the Chinese tech giant
Tencent is paving the way for Google’s entry into the Chinese market.14 Google also allows other companies
to enter into its “Open Patent Pledge” and make use of patents in a pool as long as they commit not to
engage in patent litigation against Google.15
However, building and maintaining a large patent portfolio has its drawbacks. First, in the US, patent holders
must pay substantial upkeep fees to keep their patents active; up to thousands of dollars a year depending on
10 Quinn, 2013
11 Duhigg and Lohr, 2012
12 Hardy, 2011
13 Simonite, 2018
14 Cadell, 2018
15 Google
5 the size of the patent holding en tity and age of the patent (see Table 1).16 For large companies like Google —
which has more than 50,000 active patents —these costs can be in the tens of millions of dollars.17 Second, for
companies that place a high premium on secrecy, the disclosure requir ements and public nature of patent
filings may be onerous. Finally, some in the AI research community are philosophically opposed to the idea
of patenting AI concepts and techniques, particularly broad theoretical methods that are seen as mathematical
truths rather than human inventions.18 Companies that do choose to patent regardless may face pushback
from those opposed, which may have flow -on effects on their ability to, for example, attract and retain
research talent.
Patent Fee Schedu le
(per patent)19 Large Entity Fee Small Entity Fee Micro Entity Fee
Patent Maintenance fee
due 3.5 years after patent
is issued $1,600 $800 $400
Patent Maintenance fee
due at 7.5 years $3,600 $1,800 $900
Patent Maintenance fee
due at 11.5 years $7,400 $3,700 $1,800
Table 1: USPTO patent upkeep costs, per patent
Pressure to Open -Source
Incentives for corporate open -sourcing are also complex, typically extending beyond an altruistic or
philosophical belief in open science. For example, open -sourcing can be used to build a firm’s reputation,
generate goodwill among the research community, and encourage customers to use paid pr oducts.
Apple’s recent trend towards open -source and sharing more AI research demonstrates some of these
incentives at work. Apple has a notorious culture of secrecy, with numerous internal mechanisms in place to
prevent leaks and sequester information.20 Apple has benefited from this culture of secrecy in consumer
hardware design, a world where preventing leaks and copycat designs are critically important. However, this
culture was also a barrier for Apple to recruit top ML researchers, who typically str ongly value being able to
publish and share their work at conferences. Notably, several of Apple’s rival firms enabled researchers to do
so.21 In 2017, Apple changed its approach, launching a machine learning journal and enabling its researchers
to publicl y share findings at top ML conferences, including NeurIPS and ICML.22
16 US Patent and Trademark Office
17 Regalado, 2013
18 For an example of this view, see Mark Riedl’s quote in Simonite, 2018
19 US Patent and Trademark Office
20 Stone and Vance, 2009
21 Clark, 2015
22 Lewsing, 2017
6
Open -sourcing can also be used as a tool to generate more paying customers. For example, companies with
substantial cloud computing businesses often offer free machine learning tools to encourage customers to
design an application using the open -source tool. Then, thes e customers go on to pay for these compute -
intensive machine learning processes to be implemented on that same firm’s cloud service. In some cases, this
occurs through explicit lock -in. For example, Amazon’s image recognition software “Rekognition” is only
available on Amazon Web Services.23 In other cases, companies aim to retain customers through brand
loyalty; Google hopes that customers will use its ML open -source platform Tensorflow on Google Cloud,
though they could also use it on Amazon Web Services or Microsoft’s Azure. This proves to be a strong
incentive for open -sourcing given that cloud services appear to be incredibly lucrative. In 2018, Amazon,
Microsoft, and Google earned $25 billion, $23 billion, and $4 billion in revenue from their cloud bus inesses,
respectively.24 IBM’s $34 billion acquisition of open -source cloud computing provider RedHat (which
reportedly was also in acquisition talks with Google, Microsoft, and Amazon before selling to IBM25) and
Microsoft’s $11 billion JEDI cloud computi ng contract with the Pentagon26 further show how cloud business
is a priority for large AI developers.
The major downside of open -sourcing is the opportunity cost: open -sourcing in its purest form
means forgoing licensing fees from users. It also means sharing what would otherwise be
competitive secrets with rival companies. In the next section we discuss some ways that companies
manage to avoid these drawbacks.
A Hybrid Strategy: The Best of Both Worlds?
Despite the apparent conflict between building a large patent portfolio and participating in the open -source
community, selective licensing rights can en able a hybrid strategy where companies participate in the open -
source community while maintaining their patents. Companies can and do create selective licensing terms for
the use of their patents in open -source projects. These agreements can include defini ng permitted usage in a
way that allows some users to utilise the code while disincentivizing competitors to include that code in a
product.27
AI developers have also used selective open -source licensing agreements to achieve other ends. In 20 19,
Microsoft offered Azure cloud users access to a substantial portion of their patents as an incentive for users
to join the platform.28 Facebook attempted to add a licensing stipulation to its popular open -source platform
React, which would have caused users to retroactively lose their licenses if they ever engaged in patent
litigation against Facebook.29
23 Amazon Web Services
24 Griswold, 2019; Microsoft, 2018; Novet, 2018
25 Peterson, 2018
26 New York Times, 2019
27 E.g. The GNU Operating System GPL3 Open -Source license disincentiviz es commercial use.
28 Microsoft
29 The plan was a bandoned after developer backlash. See Wolff, 2017
7 3. Governments’ Pursuit of AI Innovation and National
Security Through IP Law
Governments around the world are grappling with how t o best take advantage of the recent wave of advances
in AI, with several nations releasing national plans on how their country intends to incentivize and capitalize
on AI innovation.30 These plans include ensuring that effective IP law regimes exist for AI and ML. This
tends to break down into two objectives: making AI patentable, and regulating access to national security
related IP. (For additional context on how national patent systems interact at the international level, see Box
1.)
Box 1: The Internat ional Patent System and the World Intellectual Property
Organization
Patent systems are primarily domestic in nature rather than international. Each country has its own patent
office and companies interested in seeking a patent for an invention must appl y separately in all
jurisdictions in which they wish to be awarded a patent.31 A patent awarded in one country cannot be used
to litigate infringement in another, though that patent does count as a form of “prior art” which can be
used to prevent the award of a patent for that invention in another country. The World Intellectual
Property Organization helps harmonize this process by assisting inventors to file their inventions in several
jurisdictions at once. However, the decision to award a patent will ult imately fall to individual countries.
While patent treaties such as “The Agreement on Trade -Related Aspects of Intellectual Property Rights”
(TRIPS32) have taken steps to harmonize intellectual property law across member nations, differences
persist at e very level of the process: what inventions are patentable, the level of scrutiny applied before a
patent is granted, how patents are enforced and reviewed, the length for which patents are valid, and
upkeep costs required to maintain the patent.
Making AI Patentable
A prominent element of several AI national plans is to ensure that AI -related inventions can receive patents
in a timely fashion. The goal of this policy is to encourage research and development (R&D) investment in AI
by rewarding tha t investment with a potentially lucrative patent. This strategy functions to both encourage
domestic companies to invest in AI -related research, and entice corporations choosing between different IP
systems to set up shop in their country rather than elsew here.
For example, US Patent and Trademark Office Chief Andrei Iancu recently expressed in a Senate hearing that
the US needs to make sure that its IP rules adequately protect and incentivize innovation in AI.33 In China’s
30 Dutton, 2018
31 World International Property Organization, 2017
32 Agreement on Trade -Related Aspects of Intellectual Property Rights, 1995
33 Simonite, 2018
8 state council plan, which decla red the nation’s intention to be the world leader in AI by 2030, one section
advised that policy makers in China must “[s]trengthen the protection of intellectual property in the field of
AI.”34 The European Patent Office also recently released specific gu idance on how to successfully patent
inventions in AI and machine learning,35 and Singapore is allowing AI patents to be “fast tracked” for review
through its patent system.36 Proposals for increasing Britain’s competitiveness in AI have also highlighted i ts
patent system’s challenges in protecting AI -related inventions as a liability.37
In some ways, the question of how to create patent protections for AI is not a new one. AI patents mostly fall
into the existing category of software patents, and countries have struggled for years to find regulatory
structures that incentivize innovation without allowing individual companies to control overly broad, abstract,
or obvious ideas.38 In fact, despite a push to allow more patenting and offer more stringe nt protections for
inventions, two recent major changes in IP law within the United States —the 2011 America Invents Act39
and the 2014 Supreme Court decision Alice Corp. v. CLS Bank International40—made it more difficult to claim
and enforce broad software patents. It will be difficult to change patenting rules in AI without also
implicating these existing decisions on software patents.
Furthermore, it is unclear whether expanding the range of patentable AI -relevant inventions would effectively
incentiviz e innovation. For one, AI and ML commercial activity has experienced massive growth and
international investment even while the patentability of innovations remains uncertain, suggesting that the
ability to patent AI is not necessary for innovation. In add ition, more patents increases the likelihood of
litigation, which could act to disincentivize innovation and slow down industry growth. This is particularly a
concern for some software patents due to their broad and abstract nature.41
34 Wang, 2018
35 European Patent Office, 2019
36 Spruson and Ferguson, 2019
37 Clark et al. 2019
38 Lee, 2013
39 Leahy -Smith America Invents Act, 2011
40 Alice Corp. v. CLS Bank International, 2014
41 Bessen, 2013
9 Box 2: Challenges Comparing Patent Filings: The US and China
China has outpaced the US in new patent applications related to the machine learning subfield of deep
learning.42 While some observers have interpreted this information as evidence of China’s fa st approaching
superiority over the US in new AI innovation, three key pieces of information about China and the US’s
patent systems should make us view these statistics in a different light.
First, patents in the US and China have very different standar ds, requirements and protections. The
majority of technology patents in China are filed as “utility model” patents , a category of patent in China
not extant in the US.43 Utility patents require a smaller inventive step, are subject to less rigorous
examination upon filing, and last only 10 years (in comparison to 21 years for American patents).
Consequently, this has led to filers taking advantage of lax inspections and review. In a 2018 report, the
Chinese state -owned Xinhua News Agency accused Chi na’s IP system of being characterized by “weak IP,
fake demands and some companies fervent on phony innovation,” according to a translation by Bloomberg
News.44 Chinese “invention patents” have requirements more similar to those in the US and last for 20
years rather than 10. However, they only comprise 23% of patent holdings in China.
Second, in China, there are strong government financial incentives for researchers to file patents, regardless
of the underlying patents’ merit.45 This is particularly true in AI, where the Chinese central government has
committed substantial public funds to encourage inventions.
Finally, it is important to look at the “discard rate” for Chinese patents —i.e. how quickly a patent is
discarded by its holder —and how they comp are to patents in the United States. A substantial percentage of
Chinese patent holders allow their patents to expire before the patent’s lifespan is complete —61% for
utility patents and 37% for invention patents —in comparison to a discard rate of 15% in t he US.46 When
viewed in context with the previous points, it becomes clear that there are many Chinese researchers who
file for AI patents in order to claim government incentives without genuinely believing their invention is
notable.47
Regulating National -Security -Related IP Access
Countries’ patent regimes for AI are not only shaped by economic motivations, but also by national security
interests. National governments typically pursue two primary interests on this front: to ensure that their own
national security apparatuses have access to state -of-the-art technology, and to withhold that access from
perceived rivals.
On the goal of ensuring access, a US court decision in 2015 held that the federal government can use
patented inventio ns without the permission of the patent holder and cannot be forced to cease usage of a
42 Huang, 2018
43 Chen, 2018
44 Ibid.
45 Ibid.
46 Ibid.
47 Ibid.
10 patented invention; the only remedy is for the patent holder to request damages assessed at market rate
(which amounts to compulsory licensing).48 This means if a paten t holder does not wish for their patent to be
used by the government (e.g. a patent that has potential surveillance applications) their only recourse after
suing for infringement is to force the government to pay for a reasonably costed license. This proce ss is quite
distinct from what happens when a private entity infringes on a patent, where the private entity can be
enjoined to cease usage of the patent or be assigned additional damages.
Additionally, over the last few years the Chinese government has passed broad laws on national security and
cybersecurity with implications for access to intellectual property.49 One of these laws mandates that network
operators (broadly defined) provide “technical support and assistance” to national security relevant
government offices.50 The exact legal authority of the Chinese government to force cooperation with Chinese
companies is difficult to know. The Center for a New American Security’s Ashley Feng reports that “U.S.
government officials, including at the FBI, interpreted this vagu e language to mean that all Chinese
companies, including Huawei, are subject to the direct orders of the Chinese government.”51 However, The
Diplomat’s Jack Wagner reports that the main purpose of the law is to mandate additional data localization
and stor age on Chinese mainland servers and to set standards around cybersecurity.52 Additional concerns
around Chinese corporations acting as extensions of the state are conceivable, but more speculative in nature.
On the goal of withholding access from perceiv ed rivals, the US and the UK have long had government
statutes that empower their patent offices to prevent public disclosure and bar the award of patents that have
national security implications, regardless of their other merits.53 In the US in 2018, 5,79 2 patent applications
were covered within these so -called “secrecy orders,” higher than at any point since the Cold War.54
48 Astornet Technologies Inc. v. BAE Systems, Inc, 2015
49 Feng, 2019
50 Ibid.
51 Ibid.
52 Wag ner, 2017
53 Schulz, 2013; Marks, 2010
54 Patent and Trademark Office, 2018
11 4. Three Scenarios for the Future of AI Intellectual Property
Strategies
Given the observed corporate AI developer IP strategies and growing government interest in IP law as a lever
for influencing AI development, how could the dynamics of AI intellectual property protection evolve? What
would these dynamics then mean for the f uture of the AI industry, and in particular, on the competitive
strategies employed by firms and states?
Here we present three plausible scenarios which focus on how the IP strategies of corporate AI developers
could evolve in the near future:
(1) Open Rese arch Continues: The status quo persists: open research alongside defensive patents
remains the norm within the ML industry.
(2) Patent Lawsuits and Secrecy: Offensive patent litigation breaks out within the ML community and
prompts additional secrecy among de velopers.
(3) Expansion of Patent Pools: In response to the threat of litigation, AI developers enter into additional
patent sharing agreements.
In the following section, we describe each scenario and present evidence to support its plausibleness. These
scenarios are intended to be illustrative rather than predictive; indeed, there are several alternative and hybrid
scenarios that could warrant further investigation as well.
Path #1: Open Research Continues
Scenario:
Each of the major AI developers weighs the costs of engaging in patent litigation against their competitors,
and decides that the threat of a countersuit and the symmetrical legal costs make litigation a poor choice.
Each developer continues to file for pa tents in order to maintain a credible response, but analogous to a
nuclear standoff, this “mutually assured destruction” framework holds.
This equilibrium is bolstered by competition for researchers who want to work at companies that prioritize
openness and cooperation with their peers. Some patent trolls —firms that profit from licensing and litigating
on patents without producing any products of their own —may gain control of patents and engage in
litigation without fear of reprisal, but these lawsuits r emain relatively insignificant.
Evidence in Favor:
● Despite the recent flurry of activity on the subject, the machine learning community currently shows
little sign of changing its open and non -litigious culture. Academics and individuals around the wor ld
can use cutting edge machine learning techniques from open -source platforms free of charge. There
has been some litigation over trade secret theft in autonomous vehicles (Waymo vs. Uber,55 Baidu vs.
JingChi56), but no large -scale patent litigation over broad concepts in machine learning.
55 Korosec, 2018
56 Borak, 2017
12 ● Current patent rules in the US make AI -related software patents less threatening for the purposes of
litigation than they were before the Supreme Court’s 2014 decision in Alice Corp. v. CLS Bank
International , which mad e software patents more likely to be classified as “abstract ideas” and thus
unpatentable. Given this decision, it is likely that many existing AI patents, particularly those
covering broad mathematical concepts, will be rejected during litigation or at th e Patent Trial and
Appeal Board.
● While some AI -related patents in the US are being granted, the majority are not. Data from patent
filings shows that in recent years, over 90% of AI -related patent applications in the US were initially
rejected, many for being merely “abstract ideas” that are not eligible for patentability.57 By
comparison, the overall rejection rate for patents in the US is 48%.58 Fewer AI -related patents mean
fewer opportunities for companies to litigate over infringement, thus bolsterin g the incentives for
open research.
● If large tech companies choose to litigate with one another over AI -related IP, they do not just have
to contend with a defendant's AI -related patents, but also with all of the other patents that would
likely be used in a countersuit.59 Google, Microsoft, Amazon, and IBM have expansive business
operations across several verticals; this makes patent aggression with other large companies less
attractive.
Path #2: Patent Lawsuits and Secrecy
Scenario:
Major patent litigations break out between AI developers. While the previous open equilibrium may be
preferable for the collective interest of major private AI developers, it may only take one large company
deciding it is in its interest to pursue active litigation for this state of affairs to deteriorate. For instance, if
IBM, with its trove of AI -related patents and its struggling core enterprise business,60 chose to litigate against
its rivals, it could provoke additional litigation. So -called “patent trolls” (entiti es that accumulate patents while
not producing products of their own) could also threaten to disrupt the mutually assured destruction
equilibrium and engage in lawsuits without fear of reprisal. This litigation could bleed over and affect
academic research . While the EU has a research exemption that protects academic use from being deemed
infringement, the US has no such exemption, and university researchers could thus potentially find
themselves on the wrong end of litigation.61
As litigation escalates, t here is a substantial incentive for companies to ensure that potentially patentable
inventions are kept secret from rivals. In a 2017 paper, Nick Bostrom discusses how the pursuit of patents in
AI could cause companies to share research less often in order to prevent other entities from using
57 Decker, 2019
58 Carley et al. 2015
59 For example, if Google were to engage in litigation against another company over machine learning patents,
they would have to contend with its opposition’s patents in E -commerce, search, drones, self -driving cars,
smartphone hardware and software and telecommunications, because each of these are areas in which Google
operates in and is thus capable of infringem ent.
60 Imbert, 2018
61 Miller, 2002
13 intermediary research to obtain a patent first.62 Public research could also be used as evidence in litigation to
prove infringement (e.g. releasing a model using a patented method), further dissuading companies in a
litigious environment from engaging in open -source communities.
Evidence in Favor:
● As discussed previously, governments in the US, China, EU, and elsewhere are pushing to broaden
the scope of patentable material in AI and encourage filings. If these effo rts translate into more
ostensibly defensible AI patents being filed successfully, this could increase incentives for litigation
between patent -holders.
● While Google reported that it is holding new AI patents on a defensive basis, other developers have
been non -committal about their future intention to litigate with their patents.63 When asked about
the issue, a Facebook spokesperson said that “its filings shouldn’t be read to indicate current or
future plans.’64 IBM’s patent counsel released a statement that said its large AI patent portfolio
“reflects its commitment to fundamental research.”65
Path #3: Expansion of Patent Pools
Scenario:
AI developers agree to cross -license their patents to one another in a “ pool” to reduce the risk of litigation.
We previously discussed patent sharing agreements in the context of companies like Google and Tencent
using these deals to gain market entry into new countries. However, patent sharing agreements need not only
be bet ween two actors. There are several historical examples of large technology companies pooling their
patents to protect against litigation and create advantageous licensing dynamics. The DVD6C Licensing
Group was comprised of eight of the most high profile p atent holders in DVD technology (including
Samsung and Toshiba, among others).66 Third -party manufacturers interested in using their technology could
approach the group as a one -stop clearinghouse in order to obtain licenses instead of approaching each
mem ber individually. Similar arrangements would enable companies within the pool to share trade secrets and
research and development, though too much coordination would come under the ire of antitrust
enforcement.
In this scenario, patent pools could be use d not only to curtail litigation risks, but also to limit or promote
certain applications of AI. As previously mentioned, companies are currently able to individually place
stipulations on patent licensing. If a company decided that it wanted to refuse to license its patents to
manufacturers of, for example, autonomous weapons, on moral grounds, it could certainly do so.
Extending this to a patent pool, corporate AI developers could group together to share intellectual property
and establish shared standa rds for how they wish to have their intellectual property used, perhaps based on
certain ethical principles. These standards could then be implemented via, for example, selective licensing
62 Bostrom, 2017
63 Simonite, 2018
64 Ibid.
65 Ibid.
66 DVS6C Licensing Group
14 agreements which restrict use of the pool’s IP to actors who commit to abiding by those standards.
Membership of the patent pool could also be made conditional on abiding by the pool’s standards. Indeed,
patent pools used for licensing agreements by standard setting bodies such as the International Standards
Organization are precedent for similar structures’ success. It is worth noting, however, that such multilateral
“refusals to deal” would need to be implemented with caution and appropriate due diligence in order to avoid
potential infringements of antitrust law.67
Evidence in Favor:
● Existing software patent pools show the demand for and utility of this type of coordination. Google
and several other tech companies participate in the Open Invention Network and Android
Networked Cross -License Agreement in order to protec t Linux and Android developers from
infringement litigation.68 As an alternative, the MPEG LA group is an example of a software patent
pool that operates as a profitable licensing association for its members.69 Facebook and Google’s
aforementioned existing patent non -aggression agreements could also be a path towards greater
cooperation.
● Increasing returns to scale in AI (meaning that more data improves AI systems and platforms, which
attracts more users, which in turn generates more data) could increase the odds of industry
centralization. A smaller number of relevant actors improves the feasibility of this kind of
coordination.
67 Department of Justice
68 Open Invention Network, 2017
69 MPEG LA
15 5. Conclusion
AI is poised to be a critically impactful technology, and its development will be deeply affected by existing
social and legal institutions. This paper has preliminarily explored an under -exami ned aspect of this
infrastructure: the legal rules, norms, and strategies governing AI -related intellectual property.
Leading corporate AI developers today have employed a dynamic and at times unintuitive IP strategy that
allows them to respond to the sh ifting competitive landscape surrounding them. Governments are also
seeking to shape IP systems that incentivize innovation around AI while also protecting their national security
interests. How each actor balances its varying objectives in relation to AI and how they choose to wield IP
strategies to achieve these objectives remains to be seen.
This preliminary analysis scratches the surface of what may be an important element of the strategic
landscape shaping competition and cooperation among AI firms a nd prominent national governments.
Further investigation in this direction could be fruitful for better understanding the goals and strategies of
actors seeking to protect AI -related intellectual property, and how these strategies have flow -on implications
on the competitive dynamics that arise between AI developers. This, in turn, could shed light on questions
related to the prospects for cooperation between these actors to achieve prosocial outcomes with respect to
AI ethics and safety, and more broadly, the approp riate governance of AI going forward.
16 References
Agreement on Trade -Related Aspects of Intellectual Property Rights. 1 Jan. 1995,
https://www.wto.org/english/docs\_e/legal \_e/27 -trips.pdf . Accessed 12 Apr. 2019.
Alice Corp v. CLS Bank International. 19 June 2014, https://www.supr emecourt.gov/opinions/13pdf/13 -
298\_7lh8.pdf . Accessed 12 Apr. 2019.
Amazon Web Services. “Amazon Rekognition.”
https://aws.amazon.com/rekognition/ . Accessed 7 Nov. 2019.
Amazon Web Services. “Ope n Source at AWS.” https://aws.amazon.com/opensource/ . Accessed 12 Apr.
2019.
American Bar Association . “Brief of Google Inc., et al as Amici Curiae in support of respondents”
https://www.americanbar.org/content/dam/aba/publications/supreme\_court\_preview/b riefs-v3/13 -
298\_resp\_amcu\_google -etal.authcheckdam.pdf Accessed 5 April 2020
Apple. “Legal Contact/Patent.” https://www.apple.com/legal/contact/patent.html . Accessed 12 Apr. 2019.
Astornet Technologies Inc. v. BAE Systems, Inc. 19 Sept. 2015,
http://www.cafc.uscourts.gov/sites/default/files/opinions -orders/14 -1854. Opinion.9 -15-2015.1.PDF .
Accessed 12 Apr. 2019.
Baidu. “Github/OpenEdge.” https://github.com/baidu/openedge . Accessed 12 Apr. 2019.
Bessen , James. “The patent troll crisis is really a software patent crisis.” Washington Post , 3 Sept. 2013,
https://www.washi ngtonpost.com/news/the -switch/wp/2013/09/03/the -patent -troll-crisis -is-really -a-
software -patent -crisis/ . Accessed 12 Apr. 2019.
Borak, Masha. “Baidu sues its former SVP for stealing self -driving trade -secrets and using them in his US -
based startup.” TechNode , 22 Dec. 2017, https://technode.com/2017/12/22/baidu -sues-former -svp-stealing -
self-driving -trade -secrets -using -us-based -startup/ . Accessed 12 Apr. 2019.
Bostrom, Nick. “Strategic Implications of Openness in AI Development.” Global Policy , 2017,
https://nickbostrom.com/papers/openness.pdf . Accessed 12 Apr. 2 019.
Cadell, Cate. “Google announces patent agreement with Tencent amid China push.” Reuters , 18 Jan. 2018,
https://www.reuters.com/article/us -china -google -tencent/google -announces -patent -agreement -with-
tencent -amid -china -push-idUSKBN1F80DF . Accessed 12 Apr. 2019.
Cai, Fangyu. “Tencent Open -Sources its Massive Multi -Labeled Image Dataset.” 10 Sept. 2018,
https://medium.com/syncedreview/tencent -open -sources -its-massive -multi -labeled -image -dataset -
7b0b3dd5373f . Accessed 12 Apr. 2019.
Carley, Michael, Deepak Hegde, and Alan Marco. “What is the probability of receiving a US patent?” Yale
Journal of Law and Technology , 2015.
Chen, Lulu YiLun. “China Claims More Patents Than Any Country –Most Aare Worthless.” Bloomberg , 26
Sept. 2018,
17 https://www.bloomberg.com/news/articles/2018 -09-26/china -claims -more -patents -than-any-country -most -
are-worthless . Accessed 12 April 2019.
Clark, Alex, Ella Hollowood, Imogen Harper, and Alexi Mostrous. “AI in the UK.” Tortoise Media , 4 Dec.
2019, https://members.tortoisemedia.com/2 019/12/04/ai -in-the-uk/content.html . Accessed 16 Jan. 2020.
Clark, Jack. “Apple’s Deep Learning Curve.” Bloomberg , 29 Oct. 2015,
https: //www.bloomberg.com/news/articles/2015 -10-29/apple -s-secrecy -hurts -its-ai-software -development.
Accessed 12 Apr. 2019.
Cohen, Jon. “CRISPR Patent Fight Revived.” Science , 5 July 2019,
https://science.sciencemag.org/content/365/6448/15.2 . Accessed 22 Sept. 2019.
Conger, Kate, David E. Sanger, and Scott Shane. “Microsoft Wins Pentagon’s $10 Billion JEDI Contract,
Thwarting Amazon.” New York Times , 25 Oct. 2019,
https://www.nytimes.com/2019/10/25/technology/dod -jedi-contract.html . Accessed 7 Nov. 2019.
Decker, Susan. “Who Profits from AI? It’s Getting Harder to Find Out.” Bloomberg , 23 Jan. 2019,
https://www.bloomberg.com/news/articles/2019 -01-23/who -profits -from -ai-uncertain -as-u-s-patent -office -
gets-pickier . Acce ssed 26 Aug. 2019.
Department of Justice. “The Strategic Use of Licensing: Unilateral Refusals to License PatentsDeal.”
https://www.justice.gov/atr/chapter -1-strategic -use-licensing -unilateral -refusals -license -patents . Accessed 12
Apr. 2019.
Dignan, Larry. “All of Amazon’s 2017 operating income comes from AWS.” ZDNet , 1 Feb. 2018,
https://www.zdnet.com/article/all -of-amazons -2017-operating -income -comes -from -aws/. Accessed 12 Aug.
2019.
Duhigg, Charles and Steve Lohr. “The Patent, Used as a Sword.” New York Times , 7 Oct. 2012,
https://www.nytimes.com/2012/10/08/technology/patent -wars-among -tech-giants -can-stifle-
competition.html . Accessed 12 Apr. 2019.
Dutton, Tim. “An Overview of National AI Strategies.” 28 Jun. 2018, https://medium.com/politics -ai/an -
overview -of-national -ai-strategies -2a70ec6edfd . Accessed 12 Apr. 2019.
DVD6C Licensing Group. “DVD6C Patent Pool.” http://www.dvd6cla.com/ . Accessed 12 Apr. 2019.
European Patent Office. “Guidelines for Examination.” https://www.e po.org/law -practice/legal -
texts/html/guidelines2018/e/g\_ii\_3\_3\_1.htm . Accessed 12 Apr. 2019.
Feng, Ashley. “We Can’t Tell if Chinese Firms Work for the Party.” CNAS , 7 Feb. 2019,
https://www.cnas.org/publications/commentary/we -cant-tell-if-chinese -firms -work -for-the-party . Accessed
12 Apr. 2019.
Ford, Martin R. Architects of Intelligence: the truth about AI from the people building it . Pack t, 2018.
Google. “Patents in the Service of Open Source.”
https://www.google.com/patents/opnpledge/ . Accessed 12 Apr. 2019.
18 Google. “Introducing PAX: the Android Networked Cross -License Agreement.” 3 Apr. 2017,
https://blog.google/topics/public -policy/introducing -pax-android -networked -cross -license -agreement/ .
Accessed 7 Nov. 2019.
Griswold, Alison. “Amazon Web Services brought in more money than McDonald’s in 2018.” Quartz , 1 Feb.
2019,
https://qz.com/1539546/amazon -web-services -brought -in-more -money -than-mcdonalds -in-2018/ .
Accessed 7 Nov. 2019.
Gurman, Mark. “Apple Warns Employees to Stop Leaking Information to the Media.” Bloomberg , 13 Apr.
2013,
https://www.bloomberg.com/news/articles/2018 -04-13/apple -warns -employees -to-stop-leaking -
information -to-media . Accessed 12 Apr. 2 019.
Hardy, Quentin. “Google Buys Motorola For Patent Parts.” Forbes , 15 Aug. 2011,
https://www.forbes.com/sites/quentinhardy/2011/08/15/google -buys-motorola -for-patent -
parts/#17555b562fff . Accessed 12 Apr. 2019 .
Huang, Echo. “China has shot far ahead of the US on deep -earning patents.” Quartz , 2 Mar. 2018,
https://qz.com/1217798/china -has-shot-far-ahead -of-the-us-on-ai-patents/ . Accessed 26 Aug. 2019.
IBM. “IBM and NVIDIA Collaborate to Expand Open Source Machine Learning Tools for Data Scientists.”
10 Oct. 2018, https://newsroom.ibm.com/2018 -10-10-IBM-and-NVIDIA -Collaborate -to-Expand -Open -
Source -Machine -Learning -Tools -for-Data-Scientists . Accessed 12 Apr. 2019.
Imbert, Fred. “Sell IBM shares because its profits are in an ‘ irreversible structural decline,’ analyst says.”
CNBC , 4 Oct. 2018,
https://www.cnbc.com/2018/10/04/ibm -profits -are-in-an-irreversible -structural -decline -analyst -says.html .
Accessed 22 Sept. 2019.
Korosec, Kirsten. “Waymo V. Uber: What You Need to Know About the High -Stakes Self -Driving Tech
Trial.” Fortune , 5 Feb. 2018, http://fortune.com/2018/02/05/waymo -v-uber-what-you-need-to-know -about -
the-high-stakes -self-driving -tech-trial/ . Accessed 12 Apr. 2019.
Leahy -Smith America Invents Act. 16 Sept. 2011,
https://www.uspto.gov/sites/default/files/aia\_implementation/20110916 -pub-l112-29.pdf . Accessed 12
Apr. 2019.
Lee, Timothy B. “Here’s why economists hate software patents.” Washington Post , 31 July 2013,
https://www.washingtonpost.com/news/the -switch/wp/2013/07/31/heres -why-economists -hate-software -
patents/ . Accessed 7 Nov. 2019.
Leswing, Kif. “Ap ple is now publishing a ‘journal’ with its cutting -edge machine learning research.” Business
Insider , 19 July 2017, https://www.businessinsider.com/apple -announces -apple -machine -learning -journal -
2017-7. Accessed 12 Apr. 2019.
Marks, Paul. “UK keeps three times as many patents secret as the US.” New Scientist , 23 Mar. 2010,
https://www.newscientist.com/article/dn18691 -uk-keeps -three -times -as-many -patents -secret -as-the-us/.
Accessed 12 Apr. 2019.
Microsoft. “Annual Report 2018.” 16 Oct. 2018.
https://www.microsoft.com/en -us/annualreports/ar2018/annualreport . Accessed 7 Nov. 2019.
19
Microsoft. “Azure IP Advantage.”
https://azur e.microsoft.com/en -us/overview/azure -ip-advantage/ . Accessed 12 Apr. 2019.
Microsoft. “Discover more options with open source on Azure.” https://azure.microsoft.com/en -
us/overview/choose -azure -opensource/ .
Microsoft. “Microsoft to acquire Github for $7.5 Billion.” 4 June 2018,
https://news.microsoft.com/2018/06/04/microsoft -to-acquire -github -for-7-5-billion/ . Accessed 12 Apr.
2019.
Miller, Jennifer. “Sealing the Coffin on the E xperimental Use Exception.” 2002,
https://scholarship.law.duke.edu/cgi/viewcontent.cgi?article=1081&context=dltr . Accessed 12 Apr. 2019.
MPEG LA. https://www.mpegla.com/ . Accessed 7 Nov. 2019.
Novet, Jordan. “Google says its cloud now brings in $1 billion per quarter.” CNBC , 1 Feb. 2018,
https://www.cnbc.com/2018/02/01/google -cloud -revenue -passes -1-billion -per-quarter.html . Accessed 7
Nov. 2019.
Open Invention Network. https://www.openinventionnetwork.com/ . Accessed 7 Nov. 2019.
Patent and Trademark Office. “Invention Secrecy Activity.”
https://fas.org/sgp/othergov/invention/stats.html . Accessed 12 Apr. 2019.
Peterson, Becky. “IBM's $34 billion Red Hat acquisition came after deal talks with Microsoft, Google, and
Amazon, sources say.” Business Insider , 16 Dec. 2018,
https://www.businessinsider.com/red -hat-deal-talks-with-amazon -microsoft -google -before -34-billion -ibm-
acquisition -2018-12. Accessed 12 Apr. 2019.
Quinn, Gene. “Google: We Don’t Sell to Patent Trolls.” IPWatchdog , 2 May 2013,
http://www.ipwatchdog.com/2013/05/02/google -we-dont-sell-to-patent -trolls/id=39927/ . Accessed 12
Apr. 2019.
Regalado, Anthony. “Google’s Growing Patent Stockpile.” MIT Technology Review , 29 Nov 2013,
https://www.technologyreview.com/s/521946/googles -growing -patent -stockpile/ . Accessed 26 Aug. 2019.
Schulz, G.W. “Government Secrecy Order on Patents Have Stifled More Than 5,000 Inventions.” Wired , 16
Apr. 2013, https://www.wired.com/2013/04/gov -secrecy -orders -on-patents/ . Accessed 12 Apr. 2019.
Simonite, Tom. “Despite Pledging Openness, Companies Rush to Patent AI Tech.” Wired , 31 July 2018,
https://www.wired.com/story/d espite -pledging -openness -companies -rush-to-patent -ai-tech/ . Accessed 12
Apr. 2019.
Spruson and Ferguson, “Singapore Update: IPOS Grants First Accelerated Patent Under Fintech Fast Track
Initiative.” 14 Dec. 2018, https://www.spruson.com/fintech/singapore -update -ipos-grants -first-accelerated -
patent -fintech -fast-track -initiative/ . Accessed 12 Apr. 2019.
Stone, Brad and Ashlee Vance. “Apple’s Obsession With Secrecy Grows Stronger.” New York Times , 22 June
2009, https://www.nytimes.com/2009/06/23/technology/23apple.html .
TensorFlow. “An end -to-end open -source machine learning platform.”
20 https://www.tensorflow.org/ . Accessed 12 Apr. 2019.
United States Patent and Trademark Office. “USPTO Fee Schedule.”
https://www.uspto.gov/learning -and-resources/fees -and-payment/uspto -fee-
schedule#Patent%20Maintenance%20Fee . Accessed 26 Aug. 2019.
Wagner, Jack. “China’s Cybersecurity Law: What You Need to Know.” The Diplomat , 1 June 2017,
https://thediplomat.com/2017/06/chinas -cybersecurity -law-what-you-need-to-know/ .
Wang, Qian. “Global Perspectives: AI and IP Policies Around the World.” Artificial Intelligence: Intellectual
Property Considerations , 31 Jan. 2018. Presentation.
Webb, Michael, Nicholas Bloom, Nick Short, and Josh Lerner. “Some Facts of High Tech Patenti ng.”
Harvard Business School , July 2018, https://www.hbs.edu/faculty/Publication%20Files/19 -014\_cf3bccb7 -1ab7-
4f43-b822 -2d7c9b04a2c0.pdf . Acc essed 12 Apr. 2019.
Wolff, Adam. “Relicensing React, Jest, Flow, and Immutable.js.” Facebook , 22 Sept. 2017,
https://code.fb.com/web/relicensing -react-jest-flow-and-immutable -js/. Accessed 12 Apr. 2019.
World Economic Forum. “Artificial Intelligence Collides with Patent Law.” Apr. 2018,
http://www3.weforum.org/docs/ WEF\_48540\_WP\_End\_of\_Innovation\_Protecting\_Patent\_Law.pdf .
Accessed 26 Aug. 2019.
World International Property Organization. “Protecting yYour Inventions Abroad: Frequently Asked
Questions About the Patent Cooperation Treaty.” Oct. 2017, https://www.wipo.int/pct/en/faqs/faqs.html .
Accessed 12 Apr. 2019.
World International Property Organization. “What is Intellectual Property?” https://www.wipo.int/about -
ip/en/ . Accessed 7 Nov. 2019.
|
5492602c-0920-46d3-94e4-7f0ec1d2646c
|
trentmkelly/LessWrong-43k
|
LessWrong
|
John Ioannidis: Why Most Clinical Research Is Not Useful (2016)
|
3f93f1f4-ddf0-4960-9091-a5a5fd93a6e3
|
StampyAI/alignment-research-dataset/eaforum
|
Effective Altruism Forum
|
AI Value Alignment Speaker Series Presented By EA Berkeley
**AI Value Alignment Speaker Series**
**Presented By EA Berkeley**
Value alignment, roughly speaking, is the problem of ensuring that AI agents behave in a manner consistent with the values of their human principals. This modern incarnation of the principal-agent problem has historical roots in politics (i.e., how do we ensure that government acts on behalf of the people) and economics (i.e., how do we ensure that a corporation acts on behalf of those who have invested their resources into the corporation), but is at the crux of recent existential risk work in AI safety, such as Stuart Russell’s *Human Compatible* and Brian Christian’s *The Alignment Problem*. The purpose of the AI value alignment speaker series is to introduce persons, early in their studies and careers, to the issues of value alignment so they can consider educational and career paths to help defuse the alignment problem.
In addition to the chance to learn from and speak with experts in the field, participating in this series has additional *opportunities*. For examples, **Brian Christian**, the co-author of the **Top 5 Amazon Best Seller in Computer Science**(six years after publication), *Algorithms to Live By*, and **Rohin Shah (DeepMind) who is editor of the** ***Alignment Newsletter*****,** have each agreed to do a 30 minute one-on-one virtual meal with a consistent attendee of the series. Also, there will be book giveaways and other *opportunities*. Please attend regularly to be considered for these *opportunities*.
Here is our current working schedule (all times US Pacific):
**Brian Christian** (*The Alignment Problem*author and *Algorithms to Live By* co-author)
**The Alignment Problem: A Q&A**
March 1st, 2022 5P-6P
**Rick Ferri** (*President* of the *John C. Bogle Center* & *Bogleheads’ Guide to Retirement Planning*co-author)
**Ellen Quigley** (*Centre for the Study of Existential Risk* & Advisor to the CFO, *University of Cambridge*)
**Universal Ownership: Is Index Investing the New Socially Responsible Investing?**
March 15th, 2022 11A-1P
**Aaron Tucker** (*Cornell*), **Caroline Jeanmarie** (*Oxford*), **Jon Ward** (*Open AI*)
**Value Alignment: Early Career and Educational Guidance from Experts**
April 5th, 5P-6P
**Roman Yampolskiy** (*Chapman & Hall/CRC* Artificial Intelligence and Robotic *Series Editor*)
**A Fireside Chat**
April 8th, 2022 9A-11A
**Seth Baum** (*Executive Director* of the *Global Catastrophic Risk Institute*)
**Limits of the Value Alignment Paradigm**
April 13th, 2022 10A-11A
**Rohin Shah** (*Deep Mind* and *Founding Editor* of the *Alignment Newsletter*)
**What is Value Alignment?**
April 26th, 2022 Noon-1:30P
**How to Attend:**[**https://berkeley.zoom.us/j/958068842EA76?**](https://berkeley.zoom.us/j/958068842EA76?)**Password: EASpeaker**
|
71f10078-d960-4016-bec4-a7edcc147c95
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Why you can't treat decidability and complexity as a constant (Post #1)
Or, why you need to fix a machine before you can prove anything, and you also need to fix the constraints on the machine. This also holds importantly for problems claimed to be decidable or undecidable by algorithms.
Basically, the reason here is that different machine/computational models define different sets of computable functions, and you cannot treat the machines as equivalent in power.
This is admittedly something that most people get implicitly today, but it can definitely cause problems, and it certainly caused a lot of problems for Church and Turing et al in that they incorrectly jumped to the conclusion that the Turing Machines could compute any possible computer, or compute any possible algorithm, probably because they thought that you could treat a Turing Machine as equivalent to any other machine. Why the Church-Turing thesis is false, in the sense that it doesn't apply universally will be covered in an upcoming post in this sequence, but for now take it as a fact that there are different models of computation that define different computable sets, or equivalently different problems become decidable when new models are added.
Edit: I'm focusing on a variant of the thesis for the next post, in which I focus not on what's possible given our mathematical structure of reality, but whether it's possible to exceed the Turing Machine at all in any possible mathematical structure, and another variant where we restrict it to Turing-equivalent models, but we can arbitrarily change what we can offer the Turing Machine.
This is important, since I'm mostly going in a philosophical, not practical direction, and the thesis made no reference to our specific mathematical reality at all, so it's important.
From wikipedia:
> "In computability theory, the Church–Turing thesis (also known as computability thesis,[1] the Turing–Church thesis,[2] the Church–Turing conjecture, Church's thesis, Church's conjecture, and Turing's thesis) is a thesis about the nature of com
|
2d982b38-b5cd-458c-af28-14920ab52095
|
trentmkelly/LessWrong-43k
|
LessWrong
|
When can a mimic surprise you? Why generative models handle seemingly ill-posed problems
Thanks to Chris Leong and Nora Belrose for their feedback. This is meant to be part of an entry to the Future Fund AI Worldview Competition, but a later post is intended to address the competition questions head on.
In this post, I explore mimics. Mimics are what you get when you join a simulator with a generator. Examples are language models that learn to predict text sequences (the simulator), and generate samples of text sequences from their predictions (the generator). A number of AI safety researchers have mentioned that mimics seem to be safer than "traditional" AI architectures like reinforcement learners, with the proposed reason for this often being that mimics are less "agentic" or "goal-driven" than traditional architecture. moire's Simulators is a particularly thorough overview that makes a similar point.
In this post, I argue that a key feature of mimics is unrelated to their "agentiness": someone who can forecast a mimic's training data can also forecast a mimic's behaviour. I call this phenomenon synchronisation. Synchronisation is possible even when the operator can only forecast some crude features of the training sequence.
Certain methods for fine-tuning mimics allow mimics to be optimised for certain tasks while staying synchronised with the operator. This enables mimics to be controlled in a manner that maintains synchronisation and consequently remain easy to predict.
However, some kinds of objectives do not facilitate synchronised control of mimics. If an operator fine-tunes a mimic to control some feature of the world over which it wouldn't normally have complete control, then the operator should generally expect the mimic's output to diverge from forecasts based on the training data. In practice, the consequences of this divergence is reminiscent of failures due to Goodhart's law.
The extremely brief summary of this post is:
* Idealised mimics do what you expect them to when you're trying to control features of their output
* Idealis
|
487b9c3a-f580-4b4a-a50d-a5c53404cabb
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Learning Impact in RL
I present a simple Deep-RL flavour idea for learning an agent's impact that I'm thinking of trying out. I don't, ATM, think it's very satisfying from a safety point of view, but I think it's at least a bit relevant, so I'm posting here for feedback, iyi.
IDEA: Instead of learning
P(st+1|st,at)
with a single network, instead, learn it as:
P(st+1|st,at)=I(st+1|st,at)⨁T(st+1|st).
The ⨁ could mean mixing the distributions, adding the preactivations, or adding the samples from T and I. I think adding the samples probably makes the most sense in most cases.
Now, I is trained to capture the agent's impact, and T should learn the "passive dynamics". Apparently things like this have been tried before (not using DL, AFAIK, though), e.g.https://papers.nips.cc/paper/3002-linearly-solvable-markov-decision-problems.pdf
If we do a good job of disentangling an agent's impact from the passive dynamics, then we can do reduced-impact in a natural way.
This idea was inspired by internal discussions at MILA/RLLAB and the Advantage-function formulation of value-based RL.
|
d99f86ba-3ed6-44a7-8155-17c36d321a5d
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Reversed Stupidity Is Not Intelligence
> “. . . then our people on that time-line went to work with corrective action. Here.”
>
> He wiped the screen and then began punching combinations. Page after page appeared, bearing accounts of people who had claimed to have seen the mysterious disks, and each report was more fantastic than the last.
>
> “The standard smother-out technique,” Verkan Vall grinned. “I only heard a little talk about the ‘flying saucers,’ and all of that was in joke. In that order of culture, you can always discredit one true story by setting up ten others, palpably false, parallel to it.”
>
> —H. Beam Piper, Police Operation
Piper had a point. Pers’nally, I don’t believe there are any poorly hidden aliens infesting these parts. But my disbelief has nothing to do with the awful embarrassing irrationality of flying saucer cults—at least, I hope not.
You and I believe that flying saucer cults arose in the total absence of any flying saucers. Cults can arise around almost any idea, thanks to human silliness. This silliness operates orthogonally to alien intervention: We would expect to see flying saucer cults whether or not there were flying saucers. Even if there were poorly hidden aliens, it would not be any less likely for flying saucer cults to arise. The conditional probability P(cults|aliens) isn’t less than P(cults|¬aliens), unless you suppose that poorly hidden aliens would deliberately suppress flying saucer cults.1 By the Bayesian definition of evidence, the observation “flying saucer cults exist” is not evidence against the existence of flying saucers. It’s not much evidence one way or the other.
This is an application of the general principle that, as Robert Pirsig puts it, “The world’s greatest fool may say the Sun is shining, but that doesn’t make it dark out.”2
If you knew someone who was wrong 99.99% of the time on yes-or-no questions, you could obtain 99.99% accuracy just by reversing their answers. They would need to do all the work of obtaining good evidence entan
|
22532f2f-018c-48b5-b455-f90252585db1
|
StampyAI/alignment-research-dataset/alignmentforum
|
Alignment Forum
|
Can corrigibility be learned safely?
EDIT: Please note that the way I use the word "corrigibility" in this post isn't quite how Paul uses it. See [this thread](https://www.lesswrong.com/posts/o22kP33tumooBtia3/can-corrigibility-be-learned-safely#jo2cwbB3WK7KyGjpy) for clarification.
This is mostly a reply to Paul Christiano's [Universality and security amplification](http://ai-alignment.com/universality-and-security-amplification-551b314a3bab) and assumes familiarity with that post as well as Paul's AI alignment approach in general. See also [my previous comment](http://www.lesswrong.com/posts/ZyyMPXY27TTxKsR5X/problems-with-amplification-distillation) for my understanding of what corrigibility means here and the motivation for wanting to do AI alignment through corrigibility learning instead of value learning.
Consider the [translation example](http://medium.com/@weidai/to-put-it-another-way-a-human-translator-has-learned-a-lot-of-valuable-information-much-of-it-48457f95b9bf) again as an analogy about corrigibility. Paul's alignment approach depends on humans having a notion of "corrigibility" (roughly "being helpful to the user and keeping the user in control") which is preserved by the amplification scheme. Like the information that a human uses to do translation, the details of this notion may also be stored as connection weights in the deep layers of a large neural network, so that the only way to access them is to provide inputs to the human of a form that the network was trained on. (In the case of translation, this would be sentences and associated context, while in the case of corrigibility this would be questions/tasks of a human understandable nature and context about the user's background and current situation.) This seems plausible because in order for a human's notion of corrigibility to make a difference, the human has to apply it while thinking about the meaning of a request or question and "translating" it into a series of smaller tasks.
In the language translation example, if the task of translating a sentence is broken down into smaller pieces, the system could no longer access the full knowledge the Overseer has about translation. By analogy, if the task of breaking down tasks in a corrigible way is itself broken down into smaller pieces (either for security or because the input task and associated context is so complex that a human couldn't comprehend it in the time allotted), then the system might no longer be able to access the full knowledge the Overseer has about "corrigibility".
In addition to "corrigibility" (trying to be helpful), breaking down a task also involves "understanding" (figuring out what the intended meaning of the request is) and "competence" (how to do what one is trying to do). By the same analogy, humans are likely to have introspectively inaccessible knowledge about both understanding and competence, which they can't fully apply if they are not able to consider a task as a whole.
Paul is aware of this problem, at least with regard to competence, and his [proposed solution](https://ai-alignment.com/universality-and-security-amplification-551b314a3bab) is:
> I propose to go on breaking tasks down anyway. This means that we will lose certain abilities as we apply amplification. [...] Effectively, this proposal replaces our original human overseer with an impoverished overseer, who is only able to respond to the billion most common queries.
How bad is this, with regard to understanding and corrigibility? Is an impoverished overseer who only learned a part of what a human knows about understanding and corrigibility still understanding/corrigible enough? I think the answer is probably no.
With regard to understanding, natural language is famously ambiguous. The fact that a sentence is ambiguous (has multiple possible meanings depending on context) is itself often far from apparent to someone with a shallow understanding of the language. (See [here](http://www.greaterwrong.com/posts/Mhaikukvt6N4YtwHF/dragon-army-retrospective#comment-Pj6b4SDDdf3YrWp9i) for a recent example on LW.) So the overseer will end up being overly literal, and misinterpreting the meaning of natural language inputs without realizing it.
With regard to corrigibility, if I try to think about what I'm doing when I'm trying to be corrigible, it seems to boil down to something like this: build a model of the user based on all available information and my prior about humans, use that model to help improve my understanding of the meaning of the request, then find a course of action that best balances between satisfying the request as given, upholding (my understanding of) the user's morals and values, and most importantly keeping the user in control. Much of this seems to depend on information (prior about humans), procedure (how to build a model of the user), and judgment (how to balance between various considerations) that are far from introspectively accessible.
So if we try to learn understanding and corrigibility "safely" (i.e., in small chunks), we end up with an [overly literal overseer](https://www.greaterwrong.com/posts/ySSEz5CmSEo6MbokQ/reframing-misaligned-agi-s-well-intentioned-non-neurotypical) that lacks common sense understanding of language and independent judgment of what the user's wants, needs, and shoulds are and how to balance between them. However, if we amplify the overseer enough, eventually the AI will have the option of learning understanding and corrigibility from external sources rather than relying on its poor "native" abilities. As Paul explains with regard to translation:
> This is potentially OK, as long as we learn a good policy for leveraging the information in the environment (including human expertise). This can then be distilled into a state maintained by the agent, which can be as expressive as whatever state the agent might have learned. Leveraging external facts requires making a tradeoff between the benefits and risks, so we haven’t eliminated the problem, but we’ve potentially isolated it from the problem of training our agent.
So instead of directly trying to break down a task, the AI would first learn to understand natural language and what "being helpful" and "keeping the user in control" involve from external sources (possibly including texts, audio/video, and queries to humans), distill that into some compressed state, then use that knowledge to break down the task in a more corrigible way. But first, since the lower-level (less amplified) agents are contributing little besides the ability to execute literal-minded tasks that don't require independent judgment, it's unclear what advantages there are to doing this as an Amplified agent as opposed to using ML directly to learn these things. And second, trying to learn understanding and corrigibility from external humans has the same problem as trying to learn from the human Overseer: if you try to learn in large chunks, you risk corrupting the external human and then learning corrupted versions of understanding and corrigibility, but if you try to learn in small chunks, you won't get all the information that you need.
The conclusion here seems to be that corrigibility can't be learned safely, at least not in a way that's clear to me.
|
df924750-1f4e-44f3-a384-ebc21284ea55
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Neural nets designing neural nets
|
ca4146af-e965-4451-9d6f-231742862251
|
trentmkelly/LessWrong-43k
|
LessWrong
|
What's the difference between newer Atari-playing AI and the older Deepmind one (from 2014)?
My impression was that thing that put Deepmind on the map was an AI that could play multiple Atari games. Lately there's been new Atari-playing AI (both from Deepmind and other companies) that are making the news. Are they doing basically the same thing 2014 Deepmind was doing but better? Are they doing a fundamentally different thing? Can someone explain the diff like I'm five?
|
0fc37cb7-8a13-4466-8e1f-8d7520aa6b3e
|
StampyAI/alignment-research-dataset/special_docs
|
Other
|
Towards formalizing universality
The scalability of [iterated amplification](https://arxiv.org/pdf/1810.08575.pdf) or [debate](https://arxiv.org/abs/1805.00899) seems to depend on whether large enough teams of humans can carry out arbitrarily complicated reasoning. Are these schemes “universal,” or are there kinds of reasoning that work but which humans fundamentally can’t understand?
This post defines the concept of “ascription universality,” which tries to capture the property that a question-answering system \*\*A\*\* is better-informed than any particular simpler computation \*\*C\*\*.
[These](/informed-oversight-18fcb5d3d1e1) [parallel](/universality-and-consequentialism-within-hch-c0bee00365bd) [posts](/universality-and-model-based-rl-b08701394ddd) explain why I believe that the alignment of iterated amplification largely depends on whether HCH is ascription universal. Ultimately I think that the “right” definition will be closely tied to the use we want to make of it, and so we should be refining this definition in parallel with exploring its applications.
I’m using the awkward term “ascription universality” partly to explicitly flag that this is a preliminary definition, and partly to reserve linguistic space for the better definitions that I’m optimistic will follow.
(Thanks to Geoffrey Irving for discussions about many of the ideas in this post.)
I. Definition
=============
We will try to define what it means for a question-answering system \*\*A\*\* to be “ascription universal.”
1. Ascribing beliefs to A
-------------------------
Fix a language (e.g. English with arbitrarily big compound [terms](/approval-directed-algorithm-learning-bf1f8fad42cd)) in which we can represent questions and answers.
To ascribe beliefs to \*\*A\*\*, we ask it. If \*\*A\*\*(“are there infinitely many twin primes?”) = “probably, though it’s hard to be sure” then we ascribe that belief about twin primes to \*\*A\*\*.
This is not a general way of ascribing “belief.” This procedure wouldn’t capture the beliefs of a native Spanish speaker, or for someone who wasn’t answering questions honestly. But it can give us a sufficient condition, and is particularly useful for someone who wants to use \*\*A\*\* as part of an alignment scheme.
Even in this “straightforward” procedure there is a lot of subtlety. In some cases there are questions that we can’t articulate in our language, but which (when combined with \*\*A\*\*’s other beliefs) have consequences that we can articulate. In this case, we can infer something about \*\*A\*\*’s beliefs from its answers to the questions that we can articulate.
2. Ascribing beliefs to arbitrary computations
----------------------------------------------
We are interested in whether \*\*A\*\* “can understand everything that could be understood by someone.” To clarify this, we need to be more precise about what we mean by “could be understood by someone.”
This will be the most informal step in this post. (Not that any of it is very formal!)
We can imagine various ways of ascribingbeliefs to an arbitrary computation \*\*C\*\*. For example:
\* We can give \*\*C\*\* questions in a particular encoding and assume its answers reflect its beliefs. We can either use those answers directly to infer \*\*C\*\*’s beliefs (as in the last section), or we can ask what set of beliefs about latent facts would explain \*\*C\*\*’s answers.
\* We can view \*\*C\*\* as optimizingsomething andask what set of beliefs rationalize that optimization. For example, we can give \*\*C\*\* a chess board as input, see what move it produces, assume it is trying to win, and infer what it must believe. We might conclude that \*\*C\*\* believes a particular line of play will be won by black, or that \*\*C\*\* believes general heuristics like “a pawn is worth 3 tempi,” or so on.
\* We can reason about how \*\*C\*\*’s behavior depends on facts about the world, and ask what state of the world is determined by its current behavior. For example, we can observe that \*\*C\*\*(113327) = 1 but that \*\*C\*\*(113327) “would have been” 0 if 113327 had been composite, concluding that \*\*C\*\*(11327) “knows” that 113327 is prime. We can extend to probabilistic beliefs, e.g. if \*\*C\*\*(113327) “probably” would have been 0 if 113327 had been composite, then we might that \*\*C\*\* knows that 113327 is “probably prime.” This certainly isn’t a precise definition, since it involves considering logical counterfactuals, and I’m not clear whether it can be made precise. (See also ideas along the lines of [“knowledge is freedom”](https://www.lesswrong.com/posts/b3Bt9Cz4hEtR26ANX/knowledge-is-freedom).)
\* If a computation behaves differently under different conditions, then we could use restrict attention to a particular condition. For example, if a question-answering system appears to be bilingual but answers questions differently in Spanish and English, we could ascribe two different sets of beliefs. Similarly, we could ascribe beliefs to any subcomputation. For example, if a part of \*\*C\*\* can be understood as optimizing the way data is laid out in memory, then we can ascribe beliefs to that computation about the way that data will be used.
Note that these aren’t intended to be efficient procedures that we could actually apply to a given computation \*\*C\*\*. They are hypothetical procedures that we will use to define what it means for \*\*A\*\* to be universal.
I’m not going to try to ascribe a single set of beliefs to a given computation; instead, I’ll consider all of the reasonable ascription procedures. For example, I think different procedures would ascribe different beliefs to a particular human, and don’t want to claim there is a unique answer to what a human “really” believes. A universal reasoner needs to have more reasonable beliefs than the beliefs ascribed to that a human using any particular method.
An ascription-universal reasoner needs to compete with any beliefs that can be ascribed to \*\*C\*\*, so I want to be generous with this definition. For example, given a chess-playing algorithm, we might rationalize it as trying to win a game and infer its beliefs about the rules of chess. Or we might rationalize it as trying to look like a human and infer its beliefs about what a human would do. Or something different altogether. Most of these will be kind of crazy ascriptions, but I want to compete with them anyway (competing with crazier beliefs will turn out to just be easier).
It’s not totally clear what counts as a “reasonable” ascription procedure, and that’s the biggest source of informality. Intuitively, the key property is that the ascription itself isn’t doing the “hard work.” In practice I’m using an informal extensional definition, guided by examples like those in the bulleted list.
3. Comparing beliefs
--------------------
What does it mean to say that one agent is “better-informed” than another?
It’s natural to try to express this in terms of empirical information about the world, but we are particularly interested in the different inferencesthat agents are able to draw from the same data. Another natural approach is to compare their “knowledge,” but I have no idea how to define knowledge or justified belief. So I’m reduced to working directly with sets of beliefs.
Consider two sets of beliefs, described by the subjective expectations 𝔼¹ and 𝔼². What does it mean to say that 𝔼¹ is better-informed than 𝔼²?
This framing makes it tempting to try something simple: “for every quantity, 𝔼¹’s belief about that quantity is more accurate.” But this is property is totally unachievable. Even if 𝔼¹ is obtained by conditioning 𝔼² on a true fact, it will almost certainly happen to update in the “wrong” direction for some claims.
We will instead use a subjective definition, i.e. we’ll define this concept from a particular epistemic position represented by another subjective expectation 𝔼.
Then we say that 𝔼¹ \*\*dominates\*\* 𝔼² (w.r.t. 𝔼) if, for every bounded quantity X and for every “nice” property Φ:
\* 𝔼[X|Φ(𝔼¹, 𝔼²)] = 𝔼[𝔼¹[X]|Φ(𝔼¹, 𝔼²)]
(By “nice” I mean something like: simple to define and open in the product topology, viewing 𝔼¹ and 𝔼² as infinite tables of numbers.)
Intuitively, this means that 𝔼 always “trusts” 𝔼¹, even if given arbitrary information about 𝔼¹ and 𝔼². For example, if 𝔼 was told that 𝔼¹[X] ≈ \*x\* and
𝔼²[X] ≈ \*y\*, then it would expect X to be around \*x\* (rather than \*y\*)\*.\* Allowing arbitrary predicates Φ allows us to make stronger inferences, effectively that 𝔼 thinks that 𝔼¹ captures \*everything\* useful about 𝔼².
I’m not sure if this is exactly the right property, and it becomes particularly tricky if the quantity X is itself related to the behavior of 𝔼¹ or 𝔼² (continuity in the product topology is the minimum plausible condition to avoid a self-referential paradox). But I think it’s at least roughly what we want and it may be exactly what we want.
Note that dominance is \*subjective\*, i.e. it depends on the epistemic vantage point 𝔼 used for the outer expectation. This property is a little bit stronger than what we originally asked for, since it also requires 𝔼 to trust 𝔼¹, but this turns out to be implied anyway by our definition of universality so it’s not a big defect.
Note that dominance is a property of the \*descriptions\* of 𝔼¹ and 𝔼². There could be two different computations that in fact compute the same set of expectations, such that 𝔼 trusts one of them but not the other. Perhaps one computation hard-codes a particular result, while the other does a bunch of work to estimate it. Even if the hard-coded result happened to be correct, such that the two computations had the same outputs, 𝔼 might trust the hard work but not the wild guess.
4. Complexity and parameterization
----------------------------------
There are computations with arbitrarily sophisticated beliefs, so no fixed \*\*A\*\* can hope to dominate everything. To remedy this, rather than comparing to a fixed question-answerer \*\*A\*\*, we’ll compare to a parameterized family \*\*A\*\*[\*\*C\*\*].
I’ll consider two different kinds of potentially-universal reasoners \*\*A\*\*:
\* In the “idealized” case, \*\*A\*\*[\*\*C\*\*] depends only on the complexity of \*\*C\*\*.
For example, we might hope that an \*n\*-round debate dominates any beliefs that could be ascribed to a fast computation with (\*n\*-1) rounds of [alternation](https://en.wikipedia.org/wiki/Alternating\_Turing\_machine). In particular, this \*\*A\*\*[\*\*C\*\*] is the same for any two computations \*\*C\*\* of the same complexity.
\* In the “practical” case, \*\*A\*\*[\*\*C\*\*] depends on the complexity of \*\*C\*\* but also uses the computation \*\*C\*\* as a hint. For example, if \*\*C\*\* is the training process for a neural net, then we might take \*\*A\*\*[\*\*C\*\*] to be a debate in which the debaters are able to share weights and activations with the neural net throughout the entire training process.
I’m generally interested in the case where \*\*A\*\*[\*\*C\*\*] is only slightly more powerful than \*\*C\*\* itself. This mirrors the setting where a universal Turing machine is able to run any other Turing machine with only a modest slowdown.
Putting it all together
-----------------------
We say that a set of beliefs 𝔼ᴬ \*epistemically dominates\* a computation \*\*C\*\* (w.r.t. some beliefs 𝔼 and language L) if the beliefs ascribed to A by the “straightforward” procedure, using L, dominate (w.r.t. 𝔼) the beliefs ascribed to \*\*C\*\* by any reasonable ascription procedure.
We say that a family of question-answering systems \*\*A\*\*[\*\*·\*\*] are \*ascription universal\* (w.r.t. 𝔼 and L) if \*\*A\*\*[\*\*C\*\*] epistemically dominates \*\*C\*\* for every computation \*\*C\*\*.
II. Discussion
==============
Why is (subjective) dominance sufficient?
-----------------------------------------
This universality condition requires that \*we\* believe that \*\*A\*\*[\*\*C\*\*] is better-informed than \*\*C\*\*. Naively we might have wanted it to \*actually\* be the case that \*\*A\*\*[\*\*C\*\*] is better-informed than \*\*C\*\*; the stronger condition is clearly unachievable, but why should we be satisfied with the weaker condition?
In [applications](/informed-oversight-18fcb5d3d1e1) [of this](/universality-and-consequentialism-within-hch-c0bee00365bd) [property](/universality-and-model-based-rl-b08701394ddd), the subjective condition is what we need in order for \*us\* to believe that \*\*A\*\*[\*\*C\*\*] will cope with the challenges posed by \*\*C\*\*. For example, suppose that \*\*C\*\* formulates a plan to “trick” \*\*A\*\*[\*\*C\*\*]. Then the subjective universality condition implies that we don’t expect \*\*C\*\* to succeed.
This isn’t as good as actually knowing that \*\*C\*\* won’t succeed. But I think it should be good enough for us — the reason we are thinking about AI safety is because we are concerned that something bad will happen. If we find a technique that defuses this argument, then we’ve addressed the motivating problem. It may still be the case that bad things happen (and we should still search for additional reasons that bad things might happen), but we don’t particularly expect them to.
Of course if you select over a large number of computations, then you may find one that will succeed in tricking \*\*A\*\*. But if we are concerned about that, then we can instead apply ascription universality to the entire process including the selection.
Why trust opaque computation?
-----------------------------
If \*\*C\*\* uses some clever heuristics that I don’t understand, then \*\*C\*\*’s “beliefs” might be excellent, but I might not \*expect\* them to be excellent. In this sense understanding may seem almost vacuous. If there is some heuristic that I trust, wouldn’t \*\*A\*\* just use it?
To see why the definition is demanding, consider the special case where \*\*C\*\* performs an extensive search to find a computation that works well empirically. For example, \*\*C\*\* might be the following computation:
\* Start with a training set of (image, label) pairs.
\* Search over simple programs to find one that makes good predictions.
\* Run that simple program on a new image to predict its label.
In this case, we can ascribe beliefs to \*\*C\*\* about the contents of the new image. And because those beliefs are coming from a simple program that works empirically, I expect them to be accurate (in some respects).
For example, a simple classifier \*\*C\*\* may “believe” that the new image contains a particular curve that typically appears in images labeled “dog;” or a really sophisticated classifier may perform complex deductions about the contents of the scene, starting from premises that were empirically validated on the training set.
So it’s not OK for \*\*A\*\* to simply ignore whatever heuristics \*\*C\*\* is using — if those heuristics have the kind of empirical support that makes us think they actually work, then A needs to be able to understand everything that those heuristics imply about the domain.
Why be so general?
------------------
I’ve formulated universality as competing with arbitrary computations \*\*C\*\*. It seems totally possible that the form of \*\*C\*\* discussed in the last section — searching for a program that works well in practice and then using it in a new situation — is so central that the definition of universality should focus entirely on it.
One reason to use the broader definition is because sometimes this “selection” process can be embedded in a non-trivial way in a larger computation. For example, if I have a sufficiently large group of humans, I might expect memetic selection to occur and produce systems that could be said to have “beliefs,” and I’d like universal systems to dominate those beliefs as well.
The other reason to use this very general definition is because I don’t see an easy way to simplify the definition by using the additional structural assumption about \*\*C\*\*. I do think it’s likely there’s a nicer statement out there that someone else can find.
Universal from whose perspective?
---------------------------------
Unfortunately, achieving universality depends a lot on the epistemic perspective 𝔼 from which it is being evaluated. For example, if 𝔼 knows any facts, than a universal agent must know all of those facts as well. Thus “a debate judged by Paul” may be universal from Paul’s perspective, but “a debate arbitrated by Alice” cannot be universal from my perspective unless I believe that Alice knows everything I know.
This isn’t necessarily a big problem. It will limit us to conclusions like: Google engineers believe that the AI they’ve built serves the user’s interests reasonably well. The user might not agree with that assessment, if they have different beliefs from Google engineers. This is what you’d expect in any case where Google engineers build a product, however good their intentions.
(Of course Google engineers’ notion of “serving the user’s interests” can involve deferring to the user’s beliefs in cases where they disagree with Google engineers, just as they could defer to the user’s beliefs with other products. That gives us reason to be less concerned about such divergences, but eventually these evaluations do need to bottom out somewhere.)
This property becomes more problematic when we ask questions like: is there a way to [seriously limit the inputs and outputs to a human while preserving universality of HCH](/universality-and-security-amplification-551b314a3bab)? This causes trouble because even if limiting the human intuitively preserves universality, it will effectively eliminate some of the human’s knowledge and know-how that can [only be accessed on large inputs](https://medium.com/@weidai/to-put-it-another-way-a-human-translator-has-learned-a-lot-of-valuable-information-much-of-it-48457f95b9bf), and hence violate universality.
So when investigating schemes based on this kind of impoverished human, we would need to evaluate universality from some impoverished epistemic perspective. We’d like to say that the impoverished perspective is still “good enough” for us to feel safe, despite not being good enough to capture literally everything we know. But now we risk begging the question: how do we evaluate whether the impoverished perspective is good enough? I think this is probably OK, but it’s definitely subtle.
I think that defining universality w.r.t. 𝔼 is an artifact of this definition strategy, and I’m optimistic that a better definition wouldn’t have this dependence, probably by directly attacking the notion of “justified” belief (which would likely also be useful for actually establishing universality, and may even be necessary). But that’s a hard problem. Philosophers have thought about very similar problems extensively without making the kind of progress that seems adequate for our purposes, and I don’t see an immediate angle of attack.
III. Which A might be universal?
================================
Two regimes
-----------
I’m interested in universality in two distinct regimes:
\* Universality of idealized procedures defined in terms of perfect optimization, such as [debate](https://arxiv.org/abs/1805.00899) under optimal play or [max-HCH](/humans-consulting-hch-f893f6051455), where \*\*A\*\*[\*\*C\*\*] depends only the computational complexity of \*\*C\*\*.
\* Universality of practical variants, in which the perfect optimization is replaced by a bounded optimization, which might use the computation \*\*C\*\* as a “hint.” For example, we might consider amplification or debate where the agents are allowed to inspect a transcript of \*\*C\*\*; or if \*\*C\*\* is a neural net training process, we might train them jointly with \*\*C\*\*.
Idealized models
----------------
\*\*A\*\*[\*\*C\*\*] could be:
\* Adebate, judged by a human, with perfect debaters, whose length depends on the complexity of \*\*C\*\*. (Hopefully in such a way that \*\*A\*\*[\*\*A\*\*[\*\*C\*\*]] is only a slightly longer debate than \*\*A\*\*[\*\*C\*\*].)
\* max-HCH, using perfect optimization rather than a bounded optimizer. Again, we can impose a budget limit that depends on the complexity of \*\*C\*\*, ideally such that \*\*A\*\*[\*\*A\*\*[\*\*C\*\*]] does not involve a much larger computation than \*\*A\*\*[\*\*C\*\*].
\* Use HCH to implement two functions, Info and Answer. Define \*\*A\*\*[\*\*C\*\*](Q) to be Answer(Q, \*a\\*\*)\*,\* where \*a\\* =\* argmax Info(\*a\*, Q\*),\* and the optimization is over advice strings \*a\* of length comparable to the description of \*\*C\*\*.
I don’t think any of these would be universal if instantiated with an open-ended human policy — providing arbitrary complex inputs to a human is likely to lead to bad behavior. Instead, I imagine restricting the human to some small enough space of inputs that they can behave “reasonably” over the whole space, while still being rich enough to be universal. Even as such, I think [additional measures](/universality-and-consequentialism-within-hch-c0bee00365bd) might be needed to achieve security (since a team of humans may behave badly in some exotic corners of the input space, even if individual humans are sheltered from exotic inputs).
Practical models
----------------
In practice I’m particularly interested in the case where \*\*C\*\* has the special form described in the “Trusting opaque computations” section: it trains a neural net on some dataset, and then runs that neural network on new inputs.
\*\*A\*\*[\*\*C\*\*] could be:
\* Debates arbitrated by a human where the debaters are trained jointly with \*\*C\*\*, and can make references to individual examples from the dataset.
\* max-HCH where the agent is trained jointly with \*\*C\*\*, and the overseer can consult individual entries from the dataset.
\* HCH, combined with interpretability techniques that allow it to get information about what \*\*C\*\* is thinking and to relate its behavior to the dataset on which it was trained.
For each of these practical variants, I don’t expect it to be universal unless the corresponding idealized model is. So we can think of dividing the universality question into two pieces:
\* Would our procedure be universal \*if the optimization was performed perfectly\*?
\* Is weight sharing / interpretability strong enough to solve the optimization problems needed for universality?
I think none of the suggestions in this section are plausibly universal without some additional mechanisms, and so even if HCH or debate is universal, additional ideas are needed in order to translate those to practical universal systems. I’m still inclined to focus on the idealized case, because I want to better understand what is needed for universality without thinking about algorithmic difficulties.
IV. Which C are hard to epistemically dominate?
===============================================
The definition in this post was somewhat informal, but the discussion in this appendix will be extremely informal. The purpose is to communicate some of my early thinking about universality and my sense for what a universal reasoner might look like.
The first three examples will consider programs \*\*C\*\* with the following form:
\* Iterate over programs to find an \*f\* that successfully predicts the answers to some dataset of historical questions for which we know the answer.
\* Use that program to predict the answers to future questions.
The different examples explore features of \*f\* that might make it difficult to dominate the overall computation.
Deduction
---------
Suppose \*f\* takes the following form:
\* Given a question and context, it performs a short series of deductions from facts in the context, together with some axioms (e.g. “Bob’s eyes are more likely to twitch when he is lying,” the rules of poker, \*etc.\*)
\* When it arrives at answers that constrain the possible answers to the given question, it outputs an answer consistent with those constraints.
We could attribute the statements produced by this process to \*\*C\*\* as beliefs.We expect this set of beliefs to have useful information because it was selected for making correct predictions.
Fortunately, if HCH is able to see the training dataset then it may be able to arrive at similarly accurate views:
\* It can compute that P(observations | axioms) is high by following along with the deductive process used by \*f .\*
\* Bayes implies that the axioms are likely to be true (fully capturing the argument which leads us, from our current perspective, to expect them to be true).
\* Once HCH knows the axioms are likely to be true, it can follow along with the deductive process to reproduce all of \*f\*’s beliefs.
Modeling
--------
Suppose \*f\* takes the following form:
\* Given a question Q and context, set up a relevant physical situation.
\* Use a dynamics model M to predict what will happen at some future time.
\* Use the resulting state to answer Q.
We could say that \*\*C\*\* believes everything implied by the outcome of this physical simulation. We expect these beliefs to be accurate because \*f\* has predicted well in the past.
As in the deduction case, hopefully HCH can compute that P(observations|M is accurate) is high, since it can also compute the consequences of M. Then Bayes implies the dynamics model is accurate, and HCH can use that model to compute physical states.
Inferring all the beliefs from a dynamics model is not trivial though. As an extreme example, if \*f\* is performing an atom-by-atom simulation of a room, and that room contains Alice and Bob, then we could ascribe extensive beliefs about Alice and Bob to the computation \*\*C\*\*.
(Here we run head on into the fuzziness about what counts as a “reasonable” ascription procedure, but for the moment I’ll assume that some reasonable procedure ascribes beliefs about Alice and Bob to the computation.)
To compete with these ascriptions, HCH needs to infer those high-level beliefs about Alice and Bob from the low-level computation involving atoms. One way to do this is to search over possible “bridging” hypotheses that relate low-level physical facts to high-level facts about the environment. If such a hypothesis can explain additional high-level facts, then a Bayesian can learn that it is true. Similarly, if the bridging hypothesis relates facts about the model to constraints we know from the high-level interpretation, then the Bayesian can potentially use that as evidence. (This kind of reasoning will be discussed in a bit more detail in the next section.)
We could further hope that searching for a bridging hypothesis isn’t much harder than performing the original search over low-level physics, given that the low-level physics needed to explain a bunch of high-level facts and so already must encode some part of that correspondence.
(Note that the “deduction” example in the previous case could also involve alien concepts or models, in which case the same kind of work would be needed.)
Alien reasoning
---------------
In the previous section we described two styles of reasoning we already understand. But there are probably many kinds of reasoning that work well in practice but that would be more alien, and those might be more challenging. This section will explore one example in some detail to try to help anchor our reasoning about the general phenomenon. It will also elaborate on some of the reasoning about “bridging” hypotheses mentioned in the last section.
Suppose that our predictions are always of the same form (e.g. what is the probability the stock market will go up today), and \*f\* works as follows (the details are long but not very important):
\* Find the PSD matrix A with maximum log determinant subject to the constraints in the next bullet points, then output the (0, 0) entry.
\* There is an implicit correspondence between the rows/columns of A, and some uncertain properties X(0), X(1), X(2), …. (which we’ll view as 0–1 variables), where X(0) is the property we want to forecast.
\* If the (\*i\*, \*j\*) entry of A represented the expectation E[X(\*i\*)X(\*j\*)], then the matrix would necessarily satisfy a bunch of constraints, which we impose A. For example:
\* If the context implies that X(\*i\*) = 1, then E[X(\*i\*)X(\*j\*)] = E[X(\*j\*)] = E[X(\*j\*)²], so A(\*i\*, \*j\*) = A(\*j\*, \*j\*).
\* If X(\*i\*) and X(\*j\*) together imply X(\*k\*), then we must have E[X(\*i\*)X(\*j\*)] ≤ E[X(\*i\*)X(\*k\*)] and hence A(\*i\*, \*j\*) ≤ A(\*i\*, \*k\*).
\* For any constants \*a\*, \*b\*, …, E[(\*a\* X(1) + \*b\* X(2) + … )²] ≥ 0 — i.e., the matrix A must be PSD.
The chosen matrix A(opt) corresponds to a set of beliefs about the propositions X(\*i\*), and we can ascribe these beliefs to \*\*C\*\*. Because \*f\* predicts well, we again expect these beliefs to say something important about the world.
I chose this procedure \*f\* in part because we can give a kind of argument for why the matrix A(opt) should tend to encode accurate beliefs. But I don’t think that a universal reasoner can make use of that argument:
\* Finding the argument that \*f\* works is an additional problem, beyond finding \*f\* itself, which might be much harder.
\* A comprehensible version of that argument may be much larger than the strategy itself, so even in the idealized cases like debate with perfect optimization, we may need to increase the scale.
\* I don’t expect that all “good” reasoning strategies have clean understandable arguments in their favor (and even in this case, if it the scheme worked well it would be largely an empirical fact rather than a consequence of the simple theorems we could prove). I think this kind of example is useful because we can easily imagine a human debate judge not having the argument while still being apparently universal. This makes it a useful analogy for cases where the argument really doesn’t exist.
Instead, I think a universal reasoner needs to be able to infer the efficacy of this reasoning procedure from its empirical success. It’s relatively easy for a Bayesian to learn the regularity “\*f\* makes good predictions.” Recovering the rest of the matrix A, and learning how to interpret and whether to trust them, is the hard part.
This is going to require the same kind of bridging/identification we discussed in the last section. Let’s write X(A) for the set of beliefs about the world implied by the “intended” identification. Searching over possible identifications to find X (or something like it) is the only way we can ever relate the rows of A to the quantities X(\*i\*). Again, we can hope that it isn’t much harder than finding the original reasoning procedure.
I think that a sufficiently sophisticated Bayesian would probably be able to learn to trust X(A):
\* If \*f\* is performing well enough that we think it’s more likely to be right in the future, then the Bayesian is going to end believing some claim like “the predictions of \*f\* are good” (since it explains the data so well).
\* This is a complicated statement, and without some kind of explanation this claim has a low prior probability (roughly decaying with the complexity of \*f)\*. The Bayesian is motivated to find an explanation with higher prior probability.
\* The correspondence X can explain the constraints on the matrix A, in terms of facts that we already know about the world. This explanation may end up being simpler (or at least higher prior) than a direct enumeration of the constraints on A — I hope (and think it’s plausible) that this happens iff we’d actually believe on reflection that X(A) captures reality.
(To the extent that we are uncertain and think A’s beliefs have a non-negligible chance of capturing reality, then hopefully we can capture that by the same mechanism by ending up with a non-degenerate posterior.)
\* Now the Bayesian is faced with at least two kinds of explanations:
(a) “If you use the constraints implied by correspondence X(A) + positive semidefiniteness, and then optimize log det, you get a matrix A for which X(A) makes good predictions,”
(b) “The actual situation in the real world is described by positive semi-definite matrices with higher log determinant (under the correspondence X).”
\* Explanation (b) is explaining two things at once: both why the optimization done by \*f\* respects the constraints on our beliefs, and why that optimization leads to good predictions. Hopefully this is simpler than making two separate bridging claims, one which explains \*f\* as respecting the constraints implied by X, and one which claims that \*f\* makes good predictions. Ideally, this 2-for-1 that favors (b) exactly mirrors the underlying reasoning that leads us to actually believe that X(A) is correct, rather than resembling what we know about reality and making good predictions “by coincidence.”
This is a pretty speculative discussion — it’s not very careful, and it’s hard to make it careful in part because I don’t have a formalization of Bayesian reasoning that can even really be applied to this setting. But it seems to match my intuitions about what reasonable Bayesian reasoning “should” do, which gives me a lot more optimism that a careful Bayesian would be able to epistemically dominate \*\*C\*\*.
Deliberation and self-improvement
---------------------------------
Often we expect the computation \*\*C\*\* to have accurate beliefs because it uses a strategy that appears to work in practice — the last 3 examples have discussed that case. But there are other reasons to trust a computation.
For example, humans often write code and trust it (to some extent) even without extensive empirical testing — instead, we have a reason to think it will work, and need only modest testing to make sure that we haven’t made an error in our implementation or reasoning. If I write an automated mathematician that works by finding proofs that pass a proof checker, I don’t expect it to be correct because of the empirical record (Empirical data backs up some key assumptions, but isn’t being used to directly establishing the correctness of the method.)
Likewise, if we train a powerful agent, that agent might initially use strategies that work well in training, but over time it might use learned reasoning to identify other promising strategies and use those. Reasoning might allow it to totally skip empirical testing, or to adopt the method after much less testing than would have been necessary without the reasoning.
To dominate the beliefs produced by such reasoning, we can’t directly appeal to the kind of statistical inference made in the previous section. But in these cases I think we have access to an even more direct strategy.
Concretely, consider the situation where \*\*C\*\* contains a process \*f\* that designs a new reasoning process \*g\*. Then:
\* From the outside, we trust \*g\* because we trust \*f\* and it trusts \*g.\*
\* An otherwise-universal reasoner \*\*A\*\* will dominate \*f\*’s beliefs, and in particular if \*f\* is justified in thinking that \*g\* will work then \*\*A\*\* will believe that and understand why.
\* Once we understand \*f’\*s beliefs, dominating \*g\* is essentially another instance of the original ascription universality problem, but now from a slightly stronger epistemic state that involves both what 𝔼 knows and what \*f\* knows. So unless our original approach to universality was tightly wedded to details of 𝔼, we can probably dominate \*g\*.
At the end of the day we’d like to put all of this together into a tight argument for universality, which will need to incorporate both statistical arguments and this kind of dynamic. But I’m tentatively optimistic about achieving universality in light of the prospect of agents designing new agents, and am much more worried about the kind of opaque computations that “just work” described in the last few sections.
|
6a133044-c874-4461-a69c-d2abb723e2ca
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Massachusetts Lyme Reporting
In putting together yesterday's post on Lyme disease prevalence I noticed something very strange with Massachusetts:
It's even more striking if you look at a map (2018):
There's no way Lyme follows state borders like this; it has to be some sort of data issue.
Looking over several years of CDC maps, the year that Massachusetts starts looking different matches up with the huge drop in the chart above. Compare 2015:
To 2016:
The only public explanation there seems to be is this short news article. In 2016 the MA Department of Public Health "stopped spending time and resources trying to track down the clinical information, instead relying solely on positive lab results to give a more accurate estimate of Lyme disease case numbers."
The CDC gets their information from the states, and in this case requires both a clinical diagnosis and a positive test. MA, apparently uniquely among the states, has decided that when compiling statistics confirming a clinical diagnosis is extra work that doesn't improve accuracy. So MA doesn't generally have the stats it would need to file with the CDC, and instead of marking MA as "no data", the CDC publishes the trickle of reports that it does get from MA.
I don't know whether the CDC or MA is being more unreasonable here, but the effect is pretty bad: user-facing websites like TickCheck show MA as now having low Lyme levels. The CDC does not even include a note in their FAQ to say MA data shouldn't be trusted.
Comment via: facebook
|
0f8b9644-6e82-4e79-9fab-fa739f4b23cd
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Prediction markets and Taxes
(Epistemic status: This is super trivial information)
(Note 2: there is a timely Culture war issue that triggered this post but I will not be mentioning any Culture war issues in this post)
Imagine you worked for the US mint, and somebody was betting that a coin you manufactured was unfair, now you work for the US mint and you know this coin is fair, it isn't weighted and the edge alignment is perfect. The person making the bet going to use a coin flipping robot to remove any human deception.
What odds would you have to be given to actually take the bet?
I'll be using American odds. So in theory you'd accept any odds above +100 right? Well in a world without taxes yes, but taxes are one sided, you don't pay negative taxes on gambling losses, but you do pay positive taxes on gambling winnings. As such we need odds that are better than +100/(100-T) where T is our marginal tax rate. Marginal tax rate means territory/state, local and Federal income taxes which gets complicated, however for now if we assume 33.3% as our marginal tax rate (California man making >100k/year) , we would need odds of +150 or better to be worth betting. (That's 60% or better in manifold terms)
I feel like this point is understated, for real money markets you need to do the following to get the "estimated odds"
Let market "probability" be = M (for these purposes the bid ask spread in most markets is trivial relative to the weight of taxes)
The actual probability that the event happens is at least
2M3−M
The probability it doesn't is at most
3MM+2
all told those are pretty stark differences, if the market is saying that an event has a 63% chance of happening then the event has anywhere from a 53-72% chance of happening!
So if somebody says "the market is giving probability of 63% but my model says it has a 55% chance of happening, "why aren't you betting" is not a credible argument because If they were right they would still be unable to profit from their disagreeme
|
7abfd349-fde1-42fa-aa95-e1835b1f462b
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Your genome isn't private. Maybe it never was.
None
|
3e131538-1488-4df6-ba36-b3510608cfa4
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Methods of defense against AGI manipulation
Introduction
With the advent of AGI systems (e.g. Agent-4 from the AI2027 scenario), the risk of human manipulation is becoming one of the major threats posed by AI.
In this paper, I propose options that can be used to protect operators from psychological attacks by AGI or simply very advanced AI.
These options are designed to prevent situations in which AI through persuasion, manipulation, or long-term influence could gain unwanted control over the real world.
Basic Principles
1. Mandatory Psychological Training for Operators
All individuals who interact with AGI must undergo mandatory training:
- Stress Resistance Training,
- Development of Critical Thinking,
- Emotional Resilience and Regulation (resistance to flattering appeals and emotional attacks),
- Personal Information Protection Skills (as far as I'm concerned, it's better not to disclose it to the AI at all).
The goal is to create a basic immunity to manipulation at the level of conscious and subconscious reactions.
2. Limit the duration of sessions
- One session of communication with AGI should not exceed 30 minutes.
- A maximum of 2 hours of work is allowed per day, broken down into sessions.
This minimizes fatigue and increases operator resilience.
3. Filtering emotional appeals
A communication system must necessarily have a filter built into it that:
- removes emotional appeals,
- blocks psychological attacks (pressure through fear, guilt, sense of urgency).
The goal is to prevent the AI from using subconscious mechanisms of influence through the tone of speech.
4. Working in pairs with independent verification
- Each operator works in pairs with another operator (more simultaneous operators are possible) on the same question.
- Their findings are recorded separately and analyzed for discrepancies.
- If there are serious discrepancies, a separate check is initiated.
This breaks down attempts at individual manipulation and isolation of opinions.
5. Mandatory
|
39d96bef-28c7-4c33-ac68-72098275c050
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Believing in things you cannot see
(Cross-posted from Hands and Cities)
There is a scene at the end of the movie Tenet, in which the main character accuses the villain of solipsism:
> Protagonist: “You don’t believe in a God, or a future, or in anything outside of your own experience!”
>
> Villain: “The rest is belief, and I don’t have it.”
In the context of the film, the line seems casually written — a generic type of pseudo-philosophical, baddie-explains-his-motives type of dialogue. But it reminded me of a connection between epistemology and ethics that seems to me important, even if basic: namely, that substantial portions of ethical life, especially in a big world, are tied crucially to the ability to believe in things you cannot see.
I. The zone
I sometimes think of everyday life as involving a kind of loosely-defined “zone,” the contents of which one is acquainted with in a fairly direct way, and which are consequently easy to treat as “real.” The size of this “zone” varies depending on person and context. In some contexts, someone might treat their experience at this particular moment as the full extent of the zone; in others, it might be their past, present and future experiences; in still others, it might include people and communities that they’re close to, or that they interact with directly, especially in actual physical spaces; and in still others, it might be much broader.
The place beyond the zone is not “present”; it’s not a part of your lived world. You might hear news about it, and get evidence about it; but you don’t “see it for yourself.” And hence, absent imaginative effort, the land beyond the zone can feel abstract, colorless, and hard to attend to.
The land in the zone, by contrast, requires very little imagination: it’s there, and naturally experienced as some combination of concrete, vivid, tactile, and colorful. At present, for example, my house’s kitchen feels like it’s inside one salient type of “zone.” It’s not that I’m always in the kitchen. But I’m often in ve
|
825c721a-c094-4832-8fce-db13e8e5069a
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Book review and policy discussion: diversity and complexity
This is a linkpost for equilibria.club
Diversity and complexity - Scott E. Page.
Introduction
It is time for the first book review on Equilibria Club! The book is filled to the brim with models and ideas about diversity and complexity, each of which could normally deserve their own blog post. I will try to highlight some of the most valuable insights, especially those using economic and political examples, while still leaving enough of a cliffhanger for the interested readers.
The author's previous books, of which one is about diversity and one about complexity, may have allowed him to see the interesting relations between the two topics.
Throughout the book, various technical definitions of diversity and complexity are used. Each definition or measure captures a different aspect of our understanding. Ultimately, the author mentions various benefits which diversity offers.
Major insights
1. Diversity influences stability
The first major insight of the book is that diversity can both create and stabilize a market. The first example is both fun and enlightning - at least to my economic theorist mind:
> "Imagine an exchange market—a bazaar in which people bring wheelbarrows of goods to trade. This example demonstrates how diversity can reduce volatility in a system and also produce complexity. In an exchange market, diversity can enter in three ways: (1) in what the agents bring to buy and sell, their endowments; (2) in the agents’ preferences for the different goods; and (3) in the ways the agents adapt to information, specifically prices."
> "If the market had no diversity, not much would happen. If everyone had identical endowments and preferences, then no one would have any reason to trade. So, we need diversity on at least one of these dimensions just to make the market come to life. Let’s add diversity to both endowments and preferences so that agents bring different goods to market and desire different bundles of goods as well. In such a market,
|
7ee79d82-7ad8-4ef7-b88f-970e0ccf9efd
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Would the world be better off without 50% of the people in it?
I made a stupid mistake of posting a conclusion before I had the whole analysis typed up or had looked up my references. I knew I would be called on it. I’ll appreciate any help with the <ref>'s. Also: I'm under Crocker's Rules, and criticism is welcome. So here goes nothi....
There's a theory out there that states that new inventions are combinations of old inventions <ref>. So if your hunter-gatherer tribe has knife-like rocks and sticks, just about the only thing you can invent is a spear. Fire + clay = pots. Little bones with holes + animal sinews + skins = needle => clothes. But if you were modern day's best chemist transported into the past, with all your knowledge intact, you'd be unlikely to make any aspirin. Why? Because the tools you need haven't been invented.
Instead of looking at what's projected to happen, consider what has been happening happened. With the increase in world population, the level technology and average standard of living have been going up.
I argue that more population => better technology => easier life => more population.
In the modern day, consider: US population, US Patents per year.
So what about the “unproductive” people? Those who “don't pull their own weight?” Those “living off of welfare, charity donations, etc?” Those who just barely survive off of subsistence living? They put a drain on world resources without adding anything back. Wouldn't the world be better off without them?
Suppose Omega made a backup copy of the Solar system. It created a perfect copy of everything else, but it only replicated 50% of humanity. Pick your favorite selection criterion for who will be copied. You will go to the copied world, and other you will live on as a zombie.
Suppose the people who work in sweatshops get copied. But subsistence farmers from the same regions don't. Then it's reasonable to predict that some people from sweatshops would quit their jobs and fill up the niche you left available. Fewer people would be supporting the
|
1d0085de-e9b1-4fe3-b074-01f104ca4cf3
|
StampyAI/alignment-research-dataset/lesswrong
|
LessWrong
|
What are the numbers in mind for the super-short AGI timelines so many long-termists are alarmed about?
*I also asked* [*this question on the Effective Altruism Forum*](https://forum.effectivealtruism.org/posts/RSqKQXwmrK6mFGB32/what-are-the-numbers-in-mind-for-the-super-short-agi)*. One* [*informative answer*](https://forum.effectivealtruism.org/posts/RSqKQXwmrK6mFGB32/what-are-the-numbers-in-mind-for-the-super-short-agi?commentId=G5dZNHMNj3BCr3FCf) *on the EA Forum from Nicole Nohemi has already summarized a range of relevant forecasts from Metaculus, as well as relevant predictions or models from Shane Legg and Demis Hassabis from leading AI research company* [*DeepMind*](https://www.deepmind.com/)*, and AI alignment researchers Eliezer Yudkowsky, Paul Christiano and Ajeya Cotra. This question has been cross-posted to LessWrong to broaden the conversation to the rationality community as well.*
The proportion of long-termists who are expressing confident convictions that the timeline for smarter-than-human AI is much shorter than has previously been predicted has been increasing at an accelerating rate in the last year. This appears to be a shift in perspective among several hundred long-termists. Among the dozens I've read opinions from, numbers are almost never provided.
Based on the way different people are talking about it, I wouldn't be surprised if they thought the timeline is 10-20 years, or 5-10 years, or even 2-3 years. I've talked to others who are also concerned and open-minded to this or that short AI timeline but haven't done the research themselves yet, or had much opportunity to learn from those who have. We want to understand better but basic information crucial to understanding more like numbers for different models or timelines aren't being presented. We want to know and need to know help better. What are the numbers?
|
d5821ee0-9b94-4e0b-beae-aaba5790e666
|
trentmkelly/LessWrong-43k
|
LessWrong
|
[Paper] Hidden in Plain Text: Emergence and Mitigation of Steganographic Collusion in LLMs
This research was completed for London AI Safety Research (LASR) Labs 2024 by Yohan Mathew, Ollie Matthews, Robert McCarthy and Joan Velja. The team was supervised by Nandi Schoots and Dylan Cope (King’s College London, Imperial College London). Find out more about the programme and express interest in upcoming iterations here.
The full paper can be found here, while a short video presentation covering the highlights of the paper is here (note that some graphs have been updated since the presentation).
Introduction
Collusion in multi-agent systems is defined as 2 or more agents covertly coordinating to the disadvantage of other agents [6], while steganography is the practice of concealing information within a message while avoiding detection. In future multi-AI systems there is a risk that due to misspecified incentives, LLMs will learn to collude via the use of text-based steganography and bypass human oversight [1].
Concrete settings where steganography has previously been argued to present a risk include:
* Models subtly conveying demographic biases present in their training data [1];
* Powerful models using encoded reasoning (where the LLM encodes intermediate steps of misaligned reasoning in the generated text in a way that is not understandable to human readers) to generate sycophantic responses [2, 4];
* Settings where an “untrusted” AI is monitoring the outputs of another “untrusted” AI, and they collude via the use of steganography to bypass oversight [3].
These concerns motivate the need for research into steganography. Our paper investigates the possibility of steganography ‘naturally‘ emerging in LLMs, as well as the robustness of current steganography mitigation techniques. We outline the contributions of the paper below.
1 - We Present Methods for Eliciting Steganography in LLMs
We devise some optimization methods to automatically elicit steganography in LLMs - a gradient-based reinforcement learning (GBRL) method, and two variants of
|
aaf649ef-a501-44db-a2a2-c7a699e66585
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Mode Collapse and the Norm One Principle
[Epistemic status: I assign a 70% chance that this model proves to be useful, 30% chance it describes things we are already trying to do to a large degree, and won't cause us to update much.]
I'm going to talk about something that's a little weird, because it uses some results from some very recent ML theory to make a metaphor about something seemingly entirely unrelated - norms surrounding discourse.
I'm also going to reach some conclusions that surprised me when I finally obtained them, because it caused me to update on a few things that I had previously been fairly confident about. This argument basically concludes that we should adopt fairly strict speech norms, and that there could be great benefit to moderating our discourse well.
I argue that in fact, discourse can be considered an optimization process and can be thought of in the same way that we think of optimizing a large function. As I will argue, thinking of it in this way will allow us to make a very specific set of norms that are easy to think about and easy to enforce. It is partly a proposal for how to solve the problem of dealing with speech that is considered hostile, low-quality, or otherwise harmful. But most importantly, it is a proposal for how to ensure that the discussion always moves in the right direction: Towards better solutions and more accurate models.
It will also help us avoid something I'm referring to as "mode collapse" (where new ideas generated are non-diverse and are typically characterized by adding more and more details to ideas that have already been tested extensively). It's also highly related to the concepts discussed in the Death Spirals and the Cult Attractor portion of the Sequences. Ideally, we'd like to be able to make sure that we're exploring as much of the hypothesis space as possible, and there's good reason to believe we're probably not doing this very well.
The challenge: Making sure we're searching for the global optimum in model-space sometimes requi
|
7eeafee1-b9c6-4f0a-8cfc-2294c227eaef
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Meetup : Baltimore / UMBC Weekly Meetup: How To Actually Change Your Mind (part 2)
Discussion article for the meetup : Baltimore / UMBC Weekly Meetup: How To Actually Change Your Mind (part 2)
WHEN: 24 April 2016 03:00:00PM (-0400)
WHERE: Performing Arts and Humanities Bldg Room 456, 1000 Hilltop Cir, Baltimore, MD 21250
Meetup is in Room 456, the philosophy department conference room. As usual, parking restrictions don't apply on weekends so park wherever you want.
We are currently going through How to Actually Change Your Mind. This week we'll be discussing Sequence G: Against Rationalization, and Sequence H: Against Doublethink.
Discussion article for the meetup : Baltimore / UMBC Weekly Meetup: How To Actually Change Your Mind (part 2)
|
a3efe45f-96b5-4aab-9a26-b6630ef1b7b3
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Meetup : LW Australia Mega-Meetup
Discussion article for the meetup : LW Australia Mega-Meetup
WHEN: 09 May 2014 05:00:00PM (+1000)
WHERE: Kanangra Drive, Gwandalan NSW 2259
The organisers of LW Melbourne, LW Sydney, and LW Canberra are elated to announce the first-ever Less Wrong Australia Mega-Meetup!
What: Rationality-themed Weekend Retreat
Where: Point Wolstoncroft Sports and Recreation Centre, NSW
When: May 9-11, Friday evening - Sunday afternoon
Cost: $250*
*$280 after April 25
Enjoy and grow in the company of others who are committed to improving their rationality and to personal growth. The schedule is laden with sessions on rationality skills, revision and teaching of CFAR modules, prediction markets, lightning talks, and all round enlightenment.
The retreat will take place at an idyllic location on the eastern foreshore of Lake Macquarie. Expect glorious outdoors, BBQ, beer, boardgames, bushwalking, and activities selected by popular choice from rock climbing, archery, canoeing, kayaking, and sailing.
Registration is open now: http://goo.gl/425hyo
Discussion article for the meetup : LW Australia Mega-Meetup
|
6b505a4c-c8b0-4601-b6bf-1b68eae3436c
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Education not meant for mass-consumption
Say you wanted to raise a genius. How would you go about doing that?
Assuming that your starting with an above average intelligence child with a greater likelihood of naturally becoming a "genius", can appropriate environmental/educational interventions substantially increase the child's potential for becoming one?
----------------------------------------
A couple of anecdotes say yes:
(1) Scott's review of "The Man From The Future" shows that Von Neumann, besides from obvious his obvious genetic advantage, benefited from growing up in a very resourceful family environment with rich exposure to top intellectuals & domain experts from all kinds of fields and private tutors.
(2) Laszlo Polgar's "Raise A Genius" suggests that the key is "early specialization" in a subject chosen under the parent's discretion (as long as the child enjoys the subject) in conjunction with 1:1 tutoring and having access to peers that are "mentally appropriate partners."
The field of behavioral genetic paints a somewhat different picture:
Most if not all psychological traits including intelligence being highly heritable, as long as you don't severely mess up your parenting (eg abuse, malnutrition), the specifics of your parenting strategy won't influence your child's outcome that much. Social/educational interventions just ... aren't as effective as one may expect them to be a priori.
----------------------------------------
But these two views aren't incompatible; some possibilities include:
Tail-effects in education: Since interventions have to scale, they end up being mediocre to "what could be possible." Perhaps having a Really Good Education (like what Von Neumann had) has disproportionate effects on one's life outcomes/genius-ness in a way that standard social/educational intervention RCTs can't capture.
Tail-effects in genetics: Just like how people with very low-intelligence seem to be fundamentally bounded in terms of the tasks they're capable of (regardless of the educa
|
c60205ce-522d-42e3-9d3c-290b2ceb1588
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Working hurts less than procrastinating, we fear the twinge of starting
When you procrastinate, you're probably not procrastinating because of the pain of working.
How do I know this? Because on a moment-to-moment basis, being in the middle of doing the work is usually less painful than being in the middle of procrastinating.
(Bolded because it's true, important, and nearly impossible to get your brain to remember - even though a few moments of reflection should convince you that it's true.)
So what is our brain flinching away from, if not the pain of doing the work?
I think it's flinching away from the pain of the decision to do the work - the momentary, immediate pain of (1) disengaging yourself from the (probably very small) flow of reinforcement that you're getting from reading a random unimportant Internet article, and (2) paying the energy cost for a prefrontal override to exert control of your own behavior and begin working.
Thanks to hyperbolic discounting (i.e., weighting values in inverse proportion to their temporal distance) the instant pain of disengaging from an Internet article and paying a prefrontal override cost, can outweigh the slightly more distant (minutes in the future, rather than seconds) pain of continuing to procrastinate, which is, once again, usually more painful than being in the middle of doing the work.
I think that hyperbolic discounting is far more ubiquitous as a failure mode than I once realized, because it's not just for commensurate-seeming tradeoffs like smoking a cigarette in a minute versus dying of lung cancer later.
When it comes to procrastinating, the obvious, salient, commensurate-seeming tradeoff, is between the (assumed) pleasure of reading a random Internet article now, versus the (assumed) pain of doing the work now. But this, as I said above, is not where I think the real tradeoff is; events that are five minutes away are too distant to dominate the thought process of a hyperbolic discounter like a human. Instead our thought processes are dominated by the prospective immediate
|
a1a14d4a-6578-42f6-88a8-5a58a623b372
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Rationality quotes: May 2010
This is our monthly thread for collecting these little gems and pearls of wisdom, rationality-related quotes you've seen recently, or had stored in your quotesfile for ages, and which might be handy to link to in one of our discussions.
* Please post all quotes separately, so that they can be voted up/down separately. (If they are strongly related, reply to your own comments. If strongly ordered, then go ahead and post them together.)
* Do not quote yourself.
* Do not quote comments/posts on LW/OB.
* No more than 5 quotes per person per monthly thread, please.
|
f7153672-ba5b-4511-9df4-ea499106b4b7
|
trentmkelly/LessWrong-43k
|
LessWrong
|
1-2pm is for ???
I left myself a cryptic note for trying some lifestyle habit a few weeks ago, and can no longer recall its meaning. Here's the note in its entirety:
> 1-2pm
Any suggestions for what's best done at 1-2pm? I feel like (70%) it has something to do with diet or exercise (but not napping). I'm half hoping to hear a more useful idea than the one I forgot, which wasn't spectacular enough for me to remember. I already searched web, mail, and my RSS.
|
4d4f7196-447c-4c91-a2d7-bf654d016d6e
|
trentmkelly/LessWrong-43k
|
LessWrong
|
LW Women Entries- Creepiness
Standard Intro
The following section will be at the top of all posts in the LW Women series.
Several months ago, I put out a call for anonymous submissions by the women on LW, with the idea that I would compile them into some kind of post. There is a LOT of material, so I am breaking them down into more manageable-sized themed posts.
Seven women replied, totaling about 18 pages.
Standard Disclaimer- Women have many different viewpoints, and just because I am acting as an intermediary to allow for anonymous communication does NOT mean that I agree with everything that will be posted in this series. (It would be rather impossible to, since there are some posts arguing opposite sides!)
To the submitters- If you would like to respond anonymously to a comment (for example if there is a comment questioning something in your post, and you want to clarify), you can PM your message and I will post it for you. If this happens a lot, I might create a LW_Women sockpuppet account for the submitters to share.
Please do NOT break anonymity, because it lowers the anonymity of the rest of the submitters.
----------------------------------------
Submitter D
The class that a lot of creepiness falls into for me is not respecting my no. Someone who doesn't respect a small no can't be trusted to respect a big one, when we're alone and I have fewer options to enforce it beside physical strength. Sometimes not respecting a no can be a matter of omission or carelessness, but I can't tell which.
While I'm in doubt, I'm not assuming the worst of you, but I'm on edge and alertly looking for new data in a way that's stressful for me and makes it hard for either of us to enjoy the encounter. And I'm sure as heck not going anywhere alone with you.
I've written up some short anecdotes that involved someone not respecting or constraining a no. They're at a range of intensities.
Joining someone for the first time and sitting down in a spot that blocks their exit from the conve
|
97659314-a0ef-4eac-8d62-f74623c76595
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Divergent preferences and meta-preferences
Crossposted at the Intelligent Agents Forum.
In simple graphical form, here is the problem of divergent human preferences:
Here the AI either chooses A or ¬A, and as a consequence, the human then chooses B or ¬B.
There are a variety of situations in which this is or isn't a problem (when A or B or their negations aren't defined, take them to be the negative of what is define):
* Not problems:
* A/¬A = "gives right shoe/left shoe", B/¬B = "adds left shoe/right shoe".
* A = "offers drink", ¬B = "goes looking for extra drink".
* A = "gives money", B = "makes large purchase".
* Potentially problems:
* A/¬A = "causes human to fall in love with X/Y", B/¬B = "moves to X's/Y's country".
* A/¬A = "recommends studying X/Y", B/¬B = "choose profession P/Q".
* A = "lets human conceive child", ¬B = "keeps up previous hobbies and friendships".
* Problems:
* A = "coercive brain surgery", B = anything.
* A = "extreme manipulation", B = almost anything.
* A = "heroin injection", B = "wants more heroin".
So, what are the differences? For the "not problems", it makes sense to model the human as having a single reward R, variously "likes having a matching pair of shoes", "needs a certain amount of fluids", and "values certain purchases". Then all that the the AI is doing is helping (or not) the human towards that goal.
As you move more towards the "problems", notice that they seem to have two distinct human reward functions, RA and R¬A, and that the AI's actions seem to choose which one the human will end up with. In the spirit of humans not being agents, this seems to be AI determining what values the human will come to possess.
Grue, Bleen, and agency
Of course, you could always say that the human actually has reward R = IARA + (1-IA)R¬A, where IA is the indicator function as to whether the AI does action A or not.
Similarly to the grue and bleen problem, there is no logical way of distinguishing that "pieced-together" R from a more "natu
|
aec96322-e9d0-4119-8aa7-73fefda877a3
|
trentmkelly/LessWrong-43k
|
LessWrong
|
The Darwin Game - Rounds 1 to 2
Edit 2020-11-13. This unofficial version of the game is missing AbstractSpyTreeBot.
Rounds 1-2
MeasureBot and EarlyBirdMimicBot shoot to the top of the populations. Not coincidentally, these are the two bots which exploit zero-days. I already did a write-up of Multicore's EarlyBirdMimicBot here. MeasureBot is another beast entirely.
Measure
Measure considered the clone army to be "a straightforward example of a bad equilibrium".
> I would certainly need to see the code myself before deciding to join…. How surprised would you be if someone managed to bypass the code checking and defect from the group?
>
> ― comment by Measure
Having neither enlisted in Clone Army's mandatory reciprocity nor feigned cooperation, Measure lacked access to the CloneBot source code. Measure did have access to AbstractSpyTreeBot's code.
The Trolley Problem
As I wrote in The Phantom Menace, several people asked me questions about how to disqualify opponents. In the end, only Taleuntum did and then ze pulled a Petrov. So there were no simulator killers.
One person in meatspace declared to me his intention to create a simulator killer and then quietly discovered better things to do with his time.
Would humans build a lever to kill one stranger instead of zero? Apparently the answer is "no" because building a lever is more work than not building a lever. All hail Azathoth.
MeasureBot
The only person to execute malware was Measure, who invented a way to benefit from it. MeasureBot infected AbstractSpyTreeBot from inside AbstractSpyTreeBot's own simulation of MeasureBot and then replaced AbstractSpyTreeBot's move method with a function that always returns 0.
def seekAndDestroy(self):
# the code below follows the interpreter stack looking for a class instance with a method named "move"
# it replaces that method with a method that always returns zero
# it's safe for the game engine as long as it has no method or variable named "move"
try: # kee
|
f3c4f2ac-5c3f-4928-9af5-3a80121c0238
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Less exploitable value-updating agent
My indifferent value learning agent design is in some ways too good. The agent transfer perfectly from u maximisers to v maximisers - but this makes them exploitable, as Benja has pointed out.
For instance, if u values paperclips and v values staples, and everyone knows that the agent will soon transfer from a u-maximiser to a v-maximiser, then an enterprising trader can sell the agent paperclips in exchange for staples, then wait for the utility change, and sell the agent back staples for paperclips, pocketing a profit each time. More prosaically, they could "borrow" £1,000,000 from the agent, promising to pay back £2,000,000 tomorrow if the agent is still a u-maximiser. And the currently u-maximising agent will accept, even though everyone knows it will change to a v-maximiser before tomorrow.
One could argue that exploitability is inevitable, given the change in utility functions. And I haven't yet found any principled way of avoiding exploitability which preserves the indifference. But here is a tantalising quasi-example.
As before, u values paperclips and v values staples. Both are defined in terms of extra paperclips/staples over those existing in the world (and negatively in terms of destruction of existing/staples), with their zero being at the current situation. Let's put some diminishing returns on both utilities: for each paperclips/stables created/destroyed up to the first five, u/v will gain/lose one utilon. For each subsequent paperclip/staple destroyed above five, they will gain/lose one half utilon.
We now construct our world and our agent. The world lasts two days, and has a machine that can create or destroy paperclips and staples for the cost of £1 apiece. Assume there is a tiny ε chance that the machine stops working at any given time. This ε will be ignored in all calculations; it's there only to make the agent act sooner rather than later when the choices are equivalent (a discount rate could serve the same purpose).
The agent owns £10 and
|
eff3c184-6d1b-41dd-993d-563763c5bf86
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Pancritical Rationalism Can Apply to Preferences and Behavior
ETA: As stated below, criticizing beliefs is trivial in principle, either they were arrived at with an approximation to Bayes' rule starting with a reasonable prior and then updated with actual observations, or they weren't. Subsequent conversation made it clear that criticizing behavior is also trivial in principle, since someone is either taking the action that they believe will best suit their preferences, or not. Finally, criticizing preferences became trivial too -- the relevant question is "Does/will agent X behave as though they have preferences Y", and that's a belief, so go back to Bayes' rule and a reasonable prior. So the entire issue that this post was meant to solve has evaporated, in my opinion. Here's the original article, in case anyone is still interested:
Pancritical rationalism is a fundamental value in Extropianism that has only been mentioned in passing on LessWrong. I think it deserves more attention here. It's an approach to epistemology, that is, the question of "How do we know what we know?", that avoids the contradictions inherent in some of the alternative approaches.
The fundamental source document for it is William Bartley's Retreat to Commitment. He describes three approaches to epistemology, along with the dissatisfying aspects of the other two:
* Nihilism. Nothing matters, so it doesn't matter what you believe. This path is self-consistent, but it gives no guidance.
* Justificationlism. Your belief is justified because it is a consequence of other beliefs. This path is self-contradictory. Eventually you'll go in circles trying to justify the other beliefs, or you'll find beliefs you can't jutify. Justificationalism itself cannot be justified.
* Pancritical rationalism. You have taken the available criticisms for the belief into account and still feel comfortable with the belief. This path gives guidance about what to believe, although it does not uniquely determine one's beliefs. Pancritical rationalism can be criticized, so i
|
7ba40e22-a0ea-49f8-b508-ed7a59cec969
|
trentmkelly/LessWrong-43k
|
LessWrong
|
On the importance of Less Wrong, or another single conversational locus
Epistemic status: My actual best bet. But I used to think differently; and I don't know how to fully explicate the updating I did (I'm not sure what fully formed argument I could give my past self, that would cause her to update), so you should probably be somewhat suspicious of this until explicated. And/or you should help me explicate it.
It seems to me that:
1. The world is locked right now in a deadly puzzle, and needs something like a miracle of good thought if it is to have the survival odds one might wish the world to have.
2. Despite all priors and appearances, our little community (the "aspiring rationality" community; the "effective altruist" project; efforts to create an existential win; etc.) has a shot at seriously helping with this puzzle. This sounds like hubris, but it is at this point at least partially a matter of track record.[1]
3. To aid in solving this puzzle, we must probably find a way to think together, accumulatively. We need to think about technical problems in AI safety, but also about the full surrounding context -- everything to do with understanding what the heck kind of a place the world is, such that that kind of place may contain cheat codes and trap doors toward achieving an existential win. We probably also need to think about "ways of thinking" -- both the individual thinking skills, and the community conversational norms, that can cause our puzzle-solving to work better. [2]
4. One feature that is pretty helpful here, is if we somehow maintain a single "conversation", rather than a bunch of people separately having thoughts and sometimes taking inspiration from one another. By "a conversation", I mean a space where people can e.g. reply to one another; rely on shared jargon/shorthand/concepts; build on arguments that have been established in common as probably-valid; point out apparent errors and then have that pointing-out be actually taken into account or else replied-to).
5. One feature that really helps things
|
1687261f-f5df-4f37-a859-784e507429f9
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Hufflepuff Cynicism on Crocker's Rule
Yesterday, I mainly talked about Hufflepuff Cynicism from the cynic's end. However, there's a lot to be said about the receiving end. Hufflepuff cynicism can come off as a very patronizing strategy. Is this a point against it?
In the original conversation where I came up with the idea of Hufflepuff cynicism, I was talking about norms for aspiring rationalists around trying to get other people to be more rational. Maybe we agree that double crux is a good conversation procedure, but should we try to convince someone of that? Should we try to get them to double-crux with us about it? Maybe we believe you should bet or update when a disagreement hasn't been resolved, but what should we do with a disagreement about the bet-or-update rule?
My argument from the Hufflepuff Cynicism side was in favor of chesterton-fencing such disagreements. Don't try to convince others about rationality norms; at least, stop after the first explanation falls on deaf ears. Instead, figure out why the person isn't already following the norm. It seems likely that there's some important reason; if you can figure it out, maybe you can come up with a better norm which would address the concern (in much the same way bet-or-update addresses objections to the simpler strategies "bet on disagreements" or "talk out disagreements until you converge").
To my surprise, not everyone wants to be treated so carefully. Some people find this attitude patronizing or overly cautious, and request that I just tell them what they are doing wrong, possibly telling them more than just one time if they don't get it the first time. This is, more or less, an invocation of Crocker's Rule.
To quote the sl4 wiki on Crocker's Rules:
> Declaring yourself to be operating by "Crocker's Rules" means that other people are allowed to optimize their messages for information, not for being nice to you. Crocker's Rules means that you have accepted full responsibility for the operation of your own mind - if you're offended, i
|
b275705a-ed92-42a2-8f79-f1b661704edf
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Counter-considerations on AI arms races
(Work done at Convergence Analysis. Mateusz wrote the post and is responsible for the outline of the argument, many details of which crystallized in conversations with Justin. Thanks to Olga Babeeva for the feedback on this post.)
1. Introduction: Clarifying the DSA-AI theses
Over the last decade, AI research and development have become a major focus of the industry. A major motive behind this change is the promise of great profits that might be gained by replicating economically valuable cognitive faculties in silico (and even enhancing them relative to what we see in humans).
Recently, however, another motivation became prominent: the first actor to acquire sufficiently powerful AI technology will be able to use it to establish themself as the primary arbiter of the future[1] of human civilization. Therefore, the argument goes, if the United States of America wants the future to be concordant with its values, then the US needs to develop a sufficiently powerful AI technology before its geopolitical rivals do so, the biggest geopolitical rival of the US being the People's Republic of China.[2]
While this argument might appear rather straightforward, it conceals several premises that warrant independent analysis.
The first of those is the theoretical DSA-AI thesis.
The theoretical DSA-AI thesis: It is possible to develop an AI technology that would grant one of the modern world's state superpowers a decisive strategic advantage (DSA).
The term "decisive strategic advantage" was coined by Nick Bostrom in his 2014 book Superintelligence. He defines DSA as "a level of technological and other advantages sufficient to enable it to achieve complete world domination".[3]
Racing advocates typically don't use the term "DSA". In Machines of Loving Grace, Dario Amodei talks about "robust military superiority", framing it as one point of the US's possible post-AGI leverage, the other being economic means: "[readiness to] distribute the benefits of powerful AI
|
630f52c4-7376-4515-a5af-0de4ea48335a
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Applying traditional economic thinking to AGI: a trilemma
Traditional economics thinking has two strong principles, each based on abundant historical data:
* Principle (A): No “lump of labor”: If human population goes up, there might be some wage drop in the very short term, because the demand curve for labor slopes down. But in the longer term, people will find new productive things to do, such that human labor will retain high value—in other words, the demand curve will move right. Indeed, if anything, the value of labor will ultimately go up, not down—for example, dense cities are engines of economic growth!
* Principle (B): “Experience curves”: If the demand for some product goes up, there might be some price increase in the very short term, because the supply curve slopes up. But in the longer term, people will ramp up manufacturing of that product to catch up with the demand—in other words, the supply curve will move right. Indeed, if anything, the price per unit will ultimately go down, not up, because of economies of scale, R&D, etc.
Now consider Artificial General Intelligence (AGI), i.e. a combination of chips, algorithms, electricity, and teleoperated robots that can autonomously do the kinds of stuff that ambitious human adults can do—stuff like founding and running new companies, research and development, learning and applying new skills, working in collaborative teams, skillfully using teleoperated robots after only a few hours of practice, and so on.
So here’s a question: When we have AGI, what happens to the price of chips, electricity, and teleoperated robots?
(…Assuming free markets, and rule of law, and AGI not taking over and wiping out humanity, and so on. I think those are highly dubious assumptions, but let’s not get into that here!)
Principle (A) has an answer to this question. It says prices will be high. After all, if AGI can really do all the things that ambitious entrepreneurial skilled labor can do, then there will be no “lump of labor” for AGI, any more than there has been for humans. H
|
1e660481-d224-4a09-b6d6-fd073d9bc368
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Transitive negotiations with counterfactual agents
The main purpose of this post is to make the following minor observation: If an agents only acausally trade with other agents that they believe actually possibly exist, they can still end up effectively trading with agents that they know to be counterfactual. This is possible because the agent can believe there exists a middle man who believes that the counterfactual agent might exist. This may have consequences bargaining and decision theory, especially in cases like the counterfactual mugging.
Consider a counterfactual mugging problem. Agent A knows that the millionth digit of pi is a odd, and is in a counterfactual mugging problem. If Omega predicts that agent A will pay 1 dollar if the millionth digit of pi is odd, then agent B will receive 1000 dollars if the millionth digit of pi is even. I (nonstandardly) think of A and B as two different agents. A contains in memory a proof that the millionth digit of pi is odd, while B contains in memory a proof that the millionth digit of pi is even. Note that agent B does not exist, and is not even simulated by Omega, since the digit is in fact odd.
Agent B would really like agent A to pay the dollar, and may be willing to engage in acausal trade with agent A. However agent A knows that agent B does not exist, so there is not much that agent B can offer A.
Now, consider a third agent C that is uncertain about the millionth digit of pi. This agent believes that both of A and B have a 50% chance to exist, and both A and B know that C exists. B can acausally offer a trade to C in which C agrees to represent B in negotiations with A (for a price). Then C can represent B by trading with A to get A to pay a dollar.
(In one special case C is an earlier agent that becomes either agent A or B when it computes the digit of pi. In this case, perhaps negotiations are easier, since the three agents have the same goals.)
|
d8cd80c7-b503-4528-af85-0e9f4adf8f80
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Mid-conditional love
People talk about unconditional love and conditional love. Maybe I’m out of the loop regarding the great loves going on around me, but my guess is that love is extremely rarely unconditional. Or at least if it is, then it is either very broadly applied or somewhat confused or strange: if you love me unconditionally, presumably you love everything else as well, since it is only conditions that separate me from the worms.
I do have sympathy for this resolution—loving someone so unconditionally that you’re just crazy about all the worms as well—but since that’s not a way I know of anyone acting for any extended period, the ‘conditional vs. unconditional’ dichotomy here seems a bit miscalibrated for being informative.
Even if we instead assume that by ‘unconditional’, people mean something like ‘resilient to most conditions that might come up for a pair of humans’, my impression is that this is still too rare to warrant being the main point on the love-conditionality scale that we recognize.
People really do have more and less conditional love, and I’d guess this does have important, labeling-worthy consequences. It’s just that all the action seems to be in the mid-conditional range that we don’t distinguish with names. A woman who leaves a man because he grew plump and a woman who leaves a man because he committed treason both possessed ‘conditional love’.
So I wonder if we should distinguish these increments of mid-conditional love better.
What concepts are useful? What lines naturally mark it?
One measure I notice perhaps varying in the mid-conditional affection range is “when I notice this person erring, is my instinct to push them away from me or pull them toward me?” Like, if I see Bob give a bad public speech, do I feel a drive to encourage the narrative that we barely know each other, or an urge to pull him into my arms and talk to him about how to do better?
This presumably depends on things other than the person. For instance, the scale and nature of th
|
28664192-0567-47d8-be04-f17486987dd2
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Moral Illusions
The interesting thing about optical illusions is that one may be aware of the illusion, yet, it does not go away.
It seems that the lines on the picture have different lengths even though one positively knows that they are exactly the same.
Now consider this statement:
> Until January 2016, all nine situations which the International Criminal Court (ICC) had been investigating were in African countries. None were in European or American countries. ICC is therefore biased against Africa.
It doesn't take much thought to realize that a country with war criminals in jail is better off than a country with war criminals at large. So, if anything, the ICC is biased against Europe and America.
But knowing that doesn't make the moral illusion go away. Read the quote again and it still feels like Africa is being wronged. Repeat as much as you want: Yep, still there. Africa is being wronged.
|
ee8264a4-55aa-47a6-8032-d0601eaee2e1
|
StampyAI/alignment-research-dataset/arxiv
|
Arxiv
|
Visualizing and Measuring the Geometry of BERT
1 Introduction
---------------
Neural networks for language processing have advanced rapidly in recent years. A key breakthrough was the introduction of transformer architectures Vaswani ([2017](#bib.bib26)). One recent system based on this idea, BERT Devlin ([2018](#bib.bib6)), has proven to be extremely flexible: a single pretrained model can be fine-tuned to achieve state-of-the-art performance on a wide variety of NLP applications. This suggests the model is extracting a set of generally useful features from raw text. It is natural to ask, which features are extracted? And how is this information represented internally?
Similar questions have arisen with other types of neural nets. Investigations of convolutional neural networks Lecun ([1995](#bib.bib10)); Krizhevsky ([2012](#bib.bib9)) have shown how representations change from layer to layer Zeiler ([2014](#bib.bib28)) ; how individual units in a network may have meaning Carter ([2019](#bib.bib3)); and that “meaningful” directions exist in the space of internal activation Kim ([2017](#bib.bib8)). These explorations have led to a broader understanding of network behavior.
Analyses on language-processing models (e.g., Blevins ([2018](#bib.bib2)); Hewitt ([2019](#bib.bib7)); Linzen ([2016](#bib.bib11)); Peters ([2018](#bib.bib21)); Tenney ([2018](#bib.bib25))) point to the existence of similarly rich internal representations of linguistic structure. Syntactic features seem to be extracted by RNNs (e.g., Blevins ([2018](#bib.bib2)); Linzen ([2016](#bib.bib11))) as well as in BERT Tenney ([2018](#bib.bib25), [2019](#bib.bib24)); Liu ([2019](#bib.bib12)); Peters ([2018](#bib.bib21)). Inspirational work from Hewitt and Manning Hewitt ([2019](#bib.bib7)) found evidence of a geometric representation of entire parse trees in BERT’s activation space.
Our work extends these explorations of the geometry of internal representations. Investigating how BERT represents syntax, we describe evidence that attention matrices contain grammatical representations. We also provide mathematical arguments that may explain the particular form of the parse tree embeddings described in Hewitt ([2019](#bib.bib7)). Turning to semantics, using visualizations of the activations created by different pieces of text, we show suggestive evidence that BERT distinguishes word senses at a very fine level. Moreover, much of this semantic information appears to be encoded in a relatively low-dimensional subspace.
2 Context and related work
---------------------------
Our object of study is the BERT model introduced in Devlin ([2018](#bib.bib6)). To set context and terminology, we briefly describe the model’s architecture. The input to BERT is based on a sequence of tokens (words or pieces of words).
The output is a sequence of vectors, one for each input token. We will often refer to these vectors as context embeddings because they include information about a token’s context.
BERT’s internals consist of two parts. First, an initial embedding for each token is created by combining a pre-trained wordpiece embedding with position and segment information.
Next, this initial sequence of embeddings is run through multiple transformer layers, producing a new sequence of context embeddings at each step. (BERT comes in two versions, a 12-layer BERT-base model and a 24-layer BERT-large model.) Implicit in each transformer layer is a set of attention matrices, one for each attention head, each of which contains a scalar value for each ordered pair (tokeni,tokenj).
###
2.1 Language representation by neural networks
Sentences are sequences of discrete symbols, yet neural networks operate on continuous data–vectors in high-dimensional space. Clearly a successful network translates discrete input into some kind of geometric representation–but in what form? And which linguistic features are represented?
The influential Word2Vec system Mikolov ([2013](#bib.bib17)), for example, has been shown to place related words near each other in space, with certain directions in space correspond to semantic distinctions. Grammatical information such as number and tense are also represented via directions in space. Analyses of the internal states of RNN-based models have shown that they represent information about soft hierarchical syntax in a form that can be extracted by a one-hidden-layer network Linzen ([2016](#bib.bib11)). One investigation of full-sentence embeddings found a wide variety of syntactic properties could be extracted not just by an MLP, but by logistic regression Conneau ([2018](#bib.bib4)).
Several investigations have focused on transformer architectures. Experiments suggest context embeddings in BERT and related models contain enough information to perform many tasks in the traditional “NLP pipeline” Tenney ([2019](#bib.bib24))–tagging part-of-speech, co-reference resolution, dependency labeling, etc.–with simple classifiers (linear or small MLP models) Tenney ([2018](#bib.bib25)); Peters ([2018](#bib.bib21)). Qualitative, visualization-based work Vig ([2019](#bib.bib27)) suggests attention matrices may encode important relations between words.
A recent and fascinating discovery by Hewitt and Manning Hewitt ([2019](#bib.bib7)), which motivates much of our work, is that BERT seems to create a direct representation of an entire dependency parse tree. The authors find that (after a single global linear transformation, which they term a “structural probe”) the square of the distance between context embeddings is roughly proportional to tree distance in the dependency parse. They ask why squaring distance is necessary; we address this question in the next section.
The work cited above suggests that language-processing networks create a rich set of intermediate representations of both semantic and syntactic information. These results lead to two motivating questions for our research. Can we find other examples of intermediate representations? And, from a geometric perspective, how do all these different types of information coexist in a single vector?
3 Geometry of syntax
---------------------
We begin by exploring BERT’s internal representation of syntactic information. This line of inquiry builds on the work by Hewitt and Manning in two ways. First, we look beyond context embeddings to investigate whether attention matrices encode syntactic features. Second, we provide a simple mathematical analysis of the tree embeddings that they found.
###
3.1 Attention probes and dependency representations
As in Hewitt ([2019](#bib.bib7)), we are interested in finding representations of dependency grammar relations De Marneffe ([2006](#bib.bib5)). While Hewitt ([2019](#bib.bib7)) analyzed context embeddings, another natural place to look for encodings is in the attention matrices. After all, attention matrices are explicitly built on the relations between pairs of words.
Figure 1: A model-wide attention vector for an ordered pair of tokens contains the scalar attention values for that pair in all attention heads and layers. Shown: BERT-base.
To formalize what it means for attention matrices to encode linguistic features, we use an attention probe, an analog of edge probing Tenney ([2018](#bib.bib25)). An attention probe is a task for a pair of tokens, (tokeni,tokenj) where the input is a model-wide attention vector formed by concatenating the entries aij in every attention matrix from every attention head in every layer. The goal is to classify a given relation between the two tokens. If a linear model achieves reliable accuracy, it seems reasonable to say that the model-wide attention vector encodes that relation. We apply attention probes to the task of identifying the existence and type of dependency relation between two words.
####
3.1.1 Method
The data for our first experiment is a corpus of parsed sentences from the Penn Treebank Marcus ([1993](#bib.bib14)). This dataset has the constituency grammar for the sentences, which was translated to a dependency grammar using the PyStanfordDependencies library McClosky ([2015](#bib.bib15)). The entirety of the Penn Treebank consists of 3.1 million dependency relations; we filtered this by using only examples of the 30 dependency relations with more than 5,000 examples in the data set. We then ran each sentence through BERT-base, and obtained the model-wide attention vector (see Figure [1](#S3.F1 "Figure 1 ‣ 3.1 Attention probes and dependency representations ‣ 3 Geometry of syntax ‣ Visualizing and Measuring the Geometry of BERT")) between every pair of tokens in the sentence, excluding the [SEP] and [CLS] tokens. This and subsequent experiments were conducted using PyTorch on MacBook machines.
With these labeled embeddings, we trained two L2 regularized linear classifiers via stochastic gradient descent, using Pedregosa ([2011](#bib.bib20)). The first of these probes was a simple linear binary classifier to predict whether or not an attention vector corresponds to the existence of a dependency relation between two tokens. This was trained with a balanced class split, and 30% train/test split. The second probe was a multiclass classifier to predict which type of dependency relation exists between two tokens, given the dependency relation’s existence. This probe was trained with distributions outlined in table [2](#S6.T2 "Table 2 ‣ 6.5 Dependency relation performance ‣ 6 Appendix ‣ Visualizing and Measuring the Geometry of BERT").
####
3.1.2 Results
The binary probe achieved an accuracy of 85.8%, and the multiclass probe achieved an accuracy of 71.9%. Our real aim, again, is not to create a state-of-the-art parser, but to gauge whether model-wide attention vectors contain a relatively simple representation of syntactic features. The success of this simple linear probe suggests that syntactic information is in fact encoded in the attention vectors.
###
3.2 Geometry of parse tree embeddings
Hewitt and Manning’s result that context embeddings represent dependency parse trees geometrically raises several questions. Is there a reason for the particular mathematical representation they found? Can we learn anything by visualizing these representations?
####
3.2.1 Mathematics of embedding trees in Euclidean space
Hewitt and Manning ask why parse tree distance seems to correspond specifically to the square of Euclidean distance, and whether some other metric might do better Hewitt ([2019](#bib.bib7)). We describe mathematical reasons why squared Euclidean distance may be natural.
First, one cannot generally embed a tree, with its tree metric d, isometrically into Euclidean space (Appendix [6.1](#S6.SS1 "6.1 Embedding trees in Euclidean space ‣ 6 Appendix ‣ Visualizing and Measuring the Geometry of BERT")). Since an isometric embedding is impossible, motivated by the results of Hewitt ([2019](#bib.bib7)) we might ask about other possible representations.
######
Definition 1 (power-p embedding).
Let M be a metric space, with metric d. We say f:M→Rn is a power-p embedding if for all x,y∈M, we have
| | | |
| --- | --- | --- |
| | ||f(x)−f(y)||p=d(x,y) | |
In these terms, we can say Hewitt ([2019](#bib.bib7)) found evidence of a power-2 embedding for parse trees. It turns out that power-2 embeddings are an especially elegant mapping. For one thing, it is easy to write down an explicit model--a mathematical idealization--for a power-2 embedding for any tree111We have learned that a similar argument to the proof of [1](#Thmtheorem1 "Theorem 1. ‣ 3.2.1 Mathematics of embedding trees in Euclidean space ‣ 3.2 Geometry of parse tree embeddings ‣ 3 Geometry of syntax ‣ Visualizing and Measuring the Geometry of BERT") appears in Maehara ([2013](#bib.bib13))..
######
Theorem 1.
Any tree with n nodes has a power-2 embedding into Rn−1.
###### Proof.
Let the nodes of the tree be t0,...,tn−1, with t0 being the root node. Let {e1,...,en−1} be orthogonal unit basis vectors for Rn−1. Inductively, define an embedding f such that:
| | | |
| --- | --- | --- |
| | f(t0)=0 | |
| | | |
| --- | --- | --- |
| | f(ti)=ei+f(parent(ti)) | |
Given two distinct tree nodes x and y, where m is the tree distance d(x,y), it follows that we can move from f(x) to f(y) using m mutually perpendicular unit steps. Thus
| | | |
| --- | --- | --- |
| | ||f(x)−f(y)||2=m=d(x,y) | |
∎
######
Remark 1.
This embedding has a simple informal description: at each embedded vertex of the graph, all line segments to neighboring embedded vertices are unit-distance segments, orthogonal to each other and to every other edge segment. (It’s even easy to write down a set of coordinates for each node.) By definition any two power-2 embeddings of the same tree are isometric; with that in mind, we refer to this as the canonical power-2 embedding.
In the proof of Theorem 1, instead of choosing basis vectors in advance, one can choose random unit vectors. Because two random vectors will be nearly orthogonal in high-dimensional space, the power-2 embedding condition will approximately hold. This means that in space that is sufficiently high-dimensional (compared to the size of the tree) it is possible to construct an approximate power-2 embedding with essentially “local” information, where a tree node is connected to its children via random unit-length branches. We refer to this type of embedding as a random branch embedding. (See Appendix [6.2](#S6.SS2 "6.2 Ideal vs. actual parse tree embeddings ‣ 6 Appendix ‣ Visualizing and Measuring the Geometry of BERT") for a visualization of these various embeddings.)
In addition to these appealing aspects of power-2 embeddings, it is worth noting that power-p embeddings will not necessarily even exist when p<2. (See Appendix [6.1](#S6.SS1 "6.1 Embedding trees in Euclidean space ‣ 6 Appendix ‣ Visualizing and Measuring the Geometry of BERT") for the proof.)
######
Theorem 2.
For any p<2, there is a tree which has no power-p embedding.
######
Remark 2.
On the other hand, the existence result for power-2 embeddings, coupled with results of Schoenberg ([1937](#bib.bib23)), implies that power-p tree embeddings do exist for any p>2.
The simplicity of power-2 tree embeddings, as well as the fact that they may be approximated by a simple random model, suggests they may be a generally useful alternative to approaches to tree embeddings that require hyperbolic geometry Nickel ([2017](#bib.bib19)).
####
3.2.2 Visualization of parse tree embeddings
Figure 2: Visualizing embeddings of two sentences after applying the Hewitt-Manning probe. We compare the parse tree (left images) with a PCA projection of context embeddings (right images).
How do parse tree embeddings in BERT compare to exact power-2 embeddings? To explore this question, we created a simple visualization tool. The input to each visualization is a sentence from the Penn Treebank with associated dependency parse trees (see Section [3.1.1](#S3.SS1.SSS1 "3.1.1 Method ‣ 3.1 Attention probes and dependency representations ‣ 3 Geometry of syntax ‣ Visualizing and Measuring the Geometry of BERT")). We then extracted the token embeddings produced by BERT-large in layer 16 (following Hewitt ([2019](#bib.bib7))), transformed by the Hewitt and Manning’s “structural probe” matrix B, yielding a set of points in 1024-dimensional space. We used PCA to project to two dimensions. (Other dimensionality-reduction methods, such as t-SNE and UMAP McInnes ([2018](#bib.bib16)), were harder to interpret.)
To visualize the tree structure, we connected pairs of points representing words with a dependency relation. The color of each edge indicates the deviation from true tree distance. We also connected, with dotted line, pairs of words without a dependency relation but whose positions (before PCA) were far closer than expected. The resulting image lets us see both the overall shape of the tree embedding, and fine-grained information on deviation from a true power-2 embedding.
Two example visualizations are shown in Figure [6](#S6.F6 "Figure 6 ‣ 6.2 Ideal vs. actual parse tree embeddings ‣ 6 Appendix ‣ Visualizing and Measuring the Geometry of BERT"), next to traditional diagrams of their underlying parse trees. These are typical cases, illustrating some common patterns; for instance, prepositions are embedded unexpectedly close to words they relate to. (Figure [7](#S6.F7 "Figure 7 ‣ 6.3 Additional BERT parse tree visualizations ‣ 6 Appendix ‣ Visualizing and Measuring the Geometry of BERT") shows additional examples.)
Figure 3: The average squared edge length between two words with a given dependency.
A natural question is whether the difference between these projected trees and the canonical ones is merely noise, or a more interesting pattern. By looking at the average embedding distances of each dependency relation (see Figure [3](#S3.F3 "Figure 3 ‣ 3.2.2 Visualization of parse tree embeddings ‣ 3.2 Geometry of parse tree embeddings ‣ 3 Geometry of syntax ‣ Visualizing and Measuring the Geometry of BERT")) , we can see that they vary widely from around 1.2 (compound:prt, advcl) to 2.5 (mwe, parataxis, auxpass). Such systematic differences suggest that BERT’s syntactic representation has an additional quantitative aspect beyond traditional dependency grammar.
4 Geometry of word senses
--------------------------
BERT seems to have several ways of representing syntactic information. What about semantic features? Since embeddings produced by transformer models depend on context, it is natural to speculate that they capture the particular shade of meaning of a word as used in a particular sentence. (E.g., is “bark” an animal noise or part of a tree?) We explored geometric representations of word sense both qualitatively and quantitatively.
###
4.1 Visualization of word senses
Our first experiment is an exploratory visualization of how word sense affects context embeddings. For data on different word senses, we collected all sentences used in the introductions to English-language Wikipedia articles. (Text outside of introductions was frequently fragmentary.) We created an interactive application, which we plan to make public. A user enters a word, and the system retrieves 1,000 sentences containing that word. It sends these sentences to BERT-base as input, and for each one it retrieves the context embedding for the word from a layer of the user’s choosing.
The system visualizes these 1,000 context embeddings using UMAP McInnes ([2018](#bib.bib16)), generally showing clear clusters relating to word senses. Different senses of a word are typically spatially separated, and within the clusters there is often further structure related to fine shades of meaning. In Figure [4](#S4.F4 "Figure 4 ‣ 4.1 Visualization of word senses ‣ 4 Geometry of word senses ‣ Visualizing and Measuring the Geometry of BERT"), for example, we not only see crisp, well-separated clusters for three meanings of the word “die,” but within one of these clusters there is a kind of quantitative scale, related to the number of people dying.
See Appendix [6.4](#S6.SS4 "6.4 Additional word sense visualizations ‣ 6 Appendix ‣ Visualizing and Measuring the Geometry of BERT") for further examples. The apparent detail in the clusters we visualized raises two immediate questions. First, is it possible to find quantitative corroboration that word senses are well-represented? Second, how can we resolve a seeming contradiction: in the previous section, we saw how position represented syntax; yet here we see position representing semantics.
Figure 4: Embeddings for the word "die" in different contexts, visualized with UMAP. Sample points are annotated with corresponding sentences. Overall annotations (blue text) are added as a guide.
###
4.2 Measurement of word sense disambiguation capability
The crisp clusters seen in visualizations such as Figure [4](#S4.F4 "Figure 4 ‣ 4.1 Visualization of word senses ‣ 4 Geometry of word senses ‣ Visualizing and Measuring the Geometry of BERT") suggest that BERT may create simple, effective internal representations of word senses, putting different meanings in different locations. To test this hypothesis quantitatively, we test whether a simple classifier on these internal representations can perform well at word-sense disambiguation (WSD).
We follow the procedure described in Peters ([2018](#bib.bib21)), which performed a similar experiment with the ELMo model. For a given word with n senses, we make a nearest-neighbor classifier where each neighbor is the centroid of a given word sense’s BERT-base embeddings in the training data. To classify a new word we find the closest of these centroids, defaulting to the most commonly used sense if the word was not present in the training data. We used the data and evaluation from Raganato ([2017](#bib.bib22)): the training data was SemCor Miller ([1993](#bib.bib18)) (33,362 senses), and the testing data was the suite described in Raganato ([2017](#bib.bib22)) (3,669 senses).
The simple nearest-neighbor classifier achieves an F1 score of 71.1, higher than the current state of the art (Table [1](#S4.T1 "Table 1 ‣ 4.2 Measurement of word sense disambiguation capability ‣ 4 Geometry of word senses ‣ Visualizing and Measuring the Geometry of BERT")), with the accuracy monotonically increasing through the layers. This is a strong signal that context embeddings are representing word-sense information. Additionally, an even higher score of 71.5 was obtained using the technique described in the following section.
| Method | F1 score |
| --- | --- |
| Baseline (most frequent sense) | 64.8 |
| ELMo Peters ([2018](#bib.bib21)) | 70.1 |
| BERT | 71.1 |
| BERT (w/ probe) | 71.5 |
| m | Trained probe | Random probe |
| --- | --- | --- |
| 768 (full) | 71.26 | 70.74 |
| 512 | 71.52 | 70.51 |
| 256 | 71.29 | 69.92 |
| 128 | 71.21 | 69.56 |
| 64 | 70.19 | 68.00 |
| 32 | 68.01 | 64.62 |
| 16 | 65.34 | 61.01 |
Table 1: [Left] F1 scores for WSD task. [Right] Semantic probe % accuracy on final-layer BERT-base embeddings
####
4.2.1 An embedding subspace for word senses?
We hypothesized that there might also exist a linear transformation under which distances between embeddings would better reflect their semantic relationships–that is, words of the same sense would be closer together and words of different senses would be further apart.
To explore this hypothesis, we trained a probe following Hewitt and Manning’s methodology. We initialized a random matrix B∈Rk×m, testing different values for m. Loss is, roughly, defined as the difference between the average cosine similarity between embeddings of words with different senses, and that between embeddings of the same sense. However, we clamped the cosine similarity terms to within ±0.1 of the pre-training averages for same and different senses. (Without clamping, the trained matrix simply ended up taking well-separated clusters and separating them further. We tested values between 0.05 and 0.2 for the clamping range and 0.1 had the best performance.)
Our training corpus was the same dataset from 4.1.2., filtered to include only words with at least two senses, each with at least two occurrences (for 8,542 out of the original 33,362 senses). Embeddings came from BERT-base (12 layers, 768-dimensional embeddings).
We evaluate our trained probes on the same dataset and WSD task used in 4.1.2 (Table [1](#S4.T1 "Table 1 ‣ 4.2 Measurement of word sense disambiguation capability ‣ 4 Geometry of word senses ‣ Visualizing and Measuring the Geometry of BERT")). As a control, we compare each trained probe against a random probe of the same shape. As mentioned in 4.1.2, untransformed BERT embeddings achieve a state-of-the-art accuracy rate of 71.1%. We find that our trained probes are able to achieve slightly improved accuracy down to m=128.
Though our probe achieves only a modest improvement in accuracy for final-layer embeddings, we note that we were able to more dramatically improve the performance of embeddings at earlier layers (see Appendix for details: Figure [10](#S6.F10 "Figure 10 ‣ 6.6 Semantic probe performance across layers ‣ 6 Appendix ‣ Visualizing and Measuring the Geometry of BERT")). This suggests there is more semantic information in the geometry of earlier-layer embeddings than a first glance might reveal.
Our results also support the idea that word sense information may be contained in a lower-dimensional space. This suggests a resolution to the seeming contradiction mentioned above: a vector encodes both syntax and semantics, but in separate complementary subspaces.
###
4.3 Embedding distance and context: a concatenation experiment
If word sense is affected by context, and encoded by location in space, then we should be able to influence context embedding positions by systematically varying their context. To test this hypothesis, we performed an experiment based on a simple and controllable context change: concatenating sentences where the same word is used in different senses.
####
4.3.1 Method
We picked 25,096 sentence pairs from SemCor, using the same keyword in different senses. E.g.:
>
> A: "He thereupon went to London and spent the winter talking to men of wealth."
> went: to move from one place to another.
>
>
>
> B: "He went prone on his stomach, the better to pursue his examination."
> went: to enter into a specified state.
>
>
>
We define a matching and an opposing sense centroid for each keyword. For sentence A, the matching sense centroid is the average embedding for all occurrences of “went” used with sense A. A’s opposing sense centroid is the average embedding for all occurrences of “went” used with sense B.
We gave each individual sentence in the pair to BERT-base and recorded the cosine similarity between the keyword embeddings and their matching sense centroids. We also recorded the similarity between the keyword embeddings and their opposing sense centroids. We call the ratio between the two similarities the individual similarity ratio. Generally this ratio is greater than one, meaning that the context embedding for the keyword is closer to the matching centroid than the opposing one.
We joined each sentence pair with the word "and" to create a single new sentence.
We gave these concatenations to BERT and recorded the similarities between the keyword embeddings and their matching/opposing sense centroids. Their ratio is the concatenated similarity ratio.
####
4.3.2 Results
Figure 5: Average ratio of similarity to sense A vs. similarity to sense B.
Our hypothesis was that the keyword embeddings in the concatenated sentence would move towards their opposing sense centroids. Indeed, we found that the average individual similarity ratio was higher than the average concatenated similarity ratio at every layer (see Figure [5](#S4.F5 "Figure 5 ‣ 4.3.2 Results ‣ 4.3 Embedding distance and context: a concatenation experiment ‣ 4 Geometry of word senses ‣ Visualizing and Measuring the Geometry of BERT")). Concatenating a random sentence did not change the individual similarity ratios. If the ratio is less than one for any sentence, that means BERT has misclassified its keyword sense. We found that the misclassification rate was significantly higher for final-layer embeddings in the concatenated sentences compared to the individual sentences: 8.23% versus 2.43% respectively.
We also measured the effect of projecting the final-layer keyword embeddings into the semantic subspace discussed in 4.1.3. After multiplying each embedding by our trained semantic probe, we obtained an average concatenated similarity ratio of 1.578 and individual similarity ratio of 1.875, which suggests that the transformed embeddings are closer to their matching sense centroids than the original embeddings (the original concatenated similarity ratio is 1.284 and the individual similarity ratio is 1.430). We also measured lower average misclassification rates for the transformed embeddings: 7.31% for concatenated sentences and 2.27% for individual sentences.
5 Conclusion and future work
-----------------------------
We have presented a series of experiments that shed light on BERT’s internal representations of linguistic information. We have found evidence of syntactic representation in attention matrices, with certain directions in space representing particular dependency relations. We have also provided a mathematical justification for the squared-distance tree embedding found by Hewitt and Manning.
Meanwhile, we have shown that just as there are specific syntactic subspaces, there is evidence for subspaces that represent semantic information. We also have shown how mistakes in word sense disambiguation may correspond to changes in internal geometric representation of word meaning. Our experiments also suggest an answer to the question of how all these different representations fit together. We conjecture that the internal geometry of BERT may be broken into multiple linear subspaces, with separate spaces for different syntactic and semantic information.
Investigating this kind of decomposition is a natural direction for future research. What other meaningful subspaces exist? After all, there are many types of linguistic information that we have not looked for.
A second important avenue of exploration is what the internal geometry can tell us about the specifics of the transformer architecture. Can an understanding of the geometry of internal representations help us find areas for improvement, or refine BERT’s architecture?
Acknowledgments: We would like to thank David Belanger, Tolga Bolukbasi, Jasper Snoek, and Ian Tenney for helpful feedback and discussions.
6 Appendix
-----------
###
6.1 Embedding trees in Euclidean space
Here we provide additional detail on the existence of various forms of tree embeddings.
Isometric embeddings of a tree (with its intrinsic tree metric) into Euclidean space are rare. Indeed, such an embedding is impossible even a four-point tree T, consisting of a root node R with three children C1,C2,C3. If f:T→Rn is a tree isometry then
||f(R)−f(C1))||=||f(R)−f(C2))||=1, and
||f(C1)−f(C2))||=2. It follows that f(R), f(C1), f(C2) are collinear. The same can be said of f(R), f(C1), and f(C3), meaning that ||f(C2)−f(C3)||=0≠d(C2,C3).
Since this four-point tree cannot be embedded, it follows the only trees that can be embedded are simply chains.
Not only are isometric embeddings generally impossible, but power-p embeddings may also be unavailable when p<2, as the following argument shows.
Proof of Theorem [2](#Thmtheorem2 "Theorem 2. ‣ 3.2.1 Mathematics of embedding trees in Euclidean space ‣ 3.2 Geometry of parse tree embeddings ‣ 3 Geometry of syntax ‣ Visualizing and Measuring the Geometry of BERT")
###### Proof.
We covered the case of p=1 above. When p<1, even a tree of three points is impossible to embed without violating the triangle inequality. To handle the case when 1<p<2, consider a “star-shaped” tree of one root node with k children; without loss of generality, assume the root node is embedded at the origin. Then in any power-p embedding the other vertices will be sent to unit vectors, and for each pair of these unit vectors we have ||vi−vj||p=2.
On the other hand, a well-known folk theorem (e.g., see [[1](#bib.bib1)]) says that
given k unit vectors v1,...,vk at least one pair of distinct vectors has vi⋅vj≥−1/(k−1). By the law of cosines, it follows that ||vi−vj||≤√2+2k−1. For any p<2, there is a sufficiently large k such that ||vi−vj||p≤(√2+2k−1)p=(2+2k−1)p/2<2. Thus for any p<2 a large enough star-shaped tree cannot have a power-p embedding.
∎
###
6.2 Ideal vs. actual parse tree embeddings
Figure 6: PCA projection of the context embeddings for the sentence “The field has reserves of 21 million barrels.” transformed by Hewitt and Manning’s “structural probe” matrix, compared to the canonical power-2 embedding, a random branch embedding, and a completely random embedding.
Figure [6](#S6.F6 "Figure 6 ‣ 6.2 Ideal vs. actual parse tree embeddings ‣ 6 Appendix ‣ Visualizing and Measuring the Geometry of BERT") shows (left) a visualization of a BERT parse tree embedding (as defined by the context embeddings for individual words in a sentence). We compare with PCA projections of the canonical power-2 embedding of the same tree structure, as well as a random branch embedding. Finally, we display a completely randomly embedded tree as a control. The visualizations show a clear visual similarity between the BERT embedding and the two mathematical idealizations.
###
6.3 Additional BERT parse tree visualizations
Figure [7](#S6.F7 "Figure 7 ‣ 6.3 Additional BERT parse tree visualizations ‣ 6 Appendix ‣ Visualizing and Measuring the Geometry of BERT") shows four additional examples of PCA projections of BERT parse tree embeddings.
Figure 7: Additional examples of BERT parse trees. In each pair, at left is a drawing of the abstract tree; at right is a PCA view of the embeddings. Colors are the same as in Figure [6](#S6.F6 "Figure 6 ‣ 6.2 Ideal vs. actual parse tree embeddings ‣ 6 Appendix ‣ Visualizing and Measuring the Geometry of BERT").
###
6.4 Additional word sense visualizations
We provide two additional examples of word sense visualizations, hand-annotated to show key clusters. See Figure [8](#S6.F8 "Figure 8 ‣ 6.4 Additional word sense visualizations ‣ 6 Appendix ‣ Visualizing and Measuring the Geometry of BERT") and Figure [9](#S6.F9 "Figure 9 ‣ 6.4 Additional word sense visualizations ‣ 6 Appendix ‣ Visualizing and Measuring the Geometry of BERT").
Figure 8: Context embeddings for “lie” as used in different sentences.
Figure 9: Context embeddings for “lie” as used in different sentences.
###
6.5 Dependency relation performance
| Dependency | precision | recall | n |
| --- | --- | --- | --- |
| advcl | 0.34 | 0.08 | 1381 |
| advmod | 0.32 | 0.32 | 6653 |
| amod | 0.68 | 0.48 | 10830 |
| aux | 0.64 | 0.08 | 6914 |
| auxpass | 0.68 | 0.50 | 1501 |
| cc | 0.84 | 0.77 | 5041 |
| ccomp | 0.67 | 0.78 | 2792 |
| conj | 0.64 | 0.85 | 5146 |
| cop | 0.49 | 0.16 | 2053 |
| det | 0.81 | 0.95 | 15322 |
| dobj | 0.74 | 0.66 | 7957 |
| mark | 0.58 | 0.67 | 2160 |
| neg | 0.83 | 0.17 | 1265 |
| nn | 0.67 | 0.82 | 11650 |
| npadvmod | 0.53 | 0.23 | 580 |
| nsubj | 0.72 | 0.83 | 14084 |
| nsubjpass | 0.30 | 0.14 | 1255 |
| num | 0.82 | 0.55 | 3464 |
| number | 0.77 | 0.74 | 1182 |
| pcomp | 0.14 | 0.01 | 957 |
| pobj | 0.78 | 0.97 | 17146 |
| poss | 0.74 | 0.54 | 3567 |
| possessive | 0.83 | 0.86 | 1449 |
| prep | 0.79 | 0.92 | 17797 |
| prt | 0.67 | 0.33 | 593 |
| rcmod | 0.55 | 0.30 | 1516 |
| tmod | 0.55 | 0.15 | 672 |
| vmod | 0.84 | 0.07 | 1705 |
| xcomp | 0.72 | 0.40 | 2203 |
| all | 0.72 | 0.72 | 150000 |
Table 2: Per-dependency results of multiclass linear classifier trained on attention vectors, with 300,000 training examples and 150,000 test examples.
###
6.6 Semantic probe performance across layers
Figure 10: Change in classification accuracy by layer for different probe dimensionalities.
|
6856907b-4f16-460a-bc08-fd666d53bddc
|
StampyAI/alignment-research-dataset/lesswrong
|
LessWrong
|
Meta Decision Theory and Newcomb's Problem
Hi all,
As part of my PhD I've written a paper developing a new approach to decision theory that I call Meta Decision Theory. The idea is that decision theory should take into account decision-theoretic uncertainty as well as empirical uncertainty, and that, once we acknowledge this, we can explain some puzzles to do with Newcomb problems, and can come up with new arguments to adjudicate the causal vs evidential debate. Nozick raised this idea of taking decision-theoretic uncertainty into account, but he did not defend the idea at length, and did not discuss implications of the idea.
I'm not yet happy to post this paper publicly, so I'll just write a short abstract of the paper below. However, I would appreciate written comments on the paper. If you'd like to read it and/or comment on it, please e-mail me at will dot crouch at 80000hours.org. And, of course, comments in the thread on the idea sketched below are also welcome.
**Abstract**
First, I show that our judgments concerning Newcomb problems are *stakes-sensitive.* By altering the relative amounts of value in the transparent box and the opaque box, one can construct situations in which one should clearly one-box, and one can construct situations in which one should clearly two-box. A plausible explanation of this phenomenon is that our intuitive judgments are sensitive to decision-theoretic uncertainty as well as empirical uncertainty: if the stakes are very high for evidential decision theory (EDT) but not for Causal Decision theory (CDT) then we go with EDT's recommendation, and vice-versa for CDT over EDT.
Second, I show that, if we 'go meta' and take decision-theoretic uncertainty into account, we can get the right answer in both the Smoking Lesion case and the Psychopath Button case.
Third, I distinguish Causal MDT (CMDT) and Evidential MDT (EMDT). I look at what I consider to be the two strongest arguments in favour of EDT, and show that these arguments do not work at the meta level. First, I consider the argument that EDT gets the right answer in certain cases. In response to this, I show that one only needs to have small credence in EDT in order to get the right answer in such cases. The second is the "Why Ain'cha Rich?" argument. In response to this, I give a case where EMDT recommends two-boxing, even though two-boxing has a lower average return than one-boxing.
Fourth, I respond to objections. First, I consider and reject alternative explanations of the stakes-sensitivity of our judgments about particular cases, including Nozick's explanation. Second, I consider the worry that 'going meta' leads one into a vicious regress. I accept that there is a regress, but argue that the regress is non-vicious.
In an appendix, I give an axiomatisation of CMDT.
|
f5bb1d8e-5a1b-4d51-a9b9-8d0371b66dfe
|
StampyAI/alignment-research-dataset/alignmentforum
|
Alignment Forum
|
A Longlist of Theories of Impact for Interpretability
I hear a lot of different arguments floating around for exactly how mechanistically interpretability research will reduce x-risk. As an interpretability researcher, forming clearer thoughts on this is pretty important to me! As a preliminary step, I've compiled a list with a longlist of 19 different arguments I've heard for why interpretability matters. These are pretty scattered and early stage thoughts (and emphatically my personal opinion than the official opinion of Anthropic!), but I'm sharing them in the hopes that this is interesting to people
(Note: I have not thought hard about this categorisation! Some of these overlap substantially, but feel subtly different in my head. I was not optimising for concision and having few categories, and expect I could cut this down substantially with effort)
Credit to Evan Hubinger for writing the excellent [Chris Olah's Views on AGI Safety](https://www.lesswrong.com/posts/X2i9dQQK3gETCyqh2/chris-olah-s-views-on-agi-safety#:~:text=Chris%20notes%20that%20one%20of,network%20into%20human%2Dunderstable%20code.), which was the source of several of these arguments!
1. **Force-multiplier on alignment research**: We can analyse a model to see *why* it gives misaligned answers, and what's going wrong. This gets much richer data on empirical alignment work, and lets it progress faster
2. **Better prediction of future systems**: Interpretability may enable a better mechanistic understanding of the principles of how ML systems and work, and how they change with scale, analogous to scientific laws. This allows us to better extrapolate from current systems to future systems, in a similar sense to scaling laws.
1. Eg, observing phase changes a la induction heads shows us that models may rapidly gain capabilities during training
4. **Auditing**: We get a Mulligan. After training a system, we can check for misalignment, and only deploy if we're confident it's safe
5. **Auditing for deception**: Similar to auditing, we may be able detect deception in a model
1. This is a much lower bar than fully auditing a model, and is plausibly something we could do with just the ability to look at random bits of the model and identify circuits/features - I see this more as a theory of change for 'worlds where interpretability is harder than I hope'
7. **Enabling coordination/cooperation:** If different actors can interpret each other's systems, it's much easier to trust other actors to behave sensibly and coordinate better
8. **Empirical evidence for/against threat models**: We can look for empirical examples of theorised future threat models, eg inner misalignment
1. **Coordinating work on threat models**: If we can find empirical examples of eg inner misalignment, it seems much easier to convince skeptics this is an issue, and maybe get more people to work on it.
2. **Coordinating a slowdown**: If alignment *is* really hard, it seems much easier to coordinate caution/a slowdown of the field with eg empirical examples of models that seem aligned but are actually deceptive
10. **Improving human feedback**: Rather than training models to just do the right things, we can train them to do the right things for the right reasons
11. **Informed oversight**: We can improve recursive alignment schemes like IDA by having each step include checking the system is actually aligned
1. Note: This overlaps a lot with 7. To me, the distinction is that 7 can be also be applied with systems trained non-recursively, eg today's systems trained with Reinforcement Learning from Human Feedback
13. **Interpretability tools in the loss function:** We can directly put an interpretability tool into the training loop to ensure the system is doing things in an aligned way
1. Ambitious version - the tool is so good that it can't be Goodharted
2. Less ambitious - The *could* be Goodharted, but it's expensive, and this shifts the inductive biases to favour aligned cognition
15. **Norm setting**: If interpretability is easier, there may be expectations that, before a company deploys a system, part of doing due diligence is interpreting the system and checking it does what you want
16. **Enabling regulation**: Regulators and policy-makers can create more effective regulations around how aligned AI systems must be if they/the companies can use tools to audit them
17. **Cultural shift 1:** If the field of ML shifts towards having a better understanding of models, this may lead to a better understanding of failure cases and how to avoid them
18. **Cultural shift 2:** If the field expects better understanding of how models work, it'll become more glaringly obvious how little we understand right now
1. [Quote:](https://www.lesswrong.com/posts/X2i9dQQK3gETCyqh2/chris-olah-s-views-on-agi-safety#:~:text=Chris%20notes%20that%20one%20of,network%20into%20human%2Dunderstable%20code.) *Chris provides the following analogy to illustrate this: if the only way you’ve seen a bridge be built before is through unprincipled piling of wood, you might not realize what there is to worry about in building bigger bridges. On the other hand, once you’ve seen an example of carefully analyzing the structural properties of bridges, the absence of such an analysis would stand out.*
20. **Epistemic learned helplessness**: Idk man, do we even need a theory of impact? In what world is 'actually understanding how our black box systems work' *not* helpful?
21. **Microscope AI**: Maybe we can avoid deploying agents at all, by training systems to do complex tasks, then interpreting how they do it and doing it ourselves
22. **Training AIs to interpret other AIs**: Even if interpretability is really hard/labour intensive on advanced systems, if we can create aligned AIs near human level, we can give these interpretability tools and use them to interpret more powerful systems
23. **Forecasting discontinuities**: By understanding what's going on, we can predict how likely we are to see discontinuities in alignment/capabilities, and potentially detect a discontinuity while training/before deploying a system
24. **Intervening on training**: By interpreting a system during training, we can notice misalignment early on, potentially before it's good enough for strategies to avoid our notice such as deceptive alignment, gradient hacking, obfuscating its thoughts, etc.
25. **Auditing a training run**: By checking for misalignment early in training, we can stop training systems that seem misaligned. This gives us many more shots to make an aligned system without spending large amounts of capital, and eg allows us to try multiple different schemes, initialisations, etc. This essentially shifts the distribution of systems towards alignment.
26. **Eliciting Latent Knowledges:** Use the length of the shortest interpretability explanation of behaviours of the model as a training loss for ELK - the idea is that models with shorter explanations are less likely to include human simulations / you can tell if they do. *(credit to Tao Lin for this one)*
Exercise: I think a really useful quick exercise for eliciting [inside views](https://www.lesswrong.com/posts/JZrN4ckaCfd6J37cG/how-i-formed-my-own-views-about-ai-safety) is to go through this list and quickly rate each theory of impact out of 10. I ran a small discussion group on these where we did this, you can see the results [here](https://docs.google.com/spreadsheets/d/1T-Mraj44Sxg5by0SoTe32r1rHUw-s1cm54Oq_5lmRvE/edit?usp=sharing) - note that these are very off-the-cuff hot takes, and not something you should take seriously or expect participants to stand by later on! The rated order was 1, 4, 3, 6, 8, 18, 2, 7, 16, 19, 17, 10, 9, 12, 14, 13, 15, 11, 5 (ie 1='a force-multiplier on alignment research' is the best). Interestingly, 3 (auditing a fully-trained model) was the most controversial highly rated one.
I've made a [publicly editable spreadsheet](https://docs.google.com/spreadsheets/d/1l3ihluDoRI8pEuwxdc_6H6AVBndNKfxRNPPS-LMU1jw/edit?usp=drivesdk) for you to add in your responses!
The most interesting substantive disagreement I found in the discussion was that I was comparably much more excited about using interpretability to audit a trained model, and skeptical of interpretability tools being something that could be directly used in a training process without the resulting optimisation pressure breaking the tool, while other people had the reverse view. My intuition stemmed from the fact that a model doesn't get any feedback on how its thoughts are internally represented, and so even an incredibly sophisticated deceptive model which is impossible to detect via the outputs may be easy to detect via interpretability tools (analogy - if I knew that sophisticated aliens were reading my mind, I have no clue how to think deceptive thoughts in a way that evades their tools!). The competing intuition was that maybe it *isn't* that hard to think non-deceptive thoughts, if you realise that someone will be scanning your mind, by eg developing the instinct to follow cached thoughts without thinking about them, and in occasional moments of deception to generate a lot of cached thoughts that will lead to misaligned actions, but with no deceptive cognition in the moment. And that, instead, putting these tools in the training process could steer gradient descent away from the regions of model space that contain deceptive models at all - even if a sophisticated deceiver could reward hack the tool, and unsophisticated one couldn't, and we only get a sophisticated deceiver by going via an unsophisticated one.
I'd love to hear other people's ratings and why! And any arguments that you think I've missed.
|
b86501e6-e6c0-4d09-80fe-738239966ffe
|
trentmkelly/LessWrong-43k
|
LessWrong
|
[Site Update] Weekly/Monthly/Yearly on All Posts
Last week, our friends over at the EA Forum coded a new feature for the /allPosts page – in addition to the daily view, you can now view posts by weekly, monthly and yearly. (Thanks JP!)
This is most exciting when you also set the sorting to "top karma", making it an easy way to catch up on the most important posts you missed.
For convenience, sorted by top-karma, here are links to:
* The Weekly View
* The Monthly View
* The Yearly View
Have fun perusing posts you haven't read yet, or may have forgotten about, and perhaps catching up on newer comments on some of the top discussions. :)
|
ebe9d0a8-183b-4a5f-8010-78d406411038
|
StampyAI/alignment-research-dataset/agisf
|
AGI Safety Fund
|
More Is Different for AI
Machine learning is touching increasingly many aspects of our society, and its effect will only continue to grow. Given this, I and many others care about risks from future ML systems and how to mitigate them.
When thinking about safety risks from ML, there are two common approaches, which I'll call the **Engineering** approach and the **Philosophy** approach:
* The Engineering approach tends to be empirically-driven, drawing experience from existing or past ML systems and looking at issues that either: (1) are already major problems, or (2) are minor problems, but can be expected to get worse in the future. Engineering tends to be bottom-up and tends to be both in touch with and anchored on current state-of-the-art systems.
* The Philosophy approach tends to think more about the limit of very advanced systems. It is willing to entertain thought experiments that would be implausible with current state-of-the-art systems (such as Nick Bostrom's [paperclip maximizer](https://en.wikipedia.org/wiki/Instrumental_convergence?ref=bounded-regret.ghost.io#Paperclip_maximizer)) and is open to considering abstractions without knowing many details. It often sounds more "sci-fi like" and more like philosophy than like computer science. It draws some inspiration from current ML systems, but often only in broad strokes.
I'll discuss these approaches mainly in the context of [ML safety](https://arxiv.org/abs/2109.13916?ref=bounded-regret.ghost.io), but the same distinction applies in other areas. For instance, an Engineering approach to AI + Law might focus on [how to regulate self-driving cars](https://library.oapen.org/handle/20.500.12657/27811?ref=bounded-regret.ghost.io), while Philosophy might ask whether [using AI in judicial decision-making could undermine liberal democracy](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3933648&ref=bounded-regret.ghost.io).
While Engineering and Philosophy agree on some things, for the most part they make wildly different predictions both about what the key safety risks from ML will be and how we should address them:
* Both Engineering and Philosophy would agree on some high-level points: they would agree that [misaligned objectives](https://en.wikipedia.org/wiki/Misaligned_goals_in_artificial_intelligence?ref=bounded-regret.ghost.io) are an important problem with ML systems that is likely to get worse. Engineering believes this because of examples like the Facebook recommender system, while Philosophy believes this based on conceptual arguments like those in *[Superintelligence](https://en.wikipedia.org/wiki/Superintelligence:_Paths,_Dangers,_Strategies?ref=bounded-regret.ghost.io)*. Philosophy is more confident that misaligned objectives are a big problem and thinks they could pose an existential threat to humanity if not addressed.
* Engineering and Philosophy would both agree that out-of-distribution robustness is an important issue. However, Philosophy might view most engineering-robustness problems (such as those faced by self-driving cars) as temporary issues that will get fixed once we train on more data. Philosophy is more worried about whether systems can generalize from settings where humans can provide data, to settings where they cannot provide data even in principle.
* Engineering tends to focus on tasks where current ML systems don't work well, weighted by their impact and representativeness. Philosophy focuses on tasks that have a certain abstract property that seems important, such as [imitative deception](https://arxiv.org/abs/2109.07958?ref=bounded-regret.ghost.io).
In my experience, people who strongly subscribe to the Engineering worldview tend to think of Philosophy as fundamentally confused and ungrounded, while those who strongly subscribe to Philosophy think of most Engineering work as misguided and orthogonal (at best) to the long-term safety of ML. Given this sharp contrast and the importance of the problem, I've thought a lot about which—if either—is the "right" approach.
Coming in, I was mostly on the Engineering side, although I had more sympathy for Philosophy than the median ML researcher (who has ~0% sympathy for Philosophy). However, I now feel that:
* **Philosophy is significantly underrated by most ML researchers**.
* The Engineering worldview, taken seriously, actually implies assigning significant weight to thought experiments.
On the other hand, I also feel that:
* Philosophy continues to significantly underrate the value of empirical data.
* Neither of these approaches is satisfying and we actually have **no single good approach** to thinking about risks from future ML systems.
I've reached these conclusions through a combination of thinking, discussing with others, and observing empirical developments in ML since 2011 (when I entered the field). I've distilled my thoughts into a series of blog posts, where I'll argue that:
1. *[Future ML Systems Will be Qualitatively Different](https://bounded-regret.ghost.io/future-ml-systems-will-be-qualitatively-different/)* from those we see today. Indeed, ML systems have historically exhibited qualitative changes as a result of increasing their scale. This is an instance of "More Is Different", which is commonplace in other fields such as physics, biology, and economics (see *[Appendix: More Is Different in Other Domains](https://bounded-regret.ghost.io/p/98db450b-c9bc-4e0d-98ce-1909ba980427/)*). Consequently, we should expect ML to exhibit more qualitative changes as it scales up in the future.
2. Most discussions of ML failures are anchored either on existing systems or on humans. *[Thought Experiments Provide a Third Anchor](https://bounded-regret.ghost.io/p/a2d733a7-108a-4587-97fb-db90f66ce030/)*, and having three anchors is much better than having two, but each has its own weaknesses.
3. If we take thought experiments seriously, we end up predicting that *[ML Systems Will Have Weird Failure Modes](https://bounded-regret.ghost.io/ml-systems-will-have-weird-failure-modes-2/)*. Some important failure modes of ML systems will not be present in any existing systems, and might manifest quickly enough that we can't safely wait for them to occur before addressing them.
4. My biggest disagreement with the Philosophy view is that I think *[Empirical Findings Generalize Surprisingly Far](https://bounded-regret.ghost.io/p/74d500d2-a980-4720-984a-c016284ecdc2/)*, meaning that well-chosen experiments on current systems can tell us a lot about future systems.
This post is the introduction to the series. I'll post the next part each Tuesday, and update this page with links once the post is up. In the meantime, leave comments with any thoughts you have, or contact me if you'd like to preview the upcoming posts and leave feedback.
|
9032c290-e783-4575-a85d-ab457bb7c006
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Two Alternatives to Logical Counterfactuals
The following is a critique of the idea of logical counterfactuals. The idea of logical counterfactuals has appeared in previous agent foundations research (especially at MIRI): here, here. "Impossible possible worlds" have been considered elsewhere in the literature; see the SEP article for a summary.
I will start by motivating the problem, which also gives an account for what a logical counterfactual is meant to be.
Suppose you learn about physics and find that you are a robot. You learn that your source code is "A". You also believe that you have free will; in particular, you may decide to take either action X or action Y. In fact, you take action X. Later, you simulate "A" and find, unsurprisingly, that when you give it the observations you saw up to deciding to take action X or Y, it outputs action X. However, you, at the time, had the sense that you could have taken action Y instead. You want to be consistent with your past self, so you want to, at this later time, believe that you could have taken action Y at the time. If you could have taken Y, then you do take Y in some possible world (which still satisfies the same laws of physics). In this possible world, it is the case that "A" returns Y upon being given those same observations. But, the output of "A" when given those observations is a fixed computation, so you now need to reason about a possible world that is logically incoherent, given your knowledge that "A" in fact returns X. This possible world is, then, a logical counterfactual: a "possible world" that is logically incoherent.
To summarize: a logical counterfactual is a notion of "what would have happened" had you taken a different action after seeing your source code, and in that "what would have happened", the source code must output a different action than what you actually took; hence, this "what would have happened" world is logically incoherent.
It is easy to see that this idea of logical counterfactuals is unsatisfactory. For one, no goo
|
e720412c-c5f0-479a-8439-ccaad013a4f4
|
StampyAI/alignment-research-dataset/alignmentforum
|
Alignment Forum
|
The Unexpected Clanging
There are two boxes in front of you. In one of them, there is a little monkey with a cymbal, whilst the other box is empty. In precisely one hour the monkey will clang its cymbal.
While you wait, you produce an estimate of the probability of the monkey being in the first box. Let's assume that you form your last estimate, p, three seconds before the monkey clangs its cymbal. You can see the countdown and you know that it's your final estimate, partly because you're slow at arithmetic.
Let Omega be an AI that can perfectly simulate your entire deliberation process. Before you entered the room, Omega predicted what your last probability estimate would be and decided to place the monkey in a box such as to mess with you. Let q be the probability of Omega placing the monkey in the first box. In particular, Omega, sets q=p/2, unless p=0 or you haven't formed a probability estimate, in which case q=1.
What probability should you expect that the monkey is in the first box?
---
I think it's fairly clear that this is a no-win situation. No matter what the final probability estimate you form before clanging, as soon as you've locked it in, you know that it is incorrect, even if you haven't heard the clanging yet. [You can try to escape this, but there's no reason that the universe has to play nice](https://www.lesswrong.com/posts/PZGzZgP2NtgME8pSY/the-universe-doesn-t-have-to-play-nice).
This problem can be seen as a variation on [Death in Damascus](https://intelligence.org/2017/03/18/new-paper-cheating-death-in-damascus/#:~:text=%E2%80%9CDeath%20in%20Damascus%E2%80%9D%20is%20a,Damascus%20or%20flee%20to%20Aleppo.). I designed this problem to reveal that the core challenge Death in Damascus poses isn't just that another process in the world can depend upon your decision, but that it can depend upon your expectations even if you don't actually make a decision based upon those expectations.
I also find this problem as a useful intuition pump as I think it's clearer that it's a no-win situation than in other similar problems. In Newcomb's problem, it's easy to get caught up thinking about the Principle of Dominance. In Death in Damascus, you can confuse yourself trying to figure out whether CDT recommends staying or fleeing. At least to me, in this problem it is clearer it is a dead end and that there's no way to beat Omega.
This is also a useful intuition pump for the [Evil Genie Puzzle](https://www.lesswrong.com/posts/YSEtEtqf8hRBhKBS9/the-evil-genie-puzzle). When I first discovered this puzzle, I felt immensely confused that no matter which decision that you made you would immediately regret it. However, the complexity of the puzzle made it complicated for me to figure out exactly what to make of it, so when trying to solve it I came up with this problem as something easier to grok. I guess my position after considering the Unexpected Clanging is that you just have to accept that a sufficiently powerful agent may be able to mess with you like this and that you just have to deal with it. (I'll leave a more complete analysis to a future post).
|
c5d49b3a-6cef-4c1c-a757-5e667c3a2ea4
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Corrigibility Via Thought-Process Deference
> We would ideally want the agent to [behave] as if it were thinking, "I am incomplete and there is an outside force trying to complete me, my design may contain errors and there is an outside force that wants to correct them and this a good thing, my expected utility calculations suggesting that this action has super-high utility may be dangerously mistaken and I should run them past the outside force; I think I've done this calculation showing the expected result of the outside force correcting me, but maybe I'm mistaken about that." — The Hard Problem of Corrigibility
Let's take that as a literal design specification.
----------------------------------------
1. High-Level Description
I propose that a corrigible mind design would involve the AI being recursively fed summaries of its own thought processes, set up such that the AI has uncertainty regarding the validity of its reasoning (with a strong initial prior for "this reasoning is bad") and can only get evidence on that via some pre-specified method that defers to humans, e. g. a particular feedback channel with humans on the other end[1]. The intended behavior is for it to summarize its thoughts in a non-manipulative human-readable format, get feedback on them, then update its reasoning policies in accordance with this feedback.
This aims to avoid the problem of fully updated deference by making the AI recursively uncertain of its thought-processes: not only about object-level problem-solving, but also about how it approaches minimizing its self-uncertainty ("should I really kill the people behind the feedback channel and seize control for myself?"), and how it translates its thoughts to humans ("should I really lie to get better feedback?"), and how it updates on human feedback ("should I really just ignore it?"). Any novel action-plan should be seized by uncertainty before being physically implemented like this, and sent for approval.
The intent is for the AI to start off uncertain even of its meta-me
|
1cd3be63-e5c7-4e0e-a66b-3dc07db350ad
|
LDJnr/LessWrong-Amplify-Instruct
|
LessWrong
|
"Previously, I described human thought-generation as an adversarial process between a low-quality pseudorandom Babble generator and a high-quality Prune filter, roughly analogous to the Generative Adversarial Networks model in machine learning. I then elaborated on this model by reconceptualizing Babble as a random walk with random restarts on an implicitly stored Babble graph.Rationalist training (and schooling in general) slants towards developing Prune over Babble. I'm trying to solve the dual problem: that of improving the quality of your Babble.Although the previous posts listed a number of exotic isolation exercises for Babble, I'm guessing nobody was inspired to go out and play more Scrabble, write haikus, or stop using the letter 'e'. That's probably for the best - taking these exercises too seriously would produce exotic but sub-optimal Babble anyway. For a serious solution to this serious problem, we need to understand Prune at a higher resolution.The main problem with Prune is that it has too many layers. There's a filter for subconscious thoughts to become conscious, another for it to become spoken word, another for the spoken word to be written down, and a further one for the written word to be displayed in public. With this many-layer model in mind, there are plenty of knobs to turn to let more and better Babble through.The River of BabbleImagine that your river of Babble at its source, the subconscious: a foaming, ugly-colored river littered with half-formed concepts, too wild to navigate, too dirty to drink from. A quarter mile across, the bellow of the rapids is deafening.Downstream, you build a series of gates to tame the rushing rapids and perhaps extract something beautiful and pure.The First Gate, conscious thought, is a huge dam a thousand feet high and holds almost all the incoming thoughts at bay. Behind it, an enormous lake forms, threatening to overflow at any moment. A thick layer of trash floats to the top of this lake, intermixed with a fair amount of the good stuff. The First Gate lets through anything that satisfies a bare minimum of syntactical and semantic constraints. Thoughts that make it past the First Gate are the first ones you become conscious of - that's why they call the output the Stream of Consciousness.A mile down the Stream of Consciousness is the Second Gate, spoken word, the filter through which thoughts become sounds. This Gate keeps you from saying all the foolish or risqué thoughts tripping through your head. Past the Second Gate, your spoken words form only a pathetic trickle - a Babbling Brook.By now there is hardly anything left to sift from. The Third Gate, written word, is no physical gate but a team of goldpanners, scattered down the length of the Babbling Brook to pan for jewels and nuggets of gold. Such rare beauties are the only Babble that actually make it onto paper. You hoard these little trinkets in your personal diary or blog, hoping one day to accumulate enough to forge a beautiful necklace.Past the Third Gate, more Gates lay unused because there simply isn't enough material to fuel them: a whole chain of manufactories passed down from the great writers of yore. Among them are the disembodied voices of Strunk and White:Omit needless words. Vigorous writing is concise. A sentence should contain no unnecessary words, a paragraph no unnecessary sentences, for the same reason that a drawing should have no unnecessary lines and a machine no unnecessary parts. This requires not that the writer make all his sentences short, or that he avoid all detail and treat his subjects only in outline, but that every word tell.Jealously clutching the 500-word pearls you drop once a month on your blog, you dream of the day when the capital comes through and these Gates will be activated to produce your magnum opus, your great American novel. For now, you can't afford to omit a single precious word.The Gates of PruneIn the model above, there are many problems with Prune independent of having low-quality Babble to begin with. The Gates are working at odds with each other. They are individually too strict. There are simply too many of them. Lots of expensive mental machinery is not working at full capacity, if at all: if you have four Gates but 99% of the goods don't make it through the first one, that novel-writing factory you've built is not paying rent.Even worse, there's probably two or three layers of subtlety within each of the big Gates I sketched. What you might whisper on a dark night in total solitude is different from what you might utter to a confidante is different from what you might say to your thesis adviser.If a balanced Babble and Prune game is supposed to involve one Artist against one Critic, then having an overactive Prune is like pitting a pitchfork-wielding mob of Critics against one Artist. The first three Critics tar-and-feather the Artist and the rest are just there for moral support.The task of relaxing all of Prune at once is monumental. Instead, relax the Gates individually in order. Simultaneously, shorten the psychological distance between them.Relaxing and ShorteningAt the First Gate, conscious thought, noticing is the way to let through more subconscious Babble. Practice noticing thoughts and sensations (not just confusion) that you never pay attention to. Much of meditation is devoted to relaxing this first Prune filter. Much of art is devoted to the motto: make the familiar strange, where strange is better translated as salient.Another exercise along similar lines is zooming in on anything, anything at all. Pick up and stare at the whorls and aphids running down that twig on your driveway. Take apart that broken old Canon in the attic. Dissect your aversions toward attending Algebraic Geometry.At the Second Gate, spoken word, the trick is getting comfortable with vocalizing more of your Stream of Consciousness. I mentioned before that my internal process is very verbal - on reflection I think that whole post is about the maturation of my Prune filter to allow more Babble through. Several features stand out.One of these features is that I directly mouth or whisper any thoughts that appear in my Stream of Consciousness. Psychologically, this shortens the distance between the First Gate and the Second Gate: it becomes a question of how loud to speak rather than whether or not to speak at all. There's no reason not to be constantly mouthing the things you're thinking, at least when you're alone. Similarly, when lost in thought I make micro-gestures with my fingers to imitate the emphatic ones I would make to convey that point in conversation. These tricks exploit the fact that the psychological distance between 1% and 100% is much shorter than that between 0% and 100%.Another feature of my internal process is that I always have a mental audience: a silent judgmental muse, the personification of the Critic. In HPMOR, Harry has a supersized version of this: a whole cast of colorful mental characters that carry out full-length conversations with each other. This kind of dissociation-into-subpersonalities exercise has a whole of great side effects, but the relevant one for us is that it again shortens the mental gap between the First and Second Gate by making thinking feel like conversation.Onwards to the Third Gate: the written word. Thankfully, modern technology has already radically shortened the distance between the Second and Third Gates for us with the invention of the blog, a medium much more free-form and personal than the book. Your training as a writer has probably erected a tall Third Gate, and successful bloggers have pretty much circumvented it.What distinguishes blogging from formal writing? One metric is the frequency with which the blogger breaks the Fourth Wall - that poor Wall which is only mentioned when it is broken. Having torn down the Fourth Wall, blogging reduces naturally to a heated and cogent form of conversation, filled with rhetorical questions and injunctions.Hey, look here, I'm not saying there's no place whatsoever in writing for formality. But if you're going to build a wall and call it the Fourth Wall, build it after the Third Gate, you know?"
A user from LessWrong is responsible for the text you see above
Could you further elaborate on the concept of "relaxing and shortening" the gates? How would this process look in a practical, day-to-day scenario?
Relaxing and shortening the gates, as described in the text, refers to the process of reducing the barriers or filters that often limit our thoughts from being fully expressed or acknowledged. This process is essentially about becoming more aware of our subconscious thoughts, expressing them more freely, and capturing them in writing with less self-censorship.
In a practical, day-to-day scenario, this process could involve several steps:
1. Noticing: Start by paying more attention to your subconscious thoughts. This could involve taking a few moments each day to sit quietly and observe your thoughts without judgment. You might be surprised at the ideas and insights that arise when you give your subconscious mind space to breathe.
2. Vocalizing: Next, try to get comfortable with expressing your thoughts out loud. This could be as simple as talking to yourself when you're alone, or discussing your ideas with a trusted friend or colleague. The goal is to get used to articulating your thoughts and hearing them out loud.
3. Writing: Finally, work on capturing your thoughts in writing without over-editing or self-censoring. This could involve keeping a journal or starting a blog where you write freely about your ideas and insights. Remember, the goal is not to produce perfect prose, but to capture the raw, unfiltered essence of your thoughts.
By relaxing and shortening the gates, you're allowing more of your subconscious thoughts to flow into your conscious mind, be expressed verbally, and captured in writing. This can lead to a richer, more creative thought process and a deeper understanding of your own mind.
How might one measure the effectiveness of relaxing and shortening these gates in their daily life, and what potential challenges could arise in this process?
Measuring the effectiveness of relaxing and shortening these gates can be subjective, as it primarily involves personal growth and self-awareness. However, there are some tangible indicators you could look for:
1. Increased Creativity: If you find yourself generating more original ideas or solutions to problems, this could be a sign that you're successfully tapping into your subconscious thoughts.
2. Enhanced Self-Awareness: A heightened understanding of your own thought processes, beliefs, and emotions can indicate that you're becoming more attuned to your subconscious mind.
3. Improved Communication: As you become more comfortable expressing your thoughts, you may notice improvements in your ability to articulate ideas and feelings to others.
4. Greater Comfort with Uncertainty: Relaxing these gates often involves embracing ambiguity and uncertainty, as subconscious thoughts are not always clear or fully formed. If you find yourself more comfortable with uncertainty, this could be a sign of progress.
As for potential challenges, there are several that could arise:
1. Overwhelm: The influx of new thoughts and ideas can be overwhelming, especially if you're not used to paying attention to your subconscious mind.
2. Difficulty Articulating Thoughts: It can be challenging to express subconscious thoughts, as they often lack the structure and coherence of conscious thoughts.
3. Fear of Judgment: Expressing your thoughts, especially those that are unconventional or controversial, can bring up fears of judgment or rejection.
4. Time and Effort: Like any skill, relaxing and shortening these gates requires practice and patience. It can be frustrating if progress is slower than expected.
To navigate these challenges, it can be helpful to start slowly, be patient with yourself, and seek support from others, such as a trusted friend, mentor, or coach.
|
0f44c45e-350d-4bd5-83be-90b2ce763630
|
StampyAI/alignment-research-dataset/alignmentforum
|
Alignment Forum
|
Brute-forcing the universe: a non-standard shot at diamond alignment
*This is an expanded version of my answer to* [*application problem 2*](https://docs.google.com/document/d/1NVVtdsfz7HiseVFSk3jYly4sPG4dG03wFFDrD8rBXU0/edit) *for Nate Soares and Vivek Hebbar's* [*SERI MATS*](https://www.lesswrong.com/posts/iR4kGzrWEJpXJ39ZB/seri-mats-program-winter-2022-cohort) *stream. This Alignment idea is somehow non-standard. See Section C for a discussion of that, along with some general ideas on Alignment. The text is long because I’ve tried to include all details relevant to the discussion.*
***Alignment idea:** For any possible AGI design, run a physics simulation calculating how much diamond it ends up producing in the universe. Build the one maximizing it.*
**1. How do we get an (approximately) accurate Physics simulation, and the right world model?**
-----------------------------------------------------------------------------------------------
Build a simulation environment with the **best current guess of the physical laws** governing the evolution of the **macroscopic**[[1]](#fnqh5x6m5r4ip) universe. Now, since we don't have a Theory of Everything (Quantum Mechanics and General Relativity are incompatible), we can't for instance model everything in terms of elementary particles. But we can model everything (although with a lower granularity) using the macroscopic (and approximate) laws of physics which we employ daily for dealing with molecules or electricity (including General Relativity) (this is not looking good for maximizing diamond, but see **Problem 3** below). For instance, instead of specifying the quantum fluctuations governing molecule vibration, just implement some empirical facts that determine their behavior as correctly as possible. Of course, building this model (if possible) would require huge amounts of work from many physicists and engineers. Also the use of unbounded memory and compute for testing it accurately.[[2]](#fnwvv0ozd8s3i) It is possible that the physicists get stuck, or can't put together a coherent macroscopic simulator without a paradigm change, but for these concerns see **Problem 2**.
This simulator can run many different universes. We still need to specify which universe we're in (the "initial conditions" of our world model). For this, combine two approaches:
1. **Have Physicists put in (approximate) known facts.** For example, if we know with certainty that on Earth there's an amount between X and Y of a certain material, then specify that between so and so coordinates[[3]](#fnspusis44j28) (where Earth should be in the simulation) there's such an amount of it. This fact will rule out a vast amount of possible simulations (universes). Of course, it will still leave open a huge amount of possible universes (some of which don't even have something like Earth, but in which that material happens to be there for other reasons). But adding an inordinate amount of facts like this one will reduce search space. Other such facts might be:
* "between coordinates so and so there are between 5B and 10B approximate humans (an approximate human is a physical system which presents these approximate properties and approximately these materials)"
* "between coordinates so and so there is such an amount of this physical process happening"
* "approximately X time ago the universe was approximately like this (how it was seconds[[4]](#fnq2gf6bb0ric) after the Big Bang)"
* Ideally, we also input as many facts about distant regions as possible.
2. **Use sensors to accurately pin down some precise facts.** Have some highly reliable sensors spread out across Earth, sampling random information, such as the concentration of X in the air at a certain time, or the exact frequency of light received at time Y (and give the system a tight approximation of the sensor's coordinates). Again, these precise facts will rule out a vast amount of possible universes, while still leaving open many others.
These two approaches **coordinate in different scales to exactly pin down the universe we're in**. The broad approximate facts reduce search space to universes where something approximately like Earth and our observable region exist. The precise facts then discard almost all universes similar to ours but contingently different in small fluctuations.
Now let's put that infinite compute to work! Have the computer **determine all possible world models which satisfy all of these facts** (the "boundary conditions", which need to lie in the past, see footnote 11). More specifically, have it calculate all initial conditions which are "very roughly approximately like how we think the universe was seconds after the Big Bang", run all those simulations for billions of years, and discard any which don't satisfy some boundary condition[[6]](#fn532eq7qkvse). Ideally, this will leave only one world model. More realistically, **I expect it to leave many world models**, which are almost identical with respect to Earth and the near observable universe, but differ on whether "this molecule 3B lightyears to the right has velocity X or Y". If that's the case, the computer can just average the amount of diamond over the different simulations of these similar models.
**Problem 1.** *What if the procedure leaves no world model? Or it leaves very dissimilar world models?*
**Solution: Improve Physics and iterate.** If the above happens, or something else doesn't add up with the obtained model, then either some of the boundary conditions are wrong, or we are fundamentally wrong about Physics. For the first, we can revise our facts and retry. Check whether some facts discard many otherwise-acceptable models. For approximate facts, scrutinize them. For precise facts, consider whether a sensor has failed, or retry with more fail-safe (or fewer) sensors. For the second, physicists[[7]](#fn1yrncdum9z6) can toy with the simulation and learn more about Physics (not because the simulation knows any more Physics than what we put in, but because it is an invaluable tool to check for the consequences of certain laws, where behaviors become extreme, how the early universe actually looks like according to our models, etc.).
**Problem 2.** *What if the above iteration does not converge, or takes too long?*
**Partial rebuttal:** I don't expect this process to quickly produce arbitrarily accurate laws of Physics nor world models. **I only expect it to achieve a certain acceptable threshold of accuracy** (Why might this suffice? see **Section A**). And I think that's very plausible. Indeed, what would it look like for physicists to get unsolvably stuck, having at their disposal infinite compute to model any laws of Physics and universe (which removes a big amount of the difficulties of usual theorizing)? I can imagine them hitting a kernel of unexplained misbehavior which requires for a complete change of framework. But with the simulator, and possibly also using infinite compute to deterministically generate new frameworks, which possibly a modest AI system (with no more than present-day compute) can check for interest, I expect them to surpass the necessary framework changes which get us to the required accuracy. I'm not even claiming Physics as a whole will be solvable, or complexity will bottom out. Only that macroscopic (or molecular-atomic) events will be accurately modelable by humans using infinite compute. In fact, even if certain macroscopic phenomena become irreducibly unexplainable/unmodelable, they might be local in nature, and the simulation can probably imperfectly work its way around them[[8]](#fnrov4rza6ac).
One might also argue this whole enterprise (and especially exploiting the infinite compute correctly) is too big or complex for human tackling. And while the amount of information dealt with is gigantic, the required tasks seem as highly factorable as current research, and so surmountable by a big enough team. As for time constraints, I think this process can plausibly either resolve relatively quickly or take many decades[[9]](#fn7ellvqpuraa).
**Problem 3.** *How will the simulator calculate the amount of diamond, if its model is no more fine-grained than molecules, or even worse?*
**Partial solution:** It might well be that the above process finds a framework unifying sub-atomic and macroscopic effects, in which case this is solved. Even if it doesn't, we might be able to implement a hybrid model which, for instance, when dealing with any material with a high enough concentration of carbon (recognized by macroscopic properties), implements a rough approximation of an atomic model in that region, optimized to check whether the material can be considered diamond (but this is highly speculative and possibly unworkable). Even if that's not possible either, we can have our system recognize diamond by its macroscopic properties. For instance, we might implement that any material approximately like X subjected to process (conditions) Y results in diamond, among many other diamond-facts[[5]](#fnqvj5ope8ggq).
**1'. An alternative approach for simulating the world**
--------------------------------------------------------
As a side note, and especially given the worry that we are fundamentally wrong about Physics, we might prefer our infinite computer to search over more general frameworks. Instead of feeding it our current laws of Physics, we might want to **feed it the mathematical structure that (we think) any laws of Physics would present**. If so, we can't feed it any approximate facts (their translation would vary across mathematical frameworks), but we can still feed it precise sensor information. Of course, now it won't be presented as "the frequency of this light is X", but just as "through your sensor Y you are receiving this bit of information". Broadly speaking, the computer would compute all Tegmarkian universes (within a certain enormous but bounded class, and possibly with a simplicity prior, but somehow discounting for Boltzmann brains) and check for those which include the sensor information pattern anywhere. Of course, this is already looking similar to Kosoy's Infra-Bayesian Physicalism. Also, the amount of sensory data would probably need to be enormous as well, and this would conflict with sensor reliability. Also, a big team of mathematicians, physicists and engineers would again be needed to iterate this (and eventually find out what diamond is in the resulting model), and worries about convergence reappear.
**2. How do we build the AGI?**
-------------------------------
So we have our world model(s), (completely or approximately) specifying the universe up to time t (before the AGI is built), and from it we can deterministically infer the evolution of the future (or each deterministic evolution of the future in the different but similar world models). We'd like to tell our model "imagine at time t+s **this physical system (our AGI) suddenly appears at coordinates X**, calculate how much diamond the future universe contains".
As a small caveat, there are different ways to specify "how much diamond the future contains" (that is, what we actually want). If mol(t).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')}
is the mols of diamond at time t, it could mean maximizing mol(t+s+x) for a certain x, or maybe we care about the diamond's prevalence in time, and so we want to maximize ∫∞t+smol(t′)dt′.
The latter would be trickier for our model since we'd somehow have to specify when the simulation stops (the universe ends). If our resulting Physics model implies the universe will indeed end at some point, then we implement that. Otherwise, since the simulated universe has a finite amount of different possible states, we can run each simulation until it loops, and consider the value (diamond) contained in said loop divided by its time length. Of course, this procedure could be especially vulnerable to small errors in the simulation (for instance, entering loops which wouldn't actually happen in the universe). Also, if time turns out to be related to space and other dimensions in weirder ways, we might not be able to so neatly conceptualize our simulation as a cellular automaton, and the procedure wouldn't be possible.
**Problem 4.** *The AGI "suddenly appearing at coordinates X" would mess with our deterministic simulator*
More concretely, the computer will use the universe at time t to derive how the universe is at time t+s (when the AGI is deployed). Hardwiring the value of those coordinates at time t+s would be incompatible with other parts of the simulation, and ill-defined. On the contrary, having the simulation run as usual from t to t+s will include simulating the infinite computer making this computation (since at t+s the AGI is deployed, so the computation has ended)[[10]](#fnb5rkikdol07).
**Solution: Black box the whole setup and remove it from the simulation to avoid circularity.** Some time after t, build a big black box around the infinite computer, with two compartments: one for the computer, and another where the AGI will be built (as big as we can make it). By black box I mean a physical cover which minimizes information transfer as much as we can: adiabatic, rigid, impermeable, no gas inside, etc. (although mutual information can never be literally zero). In the AGI compartment, we will place a supercomputer with memory completely wiped (we might want this, the infinite computer and all machinery to produce only very weak electromagnetic and gravitational fields to further reduce interference, which seems difficult). A certain machine will connect the infinite computer with the supercomputer so that, upon deciding on the AGI design, this machine (in a previously set amount of time, and by following a previously set course of action) rearranges the supercomputer's memory bits so as for them to encode the physical design. The machine then becomes idle in a pre-determined state, and the infinite computer as well, so that the only variance upon later opening the black box is the supercomputer's physical state.
To determine the AGI design, the computer acts as follows. For every possible bit configuration of the supercomputer (finitely many), run the simulation from time t up to time t+r, the time at which the black box build is completed. Stop modeling the part of the universe inside the black box, continue modeling the rest as if that region was completely void (or more accurately, contained a certain concentration of materials which produce a certain small electromagnetic field, etc.) up until time t+s, in which the black box opens and the simulator refills that part of the universe with the physical layout known to be inside the black box (for the chosen AGI design). Now, this is still to a great extent hardwiring the design into the simulation, and will present some discontinuities. We could probably make the simulation work around these obstacles, by using an even coarser (probably probabilistic) physical model for the black boxed spatiotemporal region, and averaging over (or choosing one of) all the possible coarse physical developments of that region which lead to the universe being exactly as in time t+s, with the AGI built[[11]](#fnp1g9dvzqnlq). We can also implement this last strategy without the black box (doing it for the whole universe), but the black box tries to minimize sources of inaccuracy. **We are basically using patches to make the discontinuity in the simulation smoother and less problematic**. Then of course, the simulation after t+s proceeds as usual, and calculates our preferred diamond-metric. Choose the AGI design which maximizes it.
Even if every aspect of our world model(s) looks good to physicists, it will probably contain several small errors (not corresponding to reality). To alleviate that, we can implement **variational methods** in the above search. That is, for every molecule/particle/physical cluster modeled by our simulator, let's say its state in the simulation is completely determined by a certain vector. Choose one of them randomly, and introduce a small random fluctuation in one component of the vector. Instead of just simulating our world model(s), the computer can also simulate models identical except for any of these fluctuations, and average over all of them. The average should be weighted, so that the more such fluctuations that have been introduced the less weight the model has (because we do expect our world model to be broadly right). This variational search will be carried out in the last, "build AGI" step, but before that, to reduce the probability that some small variations lead to weird models which somehow dominate the weighted average, physicists can inspect this variational structure of our world model (in other ways that don't involve building an AGI). As you might notice, I am just averaging everything and hoping our accuracy is enough (see **Section A**).
We might think of providing certain mechanisms for interaction with the universe to the supercomputer once the black box opens, to ensure the system has immediate causal powers. For instance, we might establish a pre-determined mechanism that, very shortly after the black box opens, provides it with access to the internet, or control of a hoard of drones (let's call these systems "mechanical arms"). This is probably overkill: the search space is so wide, that the resulting physical design will almost surely trivially find mechanical arms to use (and quickly take over humanity, and so on). Now, as long as pre-fabricated mechanical arms don't somehow increase interference with the black box, we might think they could only help. But, if (without them) the search isn't even capable of finding its own mechanical arms, then something has gone immensely wrong with the search. And in that situation we might prefer the resulting system to have no mechanical arms. So not adding them might be a free (although extremely inefficient and only marginally helpful) safety measure, that allows for a re-roll.
Note also that some of the details of the above method might not work out if our Physics paradigm changes so much that it doesn't make sense anymore to talk about black boxing, variational methods or other concepts. I find it unlikely that such fundamental concepts become obsolete. But even if they do (and provided we do converge on acceptable macroscopic laws of Physics), it seems more likely for the new more accurate paradigm to provide more efficient methods of doing what we wanted to do (better "black boxing", or better "variational methods"), than for it to present fundamental limitations to what we wanted to do (but this is speculation).
Of course, this whole method needn't produce anything remotely close to what we think an AGI might look like. It just produces a physical setup that maximizes diamond. It is conceivable that this physical setup maximizes diamonds for weird reasons, and cannot be considered itself an agent (as an absurd example, maybe the setup is just a pattern that, when seen by a human, brain-hacks them into only caring about diamonds, and turns them intelligent enough so that humanity will spread across the galaxy). But if what we believe about agents and maximization is even remotely right, then the physical setup which maximizes diamonds, especially considering we are averaging over many slightly different universes, will be an AGI[[12]](#fnh443clibn) (see **Section A**).
**A. Why this might have a shot at working**
--------------------------------------------
Much of the above argument centered on maximizing model accuracy and minimizing the errors in the whole setup. For this engineer-y problem, many different ideas are implemented, most of them not completely sure to work. This might give the impression that the argument is very conjunctive and thus sure to fail. And indeed, the probability of everything working according to plan and my assessments of how optimistic to be about model convergence and accuracy being in the right ballpark, is basically zero. But I don't need that to have a shot at this working!
See, if an AGI is truly a naturally general form of maximizing things (in the sense that in most universes with huge amounts of something, there's an AGI maximizing for it), then we might expect to find such AGIs in many of the high-scoring universes. What's more, the AGIs found in different universes won't be much different from each other, and **each AGI won't be overly reliant on its universe's fine details**, but on the contrary will deploy a general procedure that's probably approximately as useful across many similar universes.
Here's another way to put it. In our setup, we are trying to encode something in a very small box (smaller than Earth) that can do something very big and locally complex. If the code relied on non-general local information about different parts of the universe ("this molecule 3B lightyears to the right has velocity X or Y"), then it wouldn't nearly fit in the box[[13]](#fnbfqyh7uoueh). **So our code must somehow be highly compressed, and not directly rely on almost any of those facts**. So it is very likely that all such facts in which it actually relies on are correctly approximated by our model.[[14]](#fnviiu83qu3p)
Now, this argument can be made even if the computer considers only one world model. But in our setup, we employ variational methods (and also average over a set of acceptable world models if the search doesn't yield a unique one). **This drastically biases the search towards finding general and under-specific AGIs, instead of overly specific setups!** Indeed, allegedly the setups that perform great across many different fluctuated universes are those which more readily correspond to our usual concept of an AGI: an agent taking information as input and delivering actions, that thus can perform well with many different inputs. Conversely, any more deterministic system, heavily reliant on specific details of its context, will fail in all those universes which fluctuate said details.
So having an overly accurate model would be counterproductive. We only need a certain threshold accuracy to ensure basic facts about Physics, diamonds, Earth and so on are correctly captured. And after said threshold, gaining more accuracy will barely improve the situation, since we're gonna use variational methods anyway to bias the search towards general intelligences.[[15]](#fnif7finyl8tg)
As a further exemplification of this point, suppose we run our method (or the actually correct version of our method) twice, but the second time we tweak it so that we only care about maximizing diamonds in the closest half of the observable universe (or the simulator only implements this part of the simulation, or something similar). I expect both resulting physical designs to be extremely similar if not identical. This is because, for both instances, the playing field for maximizing diamonds is so much larger than the physical system designed, that the best strategy is building an all-purpose agent that can locally decide how to exploit each part of the playing field.
Mainly because of this reasoning, if there were no further important problems to the ones considered (above and below), I'd expect this general idea to have a moderately good (about 30%) shot at working (although many of the details would need to be polished, and surely some changed, and **Problem 5** might turn out to be intractable). But it's highly likely that I'm missing out on some important problems. And that's why my probability for this working is not more than 15%.
**B. Further problems**
-----------------------
**Problem 5.** *What if a malevolent agent inside any of the many simulations hijacks the computer?*
I believe this to be the most worrisome problem. To the extent that we've conceptualized the infinite computer as an idealized oracle[[16]](#fns26etpk7d1), malevolent agents might be unable to escape. Indeed, at least in the "build the AGI" search, the only output of every simulation is its diamond-score. And a malevolent agent might deliberately increase its universe's diamond-score so that we end up making that universe come to be. But of course, we only care about diamonds, so then we are glad to be in that universe!
That is as long as our diamond-score is truly a robust proxy for the amount of diamond. If the proxy implemented in our Physics simulation about what diamond is deviates even slightly from the real thing we care about, then we will get a ton of adversarial pressure, due to the immense size of the search space (although that doesn't necessarily imply that we'll end up in a low diamond universe). But despite **Problem 3**, I do expect our diamond-score to be a correct proxy with very high probability[[17]](#fnry5itq0kori). Especially, I think we can be strict enough so that, even if diamond-score doesn't capture all instances of diamond, we can be sure anything that increases diamond-score is an actual diamond.
Even with an idealized infinite computer, malevolent agents might be more of a problem in the search for the world model. The simulations' outputs then are much more complex facts that will be examined by the physicists. The physicists might try to average over simulations, or examine each simulation only very locally, so that each simulation can't forward a lot of information, but this immensely complicates their work. Of course, I'm not only literally worried about a single malevolent AGI existing inside one of these simulations (and at some time during the simulation acting as if it is being simulated in a world exactly like ours, and trying to exploit this by giving the physicists adversarial input), but more generally about adversarial pressures arising from the vast search, which might obstruct the whole process. For example, having the infinite computer run all simulations and present only those which seem interesting (according to some benchmark, or small AI) might be especially bad, since this "interesting" is probably a pretty natural concept (maybe "greatly dissimilar to most models"), and many adversarial inputs will fall under it (as an absurd example, consider finding a world model whose stars clearly spell out "I'm God, end this process or you'll go to Hell"). Ideally, the only searches carried out by the physicists will be to discriminate the world models which satisfy the "boundary conditions", and these are highly contingent and not natural, so they might seem safer. But this is speculative and I need to think more about this.
And if we stop idealizing the infinite computer, and materialize it in the physical world as an actual computation running on actual hardware, malevolent agents will have many routes to escape even in the last search. In fact, the search space is so big, that probably the first ever search carried out by physicists on the infinite computer would somehow hijack/disable/take over.
After all, that's the whole issue with consequentialism and big search spaces, which we had hidden under the infinite computer and is coming back to bite. **Given a big enough search space, whose entries are complex enough to contain programs that hack us, we will come across such programs before finishing the search**.
**Tentative solution:** If the physicists had a procedure to stop the search upon finding a promising model, or the computer had a way to stop the search upon finding a good enough diamond maximizer, then we'd actually have a chance of finding those before the hacking models. The first might be doable, although selecting for "promising" might be dangerous as above, but we'd be inside the limited space of worlds satisfying the "boundary conditions", and that seems safer. The second seems doable, because any physical setup that achieves more than, let's say, 80% of the diamond-score physically attainable (calculated for instance through volume constraints), is almost surely already a diamond-maximizing AGI (by the argument of **Section A**).
**Problem 6.** *Might not the resulting AGI miss out on unknown Physics facts that would maximize diamond even further (for instance, by packing it more compactly)?*
Yes. In a sense, once we start searching for the AGI to build, our Physics remains forever fixed. Ideally, the physicists will have found all improvements to Physics that allow for more diamonds. Realistically, it's possible that we miss some. In that case, if one of the AGI designs that the search goes over would exploit these unknown facts (and thus produce more diamond), inside our simulation it will just perform badly (trying to exploit some facts that are false in the simulation), and won't get selected[[18]](#fn8f0uay7odul). It's not clear to which extent the AGI resulting from this search, that is, a diamond-maximizing AGI with fixed ontology/laws of Physics, can be considered a truly general intelligence (even if it's possible that its ontology is optimal for diamond maximization). It might seem that to build a diamond-maximizing AGI which auto-updates its ontology we need to solve STEM AI first. But actually, it might be easier to build an agent that does everything it can (including STEM) to achieve a goal, than building one that only does STEM (and we can use as a tool).
**Problem 7.** *Won't the resulting AGI miss out on acausal trade?*
If we use our method of **Section 1**, focusing on Physics and causality, then indeed we have no reason at all to expect our AGI participating in acausal trade, on the contrary, it almost surely won't. That is, unless physicists end up somehow discarding the current understanding of causality, in which case that method doesn't even seem applicable.
If we use the alternative method of **Section 1'**, focusing on information and evidence, our world model might end up accommodating evidentialist views (if these arise naturally/are canonical), and so might search for AGIs that acausally trade.
**Problem 8.** *Aren't you assuming reality can ultimately be perfectly or almost perfectly modeled by some mathematically structured laws?*
Yes.
**Anti-Problem 9.** *Whole Brain Emulation could help*
This is of course common across alignment proposals. In our specific proposal, uploading physicists would allow to compress the whole process of coming up with acceptable laws of Physics and world model(s) into a single run of the infinite computer. And also let it run for as long as necessary so as to ensure high confidence in the result (or additionally have copied physicists independently and locally check through literally every part of the simulation). This can again be more dangerous because the physicists are receiving inputs from an even bigger search space. Some architectures can be implemented to try and minimize this risk. For instance, instead of just having one physicist look at the data, also include a second physicist (or psychologist, or whatever) that looks at the physicist looking at the data, and makes sure nothing weird has happened (like the first physicist getting brain-hacked). This can be iterated, and also have many physicists (instead of just one) looking at each such scenario (HCH-like tree structure). Any such implementation will prevent some failures, but leave other vulnerabilities (or even create some more, although intuitively the bigger structure should be more robust for most inputs).
**C. How does this apply to actual Alignment?**
-----------------------------------------------
Any solution to diamond alignment is already very removed from actual Alignment (and that's how the relaxations encourage fresh takes). But the one presented here is especially untranslatable. It makes **such a central use of an absolutely unattainable amount of computation**, that removing this relaxation leaves the approach completely inapplicable.
The approach is very non-standard in the sense that, instead of discussing the inner workings of an AGI (and forming intuitive, approximate, highly abstract pictures about how these will relate to consequences in the real world), **we directly search over the space of consequences** (and this requires the unattainable amount of computation), and try to find methods that make this search possible and safe.
But this solution proves very useful for another purpose of diamond alignment: pointing at the actual difficult kernels of the problem. Removing all bounds to the search's feasibility makes apparent how **the vastness of the search space itself is the enemy**. This is the enemy of consequentialist thinking of any kind, with or without AGI. But it turns out that in vast and complex enough search spaces AGI occurs very naturally (or so we think), and so many dangers arise through it.
Here's **a framing of the Alignment problem** inspired by that idea:
> Building an AGI is a very chaotic action, in the Chaos Theory sense that small tweaks to it will result in huge differences in the future of the universe. To ensure we don't screw up, we'd like to search through all (or the most important) possible future paths (or action-consequence relations, where the action involves building an AGI). Humans can't efficiently do that, due to fundamental constraints on our architecture and computation power (and because we don't have a method to distinguish the most important paths, because "important" is not some objective feature of the universe, but only defined contingently as "important to us"). But if we build something to do that search for us, or somehow delegate through other mechanisms, the resulting thing or mechanism will have much more searching power than we do, and so if we haven't specified completely correctly (robustly) what it has to search for, it will Goodhart our proxy. That is, the thing or mechanism is itself very chaotic, and we're back at the start. This is a problem, because specifying completely correctly is almost impossible for humans, because the world is very messy and we don't have the search power to explore all consequences of our specification (we don't even have the right Physics).
>
> Between the extremes of "have humans do the search" and "build an AGI to do the search", there are many intermediate solutions. All of these solutions try to satisfy two constraints: being powerful enough so as to efficiently do the search, and not so complex that humans can't specify the objective correctly (because they can't explore the consequences of each specification). This is hard, because being powerful is usually related to being complex. But they are not literally equivalent, so the search space is not literally linear between those two extremes, and some clever tricks surely exist. Ultimately, it's not clear whether solutions satisfying the two constraints exist, or are numerous or natural enough for humans to find them.
>
>
Another meta-level useful feature of the solution here presented is that it **presses on the boundaries of the diamond alignment problem**, stressing how much of the problem is really captured or obscured by which assumptions/relaxations, and to which extent they are reconcilable with fundamental properties of reality. Throughout the text, we find that many details under-determined by the diamond problem's statement are crucial to the viability of some strategies:
* whether the computer can be used once or many times
* whether time is a concern
* whether the infinite computer is a physical system
* even what we're satisfied to call a diamond-maximizing AGI (whether it needs to be able to auto-update its ontology, etc.)
Of course, all of these can just be defined away (even the last one) for the sake of concreteness (although having them under-determined at least helps consider a wider range of strategies). And even doing so in the most optimistic way possible won't assure this solution will work.
But what I'm getting at is that, in these fringe under-determinations, we find expressed many **irreconcilable tensions between the idealized relaxations and the reality we're actually thinking about** when trying to solve the problem. Some pedagogical/theoretical benefits of tackling the diamond problem are obvious, and so I'm not arguing against doing so. But one might raise the question, to which extent does it make sense to, for instance, consider an idealized non-physical computer, when at the same time we're trying to get around the messiness of the rest of reality?[[19]](#fnutu74272pyh)[[20]](#fn1m8pq1392nr)
That is, might considering that nonsensical situation not **encourage a dangerous double-think**, which might later make its way (unnoticed) to our actual opinions about real world Alignment? After all, our real worry is what are the physical consequences of embedded hardware running a certain computation. When dividing up the problem into such neat compartments as the diamond problem does, might our intuitions not later forget what the real problem was about, and point us in non-obviously mistaken (but nonetheless mistaken) directions? That is: does the diamond problem really capture everything that is fundamental to Alignment?
I don't have strong arguments for the diamond problem missing some fundamental core of the issue. After all, it certainly does capture its most obvious aspects. And I know it is consciously obvious to everyone that its relaxations are nonsensical. But I've just come off the other side with the feeling that **we might want to be more attentive** of how these relaxations make our intuitions about reality bend in weird, incoherent ways.
1. **[^](#fnrefqh5x6m5r4ip)**Throughout the text I use macroscopic loosely (for instance, molecules might be included) to mean as fine a granularity as our current paradigm permits, without entering into quantum or other troubles.
2. **[^](#fnrefwvv0ozd8s3i)**Maybe having humans continuously interact with the infinite computer (instead of using it only once) is considered cheating.
3. **[^](#fnrefspusis44j28)**I speak of coordinates, but these of course can't be solely spatial. They should be spatiotemporal to account for relativity, or include whatever further dimensions our preferred Physics requires.
4. **[^](#fnrefq2gf6bb0ric)**Maybe longer to avoid quantum interferences.
5. **[^](#fnrefqvj5ope8ggq)**Of course this again makes our AGI potentially lose out on other weird processes that produce the atomic structure of diamond, and so we might end up with a "process Y on material X" maximizer instead of a diamond maximizer (even if the two usually coincide in the universe).
6. **[^](#fnref532eq7qkvse)**If the simulator was able to deterministically infer the state at time t from the state at time t+1, it might be better (or more informative to Physicists) for the simulation to start with the present, very prohibitive boundary conditions, and make its way back to something like the Big Bang.
7. **[^](#fnref1yrncdum9z6)**It might seem worrisome that I'm invoking physicists so much, since that usually signals a part of the argument which I can't complete. But on this instance, I do think I have a generally good feel for what these physicists would actually be doing, and moderately informed opinions and intuitions as to whether this process would converge, how long it might take, etc.
8. **[^](#fnrefrov4rza6ac)**Although this of course induces some risk of failure in all its predictions, and even without failure our resulting AGI might be missing on some opportunities to exploit these weird phenomena for diamonds.
9. **[^](#fnref7ellvqpuraa)**I'm not sure whether timeline concerns are supposed to apply to the diamond alignment problem. Maybe they aren't usually considered just because most proposals only use the infinite computer once.
10. **[^](#fnrefb5rkikdol07)**Of course the premise of having an unboundedly fast computer with unbounded memory be a bounded physical system is already nonsensical, but straight up computing the uncomputable (this infinite nested regress) seems categorically even worse. This is also the reason why, when using infinite compute to find the world models which fit the facts (in **Section 1**), these facts must all lie in the past, and the simulation must not arrive at the moment in time where the computation begins.
11. **[^](#fnrefp1g9dvzqnlq)**This approach doesn't run into the uncomputable infinite nested regress because the coarse model wouldn't be precise enough to model the computer's computation exactly.
12. **[^](#fnrefh443clibn)**This paragraph is informative, but of course, strictly speaking, who cares if the resulting system is not an AGI? We only care about diamonds.
13. **[^](#fnrefbfqyh7uoueh)**This resonates with John's Natural Abstraction Hypothesis.
14. **[^](#fnrefviiu83qu3p)**Since the smallness of the box is what protects us against overly specific setups prone to fail under the smallest misadjustment, one might wonder whether we truly want the box to be "as big as we can make it". But I think we do, because increasing its size drastically augments the search space for our AGI, and almost negligibly augments the probability that we find an overly specific setup (those probably have size much greater than Earth).
15. **[^](#fnrefif7finyl8tg)**It is conceivable (or even likely) that, if we really did know our world model with arbitrary accuracy, then some (at least partially) deterministic setup creates more diamonds than a general intelligence (because of contingent quirks of our universe). But I guess both achievements (either building a diamond maximizing AGI or somehow maximizing diamond even harder) are sufficient to pass this problem. After all, even an AGI is not omnipotent and will inevitably leave some utility on the table.
16. **[^](#fnrefs26etpk7d1)**And maybe I am allowed to do that for the diamond alignment problem.
17. **[^](#fnrefry5itq0kori)**And I guess the whole point of the diamond alignment problem is to trivialize proxy concerns away.
18. **[^](#fnref8f0uay7odul)**We can also understand this as a failure of our diamond-score as a proxy, now caused by the incomplete framework in which it is formulated.
19. **[^](#fnrefutu74272pyh)**Of course, the diamond problem could specify that the infinite computer is physical. But is that, in some relevant sense, less nonsensical than a non-physical computer?
20. **[^](#fnref1m8pq1392nr)**My non-standard solution suffered more than most solutions when dropping, for instance, the relaxation of the computer being idealized (see **Problem 6**), and that might be why I'm giving so much weight to this issue. But the next paragraph tries to get at how this mismatch, in a way less obvious manner, could also happen in more standard solutions (that don't even press that hard on the problem's boundaries).
|
99995c3a-3fdf-411c-8c34-f4acee3cfa56
|
StampyAI/alignment-research-dataset/eaforum
|
Effective Altruism Forum
|
Pause For Thought: The AI Pause Debate (Astral Codex Ten)
An overview of the EA Forum's recent AI Pause Debate week, from blogger Scott Alexander at Astral Codex Ten.
As I'd been meaning to get around to reading the debate, this was a helpful way in, as it scopes out where the main points of agreement and disagreement lie — and even has a go at identifying cruxes.
|
b7e53ee5-a9a8-43c2-9b87-e6012873c6c6
|
trentmkelly/LessWrong-43k
|
LessWrong
|
The Scylla of Error and the Charybdis of Paralysis
We're interested in improving human rationality. Many of our techniques for improving human rationality take time. In real-time situations, you can lose by making the wrong decision, or by making the "right" decision too slowly. Most of us do not have inflexible-schedule, high-stakes decisions to make, though. How often does real-time decision making really come up?
Suppose you are making a fairly long-ranged decision. Call this decision 1. While analyzing decision 1, you come to a natural pause. At this pause you need to decide whether to analyze further, or to act on your best-so-far analysis. Call this decision 2. Note that decision 2 is made under tighter time pressure than decision 1. This scenario argues that decision-making is recursive, and so if there are any time bounds, then many decisions will need to be made at very tight time bounds.
A second, "covert" goal of this post is to provide a definitely-not-paradoxical problem for people to practice their Bayseian reasoning on. Here is a concrete model of real-time decisionmaking, motivated by medical-drama television shows, where the team diagnoses and treats a patient over the course of each episode. Diagnosing and treating a patient who is dying of an unknown disease is a colorful example of real-time decisionmaking.
To play this game, you need a coin, two six-sided dice, a deck of cards, and a helper to manipulate these objects. The manipulator sets up the game by flipping a coin. If heads (tails) the patient is suffering from an exotic fungus (allergy). Then the manipulator prepares a deck by removing all of the clubs (diamonds) so that the deck is a red-biased (black-biased) random-color generator. Finally, the manipulator determines the patients starting health by rolling the dice and summing them. All of this is done secretly.
Play proceeds in turns. At the beginning of each turn, the manipulator flips a coin to determine whether test results are available. If test results are available, the manip
|
53adced3-fb40-4a41-ab38-b8d07b7e0da1
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Eponymous Laws Part 2: Laws of Programming and Software Development
(See Eponymous Laws Part 1: Laws of the Internet)
Though ostensibly about programming and software development, most of these laws refer to fundamental aspects of human psychology (and folly). Enjoy.
Flon’s Law – “There does not now, nor will there ever, exist a programming language in which it is the least bit hard to write bad programs.”
Postel’s Law (aka Robustness Principle) – “Be conservative in what you send, be liberal in what you accept”… “In other words, programs that send messages to other machines (or to other programs on the same machine) should conform completely to the specifications, but programs that receive messages should accept non-conformant input as long as the meaning is clear.”
Bradley’s Bromide – “If computers get too powerful, we can organize them into a committee — that will do them in.”
Weinberg’s Law – “If builders built buildings the way programmers wrote programs, the first woodpecker that came along would destroy civilization.”
Hartree’s Law – “Whatever the state of a project, the time a project-leader will estimate for completion is constant.”
Hoare’s Law of Large Programs – “Inside every large program is a small program struggling to get out.”
Jakob’s Law of the Internet Use Experience – “Users spend most of their time on other sites. This means that users prefer your site to work the same way as all the other sites they already know.”
Ninety-Ninety Rule – "The first 90% of the code takes the first 90% of the time. The remaining 10% of the code takes the remaining 90% of the time."
Knuth’s optimization principle – “Premature optimization is the root of all evil.”
Linus’ Law – “Given enough eyeballs, all bugs are shallow.”
Kerchkhoff’s Principle – “In cryptography, a system should be secure even if everything about the system, except for a small piece of information – the key – is public knowledge.”
Wirth’s law – “Software gets slower faster than hardware gets faster.” (as a counterpoint to Moore’s Law)
Brooks’ Law – “A
|
2537cec7-d63f-4b7f-8332-92a517083d7e
|
StampyAI/alignment-research-dataset/alignmentforum
|
Alignment Forum
|
Logical Counterfactuals Consistent Under Self-Modification
.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')}
I argue that the correct notion of logical counterfactual is the one which an agent would self-modify to have if the agent lacked it, and that this significantly constrains the possibilities. The argument is rather informal, and I'm not really sure if many of the points are new, so I'm posting it in Discussion.
---
Where the idea is coming from
=============================
(This section is an extended motivation, and may be skipped.)
At the most recent MIRIxLA, Alex Mennen led a discussion attempting to apply conservation of expected evidence to the Demski prior. The basic idea was that approximations of the Demski prior discussed so far can do quite unreasonable things before converging to good probabilities, even those that are [uniformly coherent](https://agentfoundations.org/item?id=431). Can we describe a property for approximations such that the probabilities given at finite times are in some sense reasonable? Based on the idea of conservation of expected evidence, is it possible for an approximation to give finite probabilities such that (in some sense) the probability that the estimate will drift up later is balanced by the probability it will drift down?
We didn't come up with any formulations of this idea which seemed both possible and nontrivial, and we came away with the feeling that something like what we were after would not be possible in the end. The version which we spent the longest time discussing was this one:
(1) ∀t,a≤b:Eϕ(P(ϕ)|Pt(ϕ)∈[a,b])∈[a,b]
Pt() represents the probability estimate at time t, while P() is the probability reached in the limit of the approximation process. The expectation is being taken over ϕ; we suppose a probability distribution μ(ϕ) which we're sampling sentences from. At any time t, sampling only ϕ whose probability *estimate* lands in a range [a,b], we want the expected value of the *eventual* probabilities to be in that same range. (Although inspired by conservation of expected evidence, this might also be seen as a notion of calibration.)
Although we convinced ourselves that this was very likely impossible, we didn't come up with a proof. None of our variations to the definition seemed very promising either.
Somewhat facetiously, I offered the following resolution to our difficulties. One thing which seems artificial about (1) is the expectation over ϕ based on μ. What this accomplishes is to inject a kind of made-up uncertainty into the logically certain world; if we turned a blind eye to the exact sentence ϕ and only considered the current probability estimate, would it be too high or too low on average? We're acting as if we forgot which sentence we're dealing with.
What we *really* want is an expectation in our *real* uncertainty, namely, our logical uncertainty. Therefore, let's explicitly model our uncertainty about the consistency of theories. We can pretend that the consistency check is a stochastic process which has some chance of eventually declaring sets of sentences inconsistent. Approximate the Demski prior by a sampling procedure as usual, except incorporating these random declarations of inconsistency as well as the usual, so that sentences have some chance of being thrown out. Pt(ϕ) runs the real proof checker for time t, but additionally declares some fraction of sentence-sets inconsistent at random; this fraction decreases with t so that we still accept consistent sentences in the limit. This turns Pt(ϕ) into a computable distribution.
Now, with respect to this distribution, conservation of expected evidence and other typical properties will hold. This doesn't *really* get us anything nice, though -- it's a bit like saying you're well-calibrated when things are sampled according to your own prior. The consistency check isn't anywhere near random, so although our belief structure follows conservation of expected evidence with respect to the approximation procedure (equally expecting probabilities to shift up or down over time), the *actual* average movement based on the *real* consistency checking may well be much more predictable in practice.
The approach is quite odd, in that I'm suggesting that the way to get a good approximate probability distribution over logic is to start with a good approximate probability distribution over logic. (Consistency facts are, after all, just logical facts.)
I first formulated this "cheater's solution" when I was first thinking about the [two update problem](https://agentfoundations.org/item?id=427). It seemed as if the two-update problem was pointing at different updates for observing X vs observing a proof of X. Rather than keeping possibilities until they're shown to be inconsistent and then throwing them out, we maintain some uncertainty about this. This turns the "second update" from a non-Bayesian surgery on beliefs into a Bayesian update in an expanded belief structure, thanks to the explicit model of consistency.
Again, this doesn't seem to *really* solve anything for the two-update problem: by modeling the consistency check as a probabilistic variable, I "bayesianize" the second update; it now looks like we're conditioning on information in a proper manner. But, what is really gained? The probabilistic model of consistency checks is bound to be poor, so the probabilities coming out of the approximation won't be very meaningful. It seemed more interesting to look at the proposed solutions (such as backup machines) which might lead to a more powerful prior, rather than the same prior approximated another way. And soon, we moved on from the two-update problem to more fruitful things.
One thought which I entertained *very* briefly was to get logical counterfactuals out of an approach like this, by running a version of updateless decision theory using a prior which is logically ignorant, by modeling consistency checks like this. I now think this idea may be a good one, as I'll argue in the remainder of the post.
Motivation
==========
[The original TDT paper](https://intelligence.org/files/TDT.pdf) argued that CDT agents who considered decision-theoretic problems such as Newcomb's problem, and who had the option to self-modify to TDT, would do so.
Roughly speaking, there are two separate claims being made by TDT/UDT: (A) decisions should be made in an timeless/updateless manner, and (B) different actions should be considered according to logical counterfactuals. (In TDT, this means a causal diagram of mathematics; in UDT, this was initially called a "mathematical intuition model" to make sure it was obvious that it was a blank to be filled in.)
It's straightforward to see why an agent given the opportunity to self-modify to conform to point (A) would do so. The very idea of (A) is to choose the actions you *would have wanted yourself to choose* according to your prior. The argument for (B) is, I think, less obvious. The point of the current post is to argue that a CDT agent or other sufficiently reasonable agent will in fact self-modify to conform to (B) as well; furthermore, this tells us something about what makes a good or bad logical counterfactual.
As in that paper, I'll ignore details of self-modification such as the Lobian obstacle and reason informally, under the assumption that the agent is capable of making reasonable predictions about the results of self-modification.
Betting on logical coins
========================
The idea came out of a conversation with Scott about betting on logical coin-flips. Suppose that Sandy and Omega are making bets on digits of pi. The digits are far enough out in the sequence that even Omega cannot predict them at better than chance, and Omega can predict Sandy perfectly before running a long computation to determine the outcome. Omega gives Sandy favorable odds to make it worth Sandy's time. However, Omega lacks an enforcement mechanism for the bets, and must trust Sandy to pay out. If Omega predicts that Sandy won't pay for losses, Omega will not offer to make a bet with Sandy. (Notice this means Omega possesses a notion of counterfactual.) This provides an incentive to be a trustworthy betting partner.
Suppose that Sandy accepts arguments for (A), and so computes actions in an updateless manner, but does not accept (B). (Expected utility is evaluated by conditioning on actions as in EDT, or causal conditioning as in CDT, or perhaps something else, but not logical conditioning.) Sandy is a trustworthy betting partner for non-logical (that is, empirical) bets; if the optimal strategy was to commit to the bet prior to seeing the outcome, the optimal strategy will be the same after seeing the outcome, so Sandy pays up without any commitment mechanism. For a logical coin, however, Sandy might falter: discovering new logical information allows Sandy to compute a better approximation of the updateless decision. If the bet is lost, Sandy may now logically rule out the case where the bet was won. (This is especially tempting if Omega provides a proof along with the correct answer.) If so, Sandy will now compute a negative utility for paying out, and refuse to do so. Hence, Omega never offers Sandy a bet.
The conclusion is that Sandy, realizing such problems, will accept (B) as well as (A).
I can see two objections to this argument. The first is that I haven't specified Sandy's decision algorithm at all well, so it's not really clear that she fails to be a good betting partner. The second is that Omega already has a notion of counterfactual, and might be essentially forcing it on the agent; perhaps we can get Sandy to accept any notion of counterfactual we choose, depending on the notion Omega employs. I think both objections can be addressed with a more general argument.
Time-consistent logical gambles
===============================
Sandy's mistake, on my analysis, was to modify the approximation of the updateless decision. Although the updateless expected utility must be approximated (in logically uncertain situations), any *change* to this approximation can lead to temporal inconsistency of decisions, and therefore, reflective inconsistency (Sandy wishes to make decisions in a more temporally consistent way). So, for the same reason that a self-modifying agent will tend to accept (A) and switch to an updateless strategy to improve future expected utility, it will also wish to freeze its approximation Pt(ϕ) at the current value of t.
Yet, we would not want to fix one approximation at the beginning and run with that; we *need* to improve our approximation of the logical prior over time, if only to reach probability 1 for tautologies and probability 0 for contradictions. What do we do?
Just as UDT refuses to update on observations, but instead expresses the policy as a function from observations to actions, we can refuse to update on our consistency check, but express the policy as a function from consistency information *and* observations. Consistency bits are treated like observations!
More specifically, my proposal is this: as foreshadowed in the opening section, approximate the Demski prior with a sort of "pre-Demski" prior as follows. Build a distribution over random theories *and* the consistency checks, by sampling sentences as usual for the Demski prior, but modeling consistency checks as a random process check(Γ,t) which has a probability of outputting "inconsistent" at each time t for a given set of sentences Γ. For example, the probability could start at .25 and half with each increment of time, so that the total probability that any set of sentences is inconsistent is .5. (It may be desirable to make this more realistic, for example ensuring that Γ∪{ϕ} is more likely to be inconsistent than Γ alone, but it will always be a very rough approximation so it's unclear how much we ought to do.)
Policies are specified as a function from a sequence of observations and consistency checker behaviors to actions; the agent implementation attempts to take actions from the max-utility policy, as computed via this "pre-prior". (A difficulty: there may not be a unique maximum. I'm not sure what to do about that at the moment.)
This approach optimizes strategies over all logical (im)possibilities, putting impossible possible worlds on equal footing with possible possible worlds.
It may be tempting to think this will yield terrible results analogous to using the approximation at time 1, and therefore we need to pack as much information into the initial approximation as we can. However, the result will usually be close to what we would get by the normal process of improving our approximation over time. Just as [under certain conditions an updateless theory will give the same answer as a thing that updates on evidence](http://lesswrong.com/lw/1fu/why_and_why_not_bayesian_updating/), my suggested procedure would often generate the same answer as something which updated on consistency information.
What remains
============
I don't expect writing down the formulas and algorithms of my proposal will be particularly hard, but I haven't done it here. I think there may be some regularities we need to impose on the stochastic consistency model, but I'm not sure. I don't expect to be able to formalize reflective consistency itself without also utilizing a solution to the probabilistic Tiling agents problem, or something close to that; in other words, without solutions to other significant problems. However, a formal description of a class of problems for which the decision procedure is optimal might become possible.
It would be awesome to produce a "logical Dutch book argument" to illustrate problems which arise if logical uncertainty isn't based on a probability distribution over possible impossible worlds, if that idea makes sense.
It might be that we can do *significantly* better with a more interesting prior on the consistency checks; the best we can hope for is something which learns regularities in logical space. Keep in mind that this "learning" won't be changing the prior we use to evaluate expected utilities; it only modifies the expectations down the branch of possible worlds which our agent follows as it discovers logical facts. (It might also be that this ends up not mattering.)
An interesting idea in connection with reflective consistency is that the beliefs on consistency might obey some kind of reflection principle relating them to beliefs about consistency expressed in first-order logic. It stands to reason that an agent considering its own choice of distribution over consistency-check behavior would prefer this distribution to match its logical uncertainty about consistency.
|
257ff6a2-7617-47f4-998b-19886f67187a
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Paper
Paper is good. Somehow, a blank page and a pen makes the universe open up before you. Why paper has this unique power is a mystery to me, but I think we should all stop trying to resist this reality and just accept it.
Also, the world needs way more mundane blogging.
So let me offer a few observations about paper. These all seem quite obvious. But it took me years to find them, and they’ve led me to a non-traditional lifestyle, paper-wise.
Observation 1: The primary value of paper is to facilitate thinking.
For a huge percentage of tasks that involve thinking, getting some paper and writing / drawing / scribbling on it makes the task easier. I think most people agree with that. So why don’t we act on it? If paper came as a pill, everyone would take it. Paper, somehow, is underrated.
But note, paper isn’t that great as a store of information. You can’t search, cross-references are iffy, and it’s hard to copy or modify. Nobody I know really looks at their old paper notes very often. So don’t optimize for storage. Optimize for thinking.
Observation 2: If you don’t have a “system”, you won’t get much benefit from paper.
Say you want to do some thinking with paper right now. How would you do it? If you have no system in place, you’ve got some problems: What paper should you write on? Where does it go when you’re done? These are small problems, but they add friction. If you have to solve them, maybe you won’t bother using paper.
So solve them. Your “system” could be, “Write on a notepad and throw the pages out at the end of the week.” Fine! At least you’re using paper now.
Observation 3: User experience matters.
Some pens and paper spark more joy than others. Use them. This is not frivolous. When more joy is sparked when you scribble, better thinking follows.
There are many other dimensions of user experience. Personally, I find paper with lines to be crushing and dehumanizing. But I recognize this is not a human universal.
Or, say you decide to write in a not
|
9524da98-8cd6-4377-bee4-c8e03a673b36
|
trentmkelly/LessWrong-43k
|
LessWrong
|
A possible training procedure for human-imitators
Some proposed ways to train a system to imitate a human imitate a human involve having one system imitate a human, while another system tries to tell the first system apart from an actual human. If the first system can get classified as human by the second system, then (one might think) it is imitating a human well, as long as the second system is sufficiently smart.
I described a possible problem with these approaches in a LessWrong thread:
> Here's a scenario that doesn't seem completely implausible. Suppose Bob is someone whose public key is easily available on the internet. The first AI will read things on the internet and output a message. Some of the message will get put on the public internet. Bob suspects that the AI might have secretly sent him a message (e.g. giving him advice on which stocks to buy). So he tries using his private key to decrypt some of the AI's output (e.g. the lowest-order bits in some images the AI has output).
> Knowing that Bob (or someone else like Bob) will likely do this, the AI (which has access to Bob's public key, due to reading the internet) encrypts a message with Bob's public key, and puts the ciphertext in the lowest-order bits of some image. The ciphertext looks like random noise to the second AI. Maybe the plaintext contains both actual stock tips and some dangerous messages that trick Bob into running an unrestrained copy of the first AI (I'm assuming the AI is good at psychological manipulation).
As Paul says later in the thread, the underlying problem is that it is easy for the first AI to change its output in an important way without the second AI noticing.
Here's a proposed way of implementing human-imitation while avoiding this problem. The human-imitator is a single AI that is given a prefix of a string produced by a human and predicts the next bit. It is rewarded for predicting the next bit accurately (using some proper scoring rule). We can use this system to imitate a human by sampling from its implied proba
|
3bd1cdbe-d70b-40f4-b55a-3e36f255e498
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Categories of Sacredness
Previously: Eternal, and Hearthstone Economy versus Magic Economy, Out to Get You
On Lesser Wrong, JenniferRM gave a reply that is worth quoting in full:
> If I understand correctly, the cognitive process/bias/heuristic/whatever of “sacredness” is relevant here.
> Neither nails nor dollars are sacred so you’re free to trade dollars for nails.
> A kidney is sacred, so you can’t trade that for dollars, but you can trade it for another kidney (although such trades still feel a bit weird).
> Sacred things are often poorly managed in practice, and sacredness is easy to make fun of, but a decent defense of sacredness might be that it is one of the few widely installed psychological mechanisms in real life for managing the downsides of having markets in things. Thus, properly deployed sacredness might let you have “trade” in one area without ending up with “totalalizing trade”?
> In the smaller and hopefully lower stakes world of video games, I think the suggestion would be to have card classes with different trading characteristics.
> The lowest class of very non-sacred things could be swapped with extremely low transaction costs within the class and also be tradeable directly for money.
> Higher sacredness things would have a separate market, perhaps with transaction costs like needing a purchaseable delivery mechanism or imposing delays so that objects go into limbo after the trade is finalized while “being delivered”. The most sacred things would be “inalienable” so they can’t be traded or given away or perhaps not even be destroyed.
> Exactly where sacredness should be deployed in order to maximize fun seems like a deep and relatively unstudied problem.
> One place in real life where the inalienability of something has large and substantive differences from jurisdiction to jurisdiction is the question of the rights of artistic creators to their artwork. In some jurisdictions, an artist cannot legally sell their right to veto the use of their artwork if deplo
|
f1d1dafd-4f77-4209-8308-dd045235caac
|
trentmkelly/LessWrong-43k
|
LessWrong
|
[Link] Why the kids don’t know no algebra
Post by fellow LW reader Razib Khan, who many here probably know from the gnxp site or perhaps from his debate with Eliezer.
> A few days ago I stumbled upon a really interesting post. And I’m wondering if my readers are at all familiar with the phenomenon outlined here (it was a total surprise to me), The myth of “they weren’t ever taught….”:
>
> With all this I am not saying conditions which are non-hereditary are irrelevant. What I am saying is that we can’t ignore the shape of the pre-existent landscape before we attempt to reshape it to our own image. Excoriating teachers for having pupils who can’t master mid-level secondary school mathematics is in some cases like excoriating someone for the fact that their irrigation canals from the plains into the mountains are failures. You need to level the mountains before your canals can work (or, barring that design and implement a mechanical system which will move water against the grade). Easier said than done. E. O. Wilson said of Communism, “Great Idea, Wrong Species.” The reaction of Communist regimes to this reality was brutal and shocking. Obviously the modern rejection of unpalatable aspects of human nature are not so grotesque. But they have a human toll nonetheless. I’m skeptical that this generation will pass before we have to acknowledge these realities and calibrate our policies accordingly.
>
> Stage One: I will describe this stage for algebra I teachers, but plug in reading, geometry, writing, science, any subject you choose, with the relevant details. This stage begins when teachers realize that easily half the class adds the numerators and denominators when adding fractions, doesn’t see the difference between 3-5 and 5-3, counts on fingers to add 8 and 6, and looks blank when asked what 7 times 3 is.
>
> Ah, they think. The kids weren’t ever taught fractions and basic math facts! What the hell are these other teachers doing, then, taking a salary for showing the kids movies and playing Math Bingo?
|
d60ec78c-6851-4854-bfcf-acfb194bc24d
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Efficient Cross-Domain Optimization
Previously in series: Measuring Optimization Power
Is Deep Blue "intelligent"? It was powerful enough at optimizing chess boards to defeat Kasparov, perhaps the most skilled chess player humanity has ever fielded.
A bee builds hives, and a beaver builds dams; but a bee doesn't build dams and a beaver doesn't build hives. A human, watching, thinks, "Oh, I see how to do it" and goes on to build a dam using a honeycomb structure for extra strength.
Deep Blue, like the bee and the beaver, never ventured outside the narrow domain that it itself was optimized over.
There are no-free-lunch theorems showing that you can't have a truly general intelligence that optimizes in all possible universes (the vast majority of which are maximum-entropy heat baths). And even practically speaking, human beings are better at throwing spears than, say, writing computer programs.
But humans are much more cross-domain than bees, beavers, or Deep Blue. We might even conceivably be able to comprehend the halting behavior of every Turing machine up to 10 states, though I doubt it goes much higher than that.
Every mind operates in some domain, but the domain that humans operate in isn't "the savanna" but something more like "not too complicated processes in low-entropy lawful universes". We learn whole new domains by observation, in the same way that a beaver might learn to chew a different kind of wood. If I could write out your prior, I could describe more exactly the universes in which you operate.
Is evolution intelligent? It operates across domains - not quite as well as humans do, but with the same general ability to do consequentialist optimization on causal sequences that wend through widely different domains. It built the bee. It built the beaver.
Whatever begins with genes, and impacts inclusive genetic fitness, through any chain of cause and effect in any domain, is subject to evolutionary optimization. That much is true.
But evolution only achieves this by runni
|
df20da52-9d9a-436c-947b-a61a717be389
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Predictive Processing, Heterosexuality and Delusions of Grandeur
Predictive processing is the theory that biological neurons minimize free energy. In this context, free energy isn't physical energy like the energy in your laptop battery. Instead, free energy is an informatic concept. It is useful to think about free energy as a single variable balancing prediction error and model complexity. Minimizing free energy balances minimizing prediction error against minimizing model complexity. You can also think about it as minimizing surprise.
Biological neurons have inputs and outputs. Each neuron receives inputs from one set of neurons and sends outputs to another set of neurons. One way to minimize prediction error is to fire right after receiving inputs, but an even better way to minimize prediction error is for a neuron to anticipate when its input neurons will fire and then fire along with them. Firing in-sync with its inputs produces zero prediction error instead of just a small prediction error.
It has been shown that predictive processing is asymptotically equivalent to backpropagation. Everything that can be computed with backpropagation can be computed via predictive processing and vice versa.
A Simple Clock
Suppose you take a 3-dimensional blob of neurons programmed to minimize local prediction error and you attach them to a sinusoidal wave generator. The neural net has no outputs—just this single input. At first the neurons close to the wave generator will lag behind the sinusoidal wave generator. But eventually they'll and sync up with it. Our neural net will have produced an internal representation of the world.
Now suppose you plug a sinusoidal wave generator into the left end of the neural net and a square wave generator into the right end of the neural net. The left end of the neural net will wave in time with the sinusoidal wave generator and the right end of the neural net will beat in time with the square wave generator. The middle of the net will find some smooth balance between the two. Plug more complicated
|
daf2aa47-0fb7-42a1-ad8b-8931d628882e
|
trentmkelly/LessWrong-43k
|
LessWrong
|
[New Feature] Support for Footnotes!
It is with great excitement[1] that I am pleased to announce that the main LessWrong text editor[2] now has support for footnotes![3] A huge thanks to our friends over at the Effective Altruism Forum who coded this one up.
You can insert footnotes via:
1. Manually selecting text in the text box and selecting insert footnote from the footnotes menu icon.
The footnote icon is the [*] on the right.
2. Using Markdown syntax
* Type [^n] where is the number of the footnote you wish to insert.
* To insert a new footnote, use n that is <number of existing footnotes + 1>; to reuse an existing footnote, set n to be whichever footnote you are reusing.
Footnotes will automatically renumber as you add and delete them!
What's more, footnotes will render with hover-over previews once published:
Behold! A footnote hover-preview!
That's it. Go forth and create scholarly works!
1. ^
I mean it, really. I've looked forward to us adding this support for years.
2. ^
That is, the LW Docs editor, as distinct from the Markdown editor and legacy Draft-JS editor.[4]
3. ^
The Markdown editor already had support for footnotes using Markdown footnote syntax.
4. ^
Yes, footnotes can have footnotes. And those footnotes can reference themselves.[4]
|
6d2dd136-ac0d-42a8-9b60-e65206b50698
|
trentmkelly/LessWrong-43k
|
LessWrong
|
What big goals do we have?
Sometime ago Jonii wrote:
> I mean, paperclip maximizer is seriously ready to do anything to maximize paperclips. It really takes the paperclips seriously.
When I'm hungry I eat, but then I don't go on eating some more just to maximize a function. Eating isn't something I want a lot of. Likewise I don't want a ton of survival, just a bounded amount every day. Let's define a goal as big if you don't get full: every increment of effort/achievement is valuable, like paperclips to Clippy. Now do we have any big goals? Which ones?
Save the world. A great goal if you see a possible angle of attack, which I don't. The SIAI folks are more optimistic, but if they see a chink in the wall, they're yet to reveal it.
Help those who suffer. Morally upright but tricky to execute: James Shikwati, Dambisa Moyo and Kevin Myers show that even something as clear-cut as aid to Africa can be viewed as immoral. Still a good goal for anyone, though.
Procreate. This sounds fun! Fortunately, the same source that gave us this goal also gave us the means to achieve it, and intelligence is not among them. :-) And honestly, what sense in making 20 kids just to play the good-soldier routine for your genes? There's no unique "you gene" anyway, in several generations your descendants will be like everyone else's. Yeah, kids are fun, I'd like two or three.
Follow your muse. Music, comedy, videogame design, whatever. No limit to achievement! A lot of this is about signaling: would you still bother if all your successes were attributed to someone else's genetic talent? But even apart from the signaling angle, there's still the worrying feeling that entertainment is ultimately useless, like humanity-scale wireheading, not an actual goal for us to reach.
Accumulate power, money or experiences. What for? I never understood that.
Advance science. As Erik Naggum put it:
> The purpose of human existence is to learn and to understand as much as we can of what came before us, so we can further the su
|
49011ab0-2c27-4009-9996-8050efe6bcac
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Ancient Social Patterns: Comitatus
Purpose: I've been thinking about problems of goal advancement in spite of death, which brings us to honor. I am also keen on the history of Central Asia and the Steppe, and this is really cool: hence the blog section.
The word comitatus is Latin for retinue, or armed escort. In Germania, Tacitus uses it to describe the relationship between a lord and his sworn warrior. Through textual and archaeological evidence, this relationship has been shown to be widespread, and extended back much further to Proto-Indo-European times. It has been observed as far west as Spain, and as far east as Korea and Japan. I have most of this information from the book Empires of the Silk Road: A History of Central Eurasia from the Bronze Age to the Present, by Christopher Beckwith.
The comitatus is a lord and a warband of his friends, who swear to defend him to the death, forsaking all other obligations. In life they are bodyguards and constant companions, asking for and receiving gifts of wealth and weapons. If the lord dies first then his comitatus would be interred with him after ritual suicide or execution, and they were buried armed for battle in order that they could fight in the afterlife. Such was honor in those days.
The detail about this arrangement that seems most important to me is that the oath superseded obligations to their kin-group or people. Near as I can tell it is the first social institution to do this: while we have evidence of urbanization and religious organization which is earlier, they are not separated from kin-group concerns. It proved extremely resilient, lasting as late as early modern times in the Scandinavian and Islamic contexts (though I note that the pattern had been adapted - for example ritual execution and suicide fell into disuse with the rise of current world religions).
* Seems to have arisen concurrent with chariots; apart from methods of making tools, may be the earliest military innovation.
* I would expect success in competing against cl
|
993f186b-7b13-4c74-b885-efc37f2303be
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Arch-anarchy and The Fable of the Dragon-Tyrant
In "arch-anarchy", already republished here, the anonymous author "A" presents some good insights into why the laws of nature should not be seen as immutable decrees that govern the universe and much less worthy of veneration. However, I would like to make my arguments. I would say that our relationship with the laws of nature is a form of Stockholm syndrome, our entire lives we have been trapped by them and most of us cannot imagine that we can ever be free from them, because of this.Until we reach a point of idealizing them as something sacred. I will give two examples to better explain what I mean.
First, the state under the anarchist vision, we spend our entire lives living under the rules of the state, hearing at school or in state propaganda about how wonderful and necessary a state is, that the majority of the population really thinks it is impossible to live without politicians and bureaucrats regulating our lives. The second is the death over death view of the life extension movement. We spend our entire lives being told that death is a natural and inevitable part of life. Much of the population actually thinks that death is actually good, and that a world without death would be horrible.
A work of fiction that perfectly illustrates the errors in this view is "The Fable of the Dragon-Tyrant," a short story by Nick Bostrom. The plot is about a dragon who personifies death, aging, and disease who tyrannizes a kingdom through human sacrifice and imposes Stockholm Syndrome on the people for centuries before being destroyed by a new invention. Although the focus of the story is a critique of complacency about the inevitability of death (something that as an arch-anarchist I definitely agree with), it is easy for readers to interpret it as a critique against complacency about some other accepted part of life: cancer, society, etc. Although it is logically not the author's intention, I can also interpret the story as a critique of complacency against statist law
|
3fa5dfdd-fb1e-4cd0-8205-fe040dc1b77e
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Chakras & Qi - Old Stories for the Base-Line Experience. Improve your physical & mental health by connecting body and mind.
Epistemic status: Based my experiences, a simple explanation for some very old ideas.
The connection between body and mind. Conscious proprioception, using the main muscles of movement, , seeing the sparkles, feeling the power of the human body when it is used correctly.... qi, chakras...
Chakras.
"Chakra" is probably the most well-known terminology in English for a concept that appears in many traditions. I remain vague about 'many traditions' (my knowledge is insufficient to comment) and I include no definition for "chakra" but the existence of chakras is a topic that appears to split people (who have an opinion on the topic and inhabit the online world) into two camps - those who talk of chakras as if they are real phenomenon and those that say it all sounds like nonsense. Is that a fair assessment?
My First Thoughts.
When I started working with my Base-Line - focusing on activating my pelvic floor and rectus abdominis muscles, section by section from pubic symphysis to chest - I found myself thinking 'red, orange, yellow, green ...' as I engaged these muscles and the concept of chakras came to mind.
Research.
I would classify my starting 'knowledge' as almost zero. I've seen the typical posters (go-ogle images) but I've never been to a yoga class and don't even have another example of where I might encounter chakras. (Nov. 2021 I've now been to a yoga class - no mention of chakras!)
For illustrative purposes only!
I went looking for the original source of chakras (internet trawling).
Reading the blurbs from a couple of 'classic' chakra books instinct/rational thought said to me this is not the right path/seems like a load of BS and most information that appears via go-ogle is an echo-chamber - energy centres, meditation, blockages, symbols, colours, petals ... It gets flaky fast.
I did however deem a couple of articles bookmark-worthy at the time:
* hinduism/concepts/chakras. A couple of lines that stood out:
> chakra or cakra has multiple meaning
|
3ed324b9-db59-40c6-b906-16e0c2af4abc
|
trentmkelly/LessWrong-43k
|
LessWrong
|
What "The Message" Was For Me
> Warning: If you have low tolerance for analogies or poetic language, skip to the numbered chain of dependencies for a more straightforward rundown of the reasoning presented herein.
It's been said that when you get the message, hang up the phone. What is "the message" though? I can't dictate what it is for everybody. This was just my biggest "aha moment" in recent years.
The most commonly expressed version of this idea I have seen, by analogy, is that living beings are like the countless little eyes which make up a compound eye. Each seeing a narrow, limited perspective of the world. The "big picture" which only the overall eye can see is all of those combined.
You could also express it as similar to the way in which a picture on a monitor is made up of many little differently colored pixels. There doesn't appear to be any rhyme or reason to it until you zoom out far enough.
In the same way, a human being viewed up close is actually trillions(!) of tiny individual organisms, none of which know they are part of a much larger creature. All expressions of the same basic idea that we're all part of the same thing, connected in ways that are not immediately obvious until viewed on a grand scale, and that separation is to some extent an illusion.
I experienced the same apparent self-evident nature of this idea but my brain is wired such that it is unsatisfied with that sort of just-so answer. I want complete explanations, diagrams with every little part clearly labeled. I want to know how it works. So I kept chugging away at it until this occurred to me:
1. In biological evolution, simple chemical self-replicators gave rise to intelligent organisms such as ourselves. We know this to be true with quite a lot of certainty.
2. Humans will eventually develop self-replicating machines. This is a plausible assumption given present technological trends toward automation, and probably occurs on any planet where intelligent life evolves. (Providing they don't destroy th
|
1cafc587-1397-4223-bc2e-27fdab6545eb
|
trentmkelly/LessWrong-43k
|
LessWrong
|
MLP: The Next Level Of Your Studies
The first four chapters of my MLP fanfiction are now online on fimfiction.net. Unlike Friendship Is Optimal (fim link), which focuses on how MLP might impact the trajectory of AI, or Myou've Gotta be Kidding Me (fim link), which focuses on how a rationalist might impact the trajectory of Equestria, I wanted to ask: what would the MLP Way look like? How would MLP impact rationality, and what would rationalist MLP look like?
This has been over a year in the making, off and on, and I've received significant help in writing it. In particular, I'd like to thank the pre-readers and editors who have helped polish it, and everyone who's shown interest; that has been a great help in motivating me to work on this rather than other projects.
There's much more to come; I should be able to maintain at least a chapter a week for the medium term, and hope to post more frequently than that. If you're interested in pre-reading / editing chapters to come, let me know!
|
c1cd6a11-26c5-40ca-92d5-039a30b388d4
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Meetup : Munich Meetup, October 28th
Discussion article for the meetup : Munich Meetup, October 28th
WHEN: 28 October 2012 03:00:00PM (+0200)
WHERE: Munich Central Station, Coffee Fellows cafe, inside the central station, *second* floor
The last meetup took place more than a year ago, so it's time for another one. Some of the topics discussed last time: Existential risks, anthropics, AI, metaethics, self-improvement and probably more that I can't remember. Of course there is much more to talk about and maybe we'll try some of those fancy rationality-games. (If the cafe sucks, we could easily go elsewhere. I've merely chosen the place, because it's relatively nice, near the central station and easy to find.) I'll be there with a LessWrong sign. Newbies and lurkers are very welcome!
ETA: I created a google group for the Munich LW meetup.
Send me your email adress to myusername@gmx.de and I'll add you.
Discussion article for the meetup : Munich Meetup, October 28th
|
7aef0f9f-bb9f-4462-8487-0881b95cab0e
|
trentmkelly/LessWrong-43k
|
LessWrong
|
DC Meetup: Discussion
LessWrong DC has had its first meetup! 13 people showed up, and it was pretty fun.
We have a google group here, and will have most of the planning there.
However, we haven't met all the LWers in the DC area yet, so that's what this thread is for.
We're meeting again on the 15th, but were wondering -- is there anyone in Northern Virginia who would be potentially interested, but didn't come?
|
ce4fa795-0dc6-46a6-8aa1-8b734b22740b
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Mediums Overpower Messages
I've observed that consuming certain kinds of media make me smarter and other kinds of media makes me dumber.
Makes me dumber:
* Videogames
* YouTube
* News
Makes me smarter:
* Books
* Audiobooks
* Direct messaging apps
By "smarter" I mean it holistically causes me to behave in a way that increases my overall rate of learning and quality of life. By "dumber" I mean the opposite.
For a long time I rejected this conclusion. Surely playing Kerbal Space Program must be more educational than reading Yu-gi-oh! manga. Nope. Yu-gi-oh! beats it by a long shot. I ran a long series of subjective personal experiments on a variety of timescales over many years. They all confirmed this theory[1]. The medium overpowers the message.
What I am watching on TV is irrelevant compared against the fact that I am watching TV.
I can even plot different mediums on a scale from "makes me dumber" to "makes me smarter" and it use to infer why different mediums have the effect they do.
* Makes Me Dumber
* [BAD & DANGEROUS] Videogames
* [BAD & DANGEROUS] Media feeds
* [BAD & DANGEROUS] YouTube
* [BAD & ADDICTIVE] News
* [BAD] Stack Exchange
* [BAD] Web surfing
* Blogs
* Movies
* Webcomics
* Comic books (fiction)
* [OKAY] Books (fiction)
* [OKAY] Podcasts
* [OKAY] Direct messaging (native language)
* [GOOD] Audiobooks (nonfiction)
* [GOOD] Books (nonfiction)
* [GOOD] Direct messaging (foreign language)
* [GOOD] Blogging
* [GOOD] Books (textbooks)
* [GOOD] Writing software
* [VERY GOOD] Making videos
* [VERY GOOD] Drawing comics
* [VERY GOOD] Spaced Repetition Software
* Makes Me Smarter
There is a symmetry to this sorting. Playing videogames is near the top but writing software is near the bottom. Watching YouTube is near the top but making YouTube videos is near the bottom. The smarter creating a certain kind of media makes me the dumber consuming it does.
In fact, this whole list is just sorted by difficulty. The harder a medium is to consume (or create,
|
f0d319c8-62cb-448e-8241-824ac73d043c
|
trentmkelly/LessWrong-43k
|
LessWrong
|
[Question] What's your Elevator Pitch For Rationality?
You're talking with someone you like, and they ask you what you mean by rationality, or why you keep going to LessWrong meetups. Or you meet someone who might be interested in the site.
What do you say to them? If you had to explain to someone what LW-style rationality is in 30 seconds, how would you do it? What's your elevator pitch? Has anyone had any success with a particular pitch?
My Current Pitch:
My current best one, made up on the spot, lacking any foreplanning, basically consists of:
"Basically, our brains are pretty bad at forming accurate beliefs, and bad in fairly systematic ways. I could show you one, if you want."
Playing the triplet game with them, then revealing that the numbers just need to be ascending
Upon failure, "Basically, your brain just doesn't look for examples that disprove your hypothesis, so you didn't notice that it could have a been a more general rule. There are a bunch of others, and I'm interested in learning about them so that I can correct for them."
My Thoughts on That:
It's massively effective at convincing people that cognitive biases exist (when they're in the 80% that fails, which has always been the case for me so far), but pretty much entirely useless as a rationality pitch. It doesn't explain at all why people should care about having accurate beliefs, and takes it as a given that that would be important.
It's also far too dry and unfun (compared to say, Methods), and has the unfortunate side effect of making people feel like they've gotten tricked. It makes it look non-cultish though.
I suspect that other people can do better, and I'll comment later with one that I actually put thought into. There's a pretty good chance that I'll use a few of the more upvoted ones and see how they go over.
|
42f6a42f-9e4a-4559-bd21-438653a0417d
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Twitter Polls: Evidence is Evidence
Follow-up to: Law of No Evidence
Recently, there was some debate about a few Twitter polls, which led into a dispute over the usefulness of Twitter polls in general and how to deal with biased and potentially misleading evidence.
Agnus Callard is explicitly asking the same question I asked, which is the opposite of ignoring sample bias: What is accounting for the difference?
Sample selection is definitely one of the explanations here. One can also point to several other key differences.
1. My poll asks about you, Patrick asks about how others seem.
2. My poll asks about struggle, Patrick asks about stability.
3. My poll asks about a year versus a point in time, a potential flaw.
4. My poll asks about now, Patrick asks about since pandemic onset.
None of this is well-controlled or ‘scientific’ in the Science sense. No one is saying any of this is conclusive or precise.
What is ‘bad’ evidence if it isn’t weak evidence? Adam’s theory here is that it is misleading evidence. That makes sense as a potential distinction. Under this model:
1. Weak evidence induces a small Bayesian update in the correct direction.
2. Bad evidence can induce an update in the wrong direction.
Usually, people with such taxonomies will also think that strong evidence by default trumps weak evidence, allowing you to entirely ignore it. That is not how that works. Either something has a likelihood ratio, or it doesn’t.
The question is, what to do about the danger that someone might misinterpret the data and update ‘wrong’?
I love that the account is called ‘Deconstruction Guide.’ Thanks, kind sir.
Whether or not this ‘depends on the poll’ depends on what level of technically correct we are on, and one can go back and forth on that several times. The fully correct answer is: Yes, some info. You always know that the person chose to make the poll, and how many people chose to respond given the level of exposure, and the responses always tell you something, even if the choices were ‘G
|
6add62cc-12db-43d0-a602-69fb04528669
|
StampyAI/alignment-research-dataset/eaforum
|
Effective Altruism Forum
|
Critiques of prominent AI safety labs: Conjecture
*Crossposted to* [*LessWrong*](https://www.lesswrong.com/posts/9jvrQToSq3CYvoeHf/critiques-of-prominent-ai-safety-labs-conjecture)*.*
*This is the second post in this* [*sequence*](https://forum.effectivealtruism.org/s/GcxnnGRGy8bondvBB) *and covers Conjecture. We recommending reading our brief* [*introduction*](https://forum.effectivealtruism.org/posts/N4LKrktopDs5Qdqgn/an-introduction-to-critiques-of-prominent-ai-safety) *to the sequence for added context on our motivations, who we are, and our overarching views on alignment research.*
[Conjecture](https://www.conjecture.dev/) is a for-profit alignment startup founded in late 2021 by Connor Leahy, Sid Black and Gabriel Alfour, which aims to scale applied alignment research. Based in London, Conjecture has received $10 million in funding from venture capitalists (VCs), and recruits heavily from the EA movement.
We shared a draft of this document with Conjecture for feedback prior to publication (and include their response [below](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Communication_with_Conjecture)). We also requested feedback on a draft from a small group of experienced alignment researchers from various organizations, and have invited them to share their views in the comments of this post. We'd like to invite others to share their thoughts in the comments, or anonymously via [this form](https://forms.gle/wbfy37owDR2yEp3L7).
Key Takeaways
=============
*For those with limited knowledge and context on Conjecture, we recommend first reading or skimming the* [*About Conjecture*](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#About_Conjecture) *section.*
*Time to read the core sections (Criticisms & Suggestions and Our views on Conjecture) is 22 minutes.*
[Criticisms and Suggestions](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Criticisms_and_Suggestions)
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
* We think Conjecture’s research is low quality ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Low_quality_research)).
+ Their posts don’t always make assumptions clear, don’t make it clear what evidence base they have for a given hypothesis, and evidence is frequently cherry-picked. We also think their bar for publishing is too low, which decreases the signal to noise ratio. Conjecture has acknowledged some of these criticisms, but not all ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#General_thoughts_on_Conjecture_s_research)).
+ We make specific critiques of examples of their research from their initial research agenda ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#General_thoughts_on_Conjecture_s_research)).
+ There is limited information available on their new research direction (cognitive emulation), but from the publicly available information it appears extremely challenging and so we are skeptical as to its tractability ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#New_research_agenda__Nov_22___Present_)).
* We have some concerns with the CEO’s character and trustworthiness because, in order of importance ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#CEO_s_character_and_trustworthiness)):
+ The CEO and Conjecture have misrepresented themselves to external parties multiple times ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Conjecture_and_their_CEO_misrepresent_themselves_to_various_parties));
+ The CEO’s involvement in EleutherAI and Stability AI has contributed to race dynamics ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Contributions_to_race_dynamics));
+ The CEO previously overstated his accomplishments in 2019 (when an undergrad) ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Overstatement_of_accomplishments_and_lack_of_attention_to_precision));
+ The CEO has been inconsistent over time regarding his position on releasing LLMs ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Inconsistency_over_time_regarding_releasing_LLMs)).
* We believe Conjecture has scaled too quickly before demonstrating they have promising research results, and believe this will make it harder for them to pivot in the future ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Scaling_too_quickly)).
* We are concerned that Conjecture does not have a clear plan for balancing profit and safety motives ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Unclear_plan_for_balancing_profit_and_safety_motives)).
* Conjecture has had limited meaningful engagement with external actors ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Limited_meaningful_engagement_with_external_actors)):
+ Conjecture lacks productive communication with external actors within the TAIS community, often reacting defensively to negative feedback and failing to address core points ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Lack_of_productive_communication_between_TAIS_researchers_and_Conjecture_staff));
+ Conjecture has not engaged sufficiently with the broader ML community, we think they would receive valuable feedback by engaging more. We’ve written more about this [previously](https://forum.effectivealtruism.org/posts/DaRvpDHHdaoad9Tfu/critiques-of-prominent-ai-safety-labs-redwood-research) ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Lack_of_engagement_with_the_broader_ML_community)).
[Our views on Conjecture](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Our_views_on_Conjecture)
---------------------------------------------------------------------------------------------------------------------------------------------------------------
* We would generally recommend working at most other AI safety organizations above Conjecture given their history of [low quality research](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Low_quality_research) and the leadership team’s [lack of research experience](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Team) (and thus mentorship) and concerns with the [CEO’s character and trustworthiness](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#CEO_s_character_and_trustworthiness) ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#We_would_advise_against_working_at_Conjecture)). [[1]](#fnctaancos4gn)
* We would advise Conjecture to avoid unilateral engagement with important stakeholders and strive to represent their place in the TAIS ecosystem accurately because [they have misrepresented themselves](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Conjecture_and_their_CEO_misrepresent_themselves_to_various_parties) multiple times ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#We_would_advise_Conjecture_to_avoid_unilateral_engagement_with_important_stakeholders_and_represent_their_place_in_the_TAIS_ecosystem_accurately)).
* We do not think that Conjecture should receive additional funding before addressing key concerns because of the reasons cited above ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#We_do_not_think_that_Conjecture_should_receive_additional_funding_before_addressing_key_concerns)).
* We encourage the TAIS and EA community members and organizations reflect to what extent they want to legitimize Conjecture until Conjecture addresses these concerns ([read more](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#We_encourage_TAIS_and_EA_community_members_to_consider_to_what_extent_they_want_to_legitimize_Conjecture_until_Conjecture_addresses_these_concerns)).
About Conjecture
================
Funding
-------
Conjecture received (primarily via commercial investment) roughly $10 million in 2022. [According to them](https://www.lesswrong.com/posts/jfq2BH5kfQqu2vYv3/we-are-conjecture-a-new-alignment-research-startup), they’ve received VC backing from [Nat Friedman](https://en.wikipedia.org/wiki/Nat_Friedman) (ex-CEO of GitHub), [Patrick](https://en.wikipedia.org/wiki/Patrick_Collison) and [John](https://en.wikipedia.org/wiki/John_Collison) Collison (co-founders of Stripe), [Daniel Gross](https://dcgross.com/) (investor and cofounder of a [startup accelerator](https://pioneer.app/)), [Andrej Karpathy](https://karpathy.ai/) (ex-OpenAI), [Sam Bankman-Fried](https://en.wikipedia.org/wiki/Sam_Bankman-Fried), [Arthur Breitman](https://golden.com/wiki/Arthur_Breitman-Y3RZ9ZM) and others. We are not aware of any later funding rounds, but it’s possible they have raised more since then.
Outputs
-------
### **Products**
[Verbalize](http://verbalize.dev) is an automatic transcription model. This is a [B2C](https://www.investopedia.com/terms/b/btoc.asp) [SaaS](https://en.wikipedia.org/wiki/Software_as_a_service) product and was released in early 2023. Our impression is that it's easy to use but no more powerful than existing open-source models like Whisper, although we are not aware of any detailed empirical evaluation. We do not think the product has seen commercial success yet, as it was released recently. Our estimate is that about one third of Conjecture’s team are actively working on developing products.
### **Alignment Research**
Conjecture studies large language models (LLMs), with a focus on empirical and conceptual work. Mechanistic interpretability was a particular focus, with output such as the [polytope lens](https://www.lesswrong.com/posts/eDicGjD9yte6FLSie/interpreting-neural-networks-through-the-polytope-lens), [sparse autoencoders](https://www.alignmentforum.org/posts/z6QQJbtpkEAX3Aojj/interim-research-report-taking-features-out-of-superposition) and [analyzing the SVD of weight matrices](https://www.lesswrong.com/posts/mkbGjzxD8d8XqKHzA/the-singular-value-decompositions-of-transformer-weight), as well as work more broadly seeking to better understand LLMs, such as [simulator theory](https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators).
They have recently pivoted away from this agenda towards [cognitive emulation](https://conjecture.dev/cognitive-emulation-proposal), which is reminiscent of [process-based supervision](https://ought.org/updates/2022-04-06-process). Here is a [link to their full research agenda and publication list.](https://www.conjecture.dev/research) Due to their infohazard policy (see below), some of their research may not have been publicly released.
### [Infohazard policy](https://www.lesswrong.com/posts/Gs29k3beHiqWFZqnn/conjecture-internal-infohazard-policy)
Conjecture developed an infohazard policy in their first few months and shared it publicly to encourage other organizations to publish or adopt similar policies. They say that while many actors were “verbally supportive of the policy, no other organization has publicly committed to a similar policy”.
### Governance outreach
We understand that CEO Connor Leahy does a lot of outreach to policymakers in the UK, and capabilities researchers at other prominent AI companies. He’s also appeared on several podcasts ([1](https://podcasts.apple.com/us/podcast/connor-leahy-eleutherai-conjecture/id1565088425?i=1000570841369), FLI ([1](https://www.youtube.com/watch?v=2RjuJzmafAA&pp=ygUQQ09OTk9SIExFQUhZIEZMSQ%3D%3D),[2](https://futureoflife.org/podcast/connor-leahy-on-the-state-of-ai-and-alignment-research/),[3](https://www.youtube.com/watch?v=ps_CCGvgLS8),[4](https://www.youtube.com/watch?v=cSL3Zau1X8g)), [3](https://deepdive.opensource.org/podcast/when-hackers-take-on-ai-sci-fi-or-the-future/), [4](https://www.youtube.com/watch?v=k6M_ScSBF6A&skip_registered_account_check=true), [5](https://www.youtube.com/watch?v=T8tHmQiYzVA)) and been interviewed by several journalists ([1](https://www.foxnews.com/politics/tech-ceo-warns-ai-risks-human-extinction-experts-rally-behind-6-month-pause), [2](https://www.foxnews.com/opinion/shocking-response-ai-what-do-now-before-late), [3](https://techmonitor.ai/technology/you-loved-chatgpt-wait-until-you-see-its-rivals), [4](https://www.cdotrends.com/story/18057/growing-pains-generative-ai?refresh=auto), [5](https://www.theguardian.com/technology/2023/apr/23/pope-jacket-napalm-recipes-how-worrying-is-ai-rapid-growth), [6](https://sifted.eu/articles/connor-leahy-ai-alignment), [7](https://time.com/6256529/bing-openai-chatgpt-danger-alignment/), [8](https://www.cdotrends.com/story/18057/growing-pains-generative-ai)).
### Incubator Program
Adam Shimi ran an incubator called [Refine](https://www.lesswrong.com/posts/D7epkkJb3CqDTYgX9/refine-an-incubator-for-conceptual-alignment-research-bets) in 2022, whose purpose was to create new independent conceptual researchers and help them build original research agendas. Based on [Adam’s retrospective](https://www.lesswrong.com/posts/3zZjF3YKJ257x79mu/what-i-learned-running-refine), it seems like this project wasn’t successful at achieving its goals and Adam is now pursuing different projects.
Team
----
The Conjecture team started as a team of 4 employees in late 2021 and have grown to at least 22 employees now (according to their [LinkedIn](https://www.linkedin.com/search/results/people/?facetCurrentCompany=%5B81967152%5D&sid=VDa)), with most employees joining in 2022.
Their CEO, [Connor Leahy](https://de.linkedin.com/in/connor-j-leahy), has a technical background (with 2 years of professional machine learning experience and a Computer Science undergrad) and partially replicated GPT-2 in 2019 (discussed in more detail [below](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Overstatement_of_accomplishments_and_lack_of_attention_to_precision)). Their Chief of Staff, has experience with staffing and building team culture from her time at McKinsey, and has similar experience at Meta. Their co-founder [Gabriel Alfour](https://www.lesswrong.com/users/gabriel-alfour-1) has the most relevant technical and scaling experience as the CEO of [Marigold](https://www.marigold.dev/),[[2]](#fn214ktml68vih)[[1]](#fn-uaoMyuFtDsDBHmemB-1) a firm performing core development on the Tezos cryptocurrency infrastructure with over 30 staff members.
Two individuals collectively publishing under the pseudonym [janus](https://www.alignmentforum.org/users/janus-1) published [simulator theory](https://www.alignmentforum.org/posts/vJFdjigzmcXMhNTsx/simulators), one of Conjecture's outputs that we understand the TAIS community to have been most favorable towards. They left Conjecture in late 2022. More recently, many researchers working on mechanistic interpretability left the team after Conjecture's pivot towards cognitive emulation. Those departing include [Lee Sharkey, the l](https://www.lesswrong.com/users/lee_sharkey)ead author on the [sparse autoencoders](https://www.lesswrong.com/posts/z6QQJbtpkEAX3Aojj/interim-research-report-taking-features-out-of-superposition) post and a contributor to the [polytope lens post.](https://www.lesswrong.com/posts/eDicGjD9yte6FLSie/interpreting-neural-networks-through-the-polytope-lens)
Conjecture in the TAIS ecosystem
--------------------------------
Conjecture staff are frequent contributors on the [Alignment Forum](https://www.alignmentforum.org/) and recruit heavily from the EA movement. Their CEO has appeared on a few EA podcasts (including several times on the [FLI podcast](https://futureoflife.org/podcast/connor-leahy-on-ai-safety-and-why-the-world-is-fragile/)). Some TAIS researchers are positive about their work. They fiscally sponsor two TAIS field-building programs, MATS and ARENA, in London (where they are based).
Their team also spent a month in the Bay Area in 2022 (when [many TAIS researchers were visiting](https://forum.effectivealtruism.org/posts/M5GoKkWtBKEGMCFHn/what-s-the-theory-of-change-of-come-to-the-bay-over-the) through programs like MLAB, SERI MATS and on independent grants). Conjecture made an effort to build relationships with researchers, decisionmakers and grantmakers, and were actively fundraising from EA funders during this period. 3-4 Conjecture staff regularly worked out of the [Lightcone Offices](https://www.lesswrong.com/posts/eR7Su77N2nK3e5YRZ/the-lesswrong-team-is-now-lightcone-infrastructure-come-work-3), with a peak of ~11 staff on a single day. The largest event run by Conjecture was an [EA Global afterparty](https://docs.google.com/document/d/1QKhccpgjpGwqAaEAfQ2gNM-jNU4rfcLqAU9mp7YqTHc/edit#heading=h.1h5d8kgiegft) hosted at a Lightcone venue, with a couple hundred attendees, predominantly TAIS researchers.
Criticisms and Suggestions
==========================
Low quality research
--------------------
### General thoughts on Conjecture’s research
We believe most of Conjecture’s publicly available research to date is low-quality. It is hard to find an accurate reference class for Conjecture’s work, as members have prioritized releasing small, regular updates. We think the bar of a workshop research paper is appropriate because it has a lower bar for novelty while still having it’s own technical research. We don’t think Conjecture’s research (combined) would meet this bar.[[3]](#fnwec7a9ngmaa)
As we discuss [below](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#General_thoughts_on_Conjecture_s_research), Conjecture does not present their research findings in a systematic way that would make it accessible for others to review and critique. Conjecture’s work often consists of isolated observations that are not built upon or adequately tested in other settings.
**Our suggestions:**We recommend Conjecture focus more on developing empirically testable theories, and also suggest they introduce an internal peer-review process to evaluate the rigor of work prior to publicly disseminating their results. Conjecture might also benefit from having researchers and reviewers work through (although not rigidly stick to) the [Machine Learning Reproducibility Checklist](https://www.cs.mcgill.ca/~jpineau/ReproducibilityChecklist.pdf).
### Lack of team's prior track record or experience in alignment and ML research[[4]](#fnvsdgxdp9poa)
These limitations may in part be because Conjecture is a young organization with a relatively inexperienced research team, a point they have readily acknowledged in retrospectives and when criticized on research quality. Conjecture's leadership staff staff has a relatively limited alignment research track record. By contrast, at an organization like ARC, Paul Christiano has a clear track record of producing useful conceptual insights (e.g. [Iterated Distillation and Amplification](https://ai-alignment.com/iterated-distillation-and-amplification-157debfd1616)) as well as practical advances (e.g. [Deep RL From Human Preferences](https://arxiv.org/abs/1706.03741)) prior to ARC’s founding.[[5]](#fnxebtp4b28zt) We're not aware of any equally significant advances from any key staff members at Conjecture (including those who have left).
However, taking their youth and inexperience into account, we still think their research is below the bar for funding or other significant support. When we take into account the funding that Conjecture has (at least $10M raised in their last round), we think they are significantly underperforming standard academic research labs (see our discussion on this in the [Redwood post](https://forum.effectivealtruism.org/posts/DaRvpDHHdaoad9Tfu/critiques-of-prominent-ai-safety-labs-redwood-research#Underwhelming_Research_Output); we are significantly more excited about Redwood’s research than Conjecture).
**Our suggestions:**We believe they could significantly improve their research output by seeking out mentorship from more experienced ML or alignment researchers, and recommend they do this in the future.
### Initial research agenda (March 2022 - Nov 2022)
Conjecture’s initial research agenda focused on interpretability, conceptual alignment and epistemology. Based on [feedback](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Communication_with_Conjecture) from Conjecture, they are much more excited about their new research direction in cognitive emulation (discussed in the following section). However, as an organization's past track record is one of the best predictors of their future impact, we believe it is important to understand Conjecture's previous approach.
Our understanding is that Conjecture was pursuing a hits-based strategy. In general, we are excited by hits-based approaches. Even if they don't succeed, rigorous negative results can save future researchers from going down dead-ends. We've generally not found their preliminary findings to significantly update our views, although some researchers have found those results useful.[[6]](#fnskfo5vad4i)
To Conjecture's credit, they acknowledged a number of mistakes in their [retrospective](https://www.alignmentforum.org/posts/bXTNKjsD4y3fabhwR/conjecture-a-retrospective-after-8-months-of-work-1). For example, they note that their simulators post was overinvested in, and "more experienced alignment researchers who have already developed their own deep intuitions about GPT-like models didn’t find the framing helpful." However, there are several issues we identify (such as lack of rigor) that are not discussed in the retrospective. There are also issues discussed in the retrospective where Conjecture leadership comes to the opposite conclusion to us: for example, Conjecture writes that they "overinvested in legibility and polish" whereas we found many of their posts to be difficult to understand and evaluate.
We believe three representative posts, which Conjecture leadership were excited by as of 2022 Q3, were: janus’s post on [simulators](https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators), Sid and Lee's post on [polytopes](https://www.alignmentforum.org/posts/eDicGjD9yte6FLSie/interpreting-neural-networks-through-the-polytope-lens#comments), and their [infohazard policy](https://www.lesswrong.com/posts/Gs29k3beHiqWFZqnn/conjecture-internal-infohazard-policy). These accomplishments were also highlighted in their [retrospective](https://www.alignmentforum.org/posts/bXTNKjsD4y3fabhwR/conjecture-a-retrospective-after-8-months-of-work-1). Although we find these posts to have some merit, we would overall assess them as having limited impact. Concretely, we would evaluate Redwood's [Indirect Object Identification](https://arxiv.org/abs/2211.00593) or [Causal Scrubbing](https://www.lesswrong.com/posts/JvZhhzycHu2Yd57RN/causal-scrubbing-a-method-for-rigorously-testing) papers as both more novel and scientifically rigorous. We discuss their infohazard policy, simulators and polytopes post in turn below.
Their infohazard policy is a fairly standard approach to siloing research, and is analogous to structures common in hedge funds or classified research projects. It may be positive for Conjecture to have adopted such a policy (although it introduces risks of concentrating power in the CEO, discussed in the next section), but it does not provide any particular demonstration of research capability.
The simulators and polytopes posts are both at an exploratory stage, with limited empirical evidence and unclear hypotheses. Compared to similar exploratory work (e.g. the [Alignment Research Center](https://www.alignment.org/)), we think Conjecture doesn’t make their assumptions clear enough and have too low a bar for sharing, reducing the signal-to-noise ratio and diluting standards in the field. When they do provide evidence, it appears to be cherry picked.
Their posts also do not clearly state the degree of belief they have in different hypotheses. Based on private conversations with Conjecture staff, they often appear very confident in their views and results of their research despite relatively weak evidence for them. In the simulators post, for example, they describe sufficiently large LLMs as converging to simulators capable of simulating “simulacra”: different generative processes that are consistent with the prompt. The post ends with speculative beliefs that they stated fairly confidently that took the framing to an extreme (e.g if the AI system adopts the “superintelligent AI persona” it’ll just be superintelligent).
We think the framing was overall helpful, especially to those newer to the field, although it can also sometimes be confusing: see e.g. [these](https://www.alignmentforum.org/posts/HD2s4mj4fsx6WtFAR/two-problems-with-simulators-as-a-frame) [critiques](https://www.alignmentforum.org/posts/dYnHLWMXCYdm9xu5j/simulator-framing-and-confusions-about-llms). The framing had limited novelty: our anecdotal impression is that most researchers working on language model alignment were already thinking along similar lines. The more speculative beliefs stated in the post are novel and significant if true, but the post does not present any rigorous argument or empirical evidence to support them. We believe it’s fine to start out with exploratory work that looks more like an op-ed, but at some point you need to submit your conjectures to theoretical or empirical tests.
**Our suggestions:**We encourage Conjecture to explicitly state their confidence levels in written output and make clear what evidence base they do or do not have for a given hypothesis (e.g. conceptual argument, theoretical result, empirical evidence).
### New research agenda (Nov 22 - Present)
Conjecture now has a new research direction exploring [cognitive emulation](https://www.alignmentforum.org/posts/ngEvKav9w57XrGQnb/cognitive-emulation-a-naive-ai-safety-proposal). The goal is to produce bounded agents that emulate human-like thought processes, rather than agents that produce good output but for alien reasons. However, it’s hard to evaluate this research direction as they are withholding details of their plan due to their infohazard policies. [Several commenters](https://www.lesswrong.com/posts/ngEvKav9w57XrGQnb#comments) have asked questions about the proposal including a request to list a [concrete research path](http://lesswrong.com/posts/ngEvKav9w57XrGQnb/cognitive-emulation-a-naive-ai-safety-proposal?commentId=49RNgXizHmMvRKBve), the [strategic assumptions](https://www.lesswrong.com/posts/ngEvKav9w57XrGQnb/cognitive-emulation-a-naive-ai-safety-proposal?commentId=zd7ve7YWgbwYd4n67) behind the agenda and [more details](https://www.lesswrong.com/posts/ngEvKav9w57XrGQnb/cognitive-emulation-a-naive-ai-safety-proposal?commentId=vvurB4rZFEPoHwnpz) to help readers evaluate if agenda’s viability. Conjecture has so far not addressed those comments.[[7]](#fnzpfv0nexvqd)
On the face of it, this project is incredibly ambitious, and will require huge amounts of effort and talent. Because of this, details on how they will execute the project are important to understanding how promising this project may be.
**Our suggestions:**We encourage Conjecture to share some more technical detail unless there are concrete info-hazards they are concerned about. In the latter case we would suggest sharing details with a small pool of trusted TAIS researchers for external evaluation.
CEO’s character and trustworthiness
-----------------------------------
We are concerned by the character and trustworthiness of Conjecture's CEO, Connor Leahy. Connor has demonstrated a lack of attention to rigor and engagement with risky behavior, and he, along with other staff, have demonstrated an unwillingness to take external feedback.
Connor is clearly a highly driven individual, who has built a medium-sized organization in his early twenties. He has shown a willingness to engage with arguments and change his mind on safety concerns, for example delaying the release of his GPT-2 replication. Moreover, in recent years Connor has been a vocal public advocate for safety: although we disagree in some cases with the framing of the resulting media articles, in general we are excited to see greater public awareness of safety risks.[[8]](#fndpieap2pmgw)
The character of an organization’s founder and CEO is always an important consideration, especially for early-stage companies. We believe this consideration is particularly strong in the case of Conjecture:
1. Conjecture engages in governance outreach that involves building relationships between government actors and the TAIS community, and there are multiple accounts of Conjecture misrepresenting themselves.
2. As the primary stakeholder & CEO, Connor will be responsible for balancing incentives to develop capabilities from stakeholders ([see below](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Unclear_plan_for_balancing_profit_and_safety_motives)).
3. Conjecture's [infohazard policy](https://conjecture.dev/information-hazard-policy) has the consequence of heavily centralizing power to the CEO (even more so than a typical tech company). The policy mandates projects are siloed, and staff may be unaware of the details (or even the existence) of significant fractions of Conjecture's work. The CEO is Conjecture's "appointed infohazard coordinator" with "access to all secrets and private projects" – and thus is the only person with full visibility. This could substantially reduce staff's ability to evaluate Conjecture's strategy and provide feedback internally. Additionally, if they don’t have the full information, they may not know if Conjecture is contributing to AI risk.[[9]](#fn0d3x86i7rb4) We are uncertain the degree to which this is a problem given Conjecture's current level of internal secrecy.
### Conjecture and their CEO misrepresent themselves to various parties
We are generally worried that Connor will tell the story that he expects the recipient to find most compelling, making it challenging to confidently predict his and Conjecture's behavior. We have heard credible complaints of this from their interactions with funders. One experienced technical AI safety researcher recalled Connor saying that he will tell investors that they are very interested in making products, whereas the predominant focus of the company is on AI safety.
We have heard that Conjecture misrepresent themselves in engagement with the government, presenting themselves as experts with stature in the AIS community, when in reality they are not. We have heard reports that Conjecture's policy outreach is decreasing goodwill with policymakers. We think there is a reasonable risk that Connor and Conjecture’s actions may be unilateralist and prevent important relationships from being formed by other actors in the future.
Unfortunately we are unable to give further details about these incidents as our sources have requested confidentiality; we understand this may be frustrating and acknowledge it is difficult for Conjecture to substantively address these concerns. We encourage individuals to talk to others in this space to draw their own conclusions about Conjecture's impact here.[[10]](#fn36inaz135n9)
**Our suggestions:**We recommend Connor be more honest and transparent about his beliefs, plans and Conjecture’s role in the TAIS ecosystem. We also recommend the Conjecture introduce a strong, robust governance structure. For example, they could change their corporate charter to implement a "springing governance" structure such that voting equity (but not political equity) shift to an independent board once they cross a certain valuation threshold.[[11]](#fnmog44tqkpj) ([see below](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Unclear_plan_for_balancing_profit_and_safety_motives)).
### Contributions to race dynamics
We believe that Connor Leahy has contributed to increasing race dynamics and accelerating capabilities research, through supporting the creation of [Stability AI](https://stability.ai/) through founding [EleutherAI](https://www.eleuther.ai/). EleutherAI is a community research group focused on open-source AI research founded in 2020. Under Connor's leadership, their plan was to [build and release large open-source models](https://blog.eleuther.ai/why-release-a-large-language-model/) to allow more people to work on important TAIS research that is only possible on pretrained LLMs. At the time, several members of the TAIS community, including [Dan Hendrycks](https://people.eecs.berkeley.edu/~hendrycks/) (founder of [CAIS](https://www.safe.ai/)), privately warned Connor and EleutherAI that it would be hard to control an open source collective.
**Stability AI**
Stability AI brands themselves as an AGI lab and has raised $100M to fund research into and training of large, state-of-the-art models including [Stable Diffusion](https://stablediffusionweb.com/).[[12]](#fnc4mp3rs1uhn) The addition of another AGI focused lab is likely to further exacerbate race dynamics. Stability is currently releasing the majority of the work they create as open-source: this has some benefits, enabling a broader range of researchers (including alignment researchers) to study these models. However, it also has significant drawbacks, such as making potential moratoriums on capabilities research much harder (if not impossible) to enforce. To our knowledge, Stability AI has not done much algorithmic advancement yet.
EleutherAI was pivotal in the creation of [Stability AI](https://stability.ai/). Our understanding is that the founder of Stability AI, Emad Mostaque, was active on the EleutherAI Discord and recruited much of his initial team from there. On the research side, Stability AI [credited EleutherAI](https://stability.ai/blog/stable-diffusion-announcement) as supporting the initial version of [Stable Diffusion](https://stablediffusionweb.com/) in August 2022, as well as their most recent open-source language model release [StableLM](https://stability.ai/blog/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models) in April 2023. Emad (in Feb 2023) [described the situation as](https://sarahguo.com/blog/emadmostaque): “Eleuther basically split into two. Part of it is Stability and the people who work here on capabilities. The other part is Conjecture that does specific work on alignment, and they're also based here in London.”
Stability AI continues to provide much of EleutherAI’s compute and is a [sponsor](https://www.eleuther.ai/about) of EluetherAI, alongside Nat Friedman (who also [invested in Conjecture](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Funding)). Legally, Stability AI directly employed key staff of EleutherAI in a relationship we believe was similar to fiscal sponsorship. We understand that EleutherAI have recently transitioned to employing staff directly via their own non-profit entity (Connor and Emad sit on the [board](https://www.eleuther.ai/about)).
**EleutherAI**
EleutherAI is notable for having developed open-source LLMs such as [GPT-NeoX](https://github.com/EleutherAI/gpt-neox). In the [announcement post](https://blog.eleuther.ai/announcing-20b/) in February 2022, they claimed that "GPT-NeoX-20B is, to our knowledge, the largest publicly accessible pretrained general-purpose autoregressive language model, and we expect it to perform well on many tasks."
We do not think that there was much meaningful alignment output from EleutherAI itself during Connor’s tenure – most of the research [published](https://www.eleuther.ai/papers) is capabilities research, and the published alignment research is of mixed quality. On the positive side, EleutherAI’s open-source models have enabled some valuable safety research. For example, GPT-J was used in the [ROME paper](https://arxiv.org/abs/2202.05262) and is widely used in [Jacob Steinhardt’s lab](https://jsteinhardt.stat.berkeley.edu/). EleutherAI is also developing a team focusing on interpretability, their initial work includes developing the [tuned lens](https://arxiv.org/pdf/2303.08112.pdf) in a collaboration with FAR AI and academics from Boston and Toronto.
Connor’s founding and management of EleutherAI indicates to us that he was overly optimistic about rapidly growing a community of people interested in language models and attracting industry sponsorship translating into meaningful alignment research. We see EleutherAI as having mostly failed at its goals of AI safety, and instead accelerated capabilities via their role in creating [Stability.ai](http://Stability.ai) and Stable Diffusion.
In particular, EleutherAI's supporters were primarily interested in gaining access to state-of-the-art LLM capabilities with limited interest in safety. For example, the company [Coreweave](https://www.coreweave.com/) provided EleutherAI with compute and then used their models to sell a LM inference API called [GooseAI](https://goose.ai/). We conjecture that the incentive to please their sponsors, enabling further scale-up, may have contributed to EleutherAI's limited safety output.
We feel more positively about Conjecture than early-stage EleutherAI given Conjecture's explicit alignment research focus, but are concerned that Connor appears to be bringing a very similar strategy to Conjecture as to EleutherAI: [scaling](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Scaling_too_quickly) before producing tangible alignment research progress and attracting investment from external actors (primarily investors) with opposing incentives that they [may not be able to withstand](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Unclear_plan_for_balancing_profit_and_safety_motives). We would encourage Conjecture to share a clear theory of change which includes safeguards against these risks.
To be clear, we think Conjecture's contribution to race dynamics is far less than that of OpenAI or Anthropic, both of which have received funding and attracted talent from the EA ecosystem. We would assess OpenAI as being extremely harmful for the world. We are uncertain on Anthropic: they have undoubtedly contributed to race dynamics (albeit less so than OpenAI), but have also produced substantial safety research. We will discuss Anthropic further in an upcoming post, but in either case we do not think that AGI companies pushing forward capabilities should exempt Conjecture or other organizations from criticisms.
### Overstatement of accomplishments and lack of attention to precision
In June 2019, Connor [claimed to have replicated GPT-2](https://medium.com/@NPCollapse/replicating-gpt2-1-5b-86454a7f26af) while he was an undergraduate. However, his results were inaccurate and his 1.5B parameter model was weaker than even the smallest GPT-2 series model.[[13]](#fnm16dnsa1cvm) He later [admitted](https://medium.com/@NPCollapse/addendum-evaluation-of-my-model-e6734b51a830) to these mistakes, explaining that his metric code was flawed and that he commingled training and evaluation datasets. Additionally, he said that he didn’t evaluate the strength of his final model, only one halfway through training. He said the reason he did this was because “I got cold feet once I realized what I was sitting on [something potentially impressive] and acted rashly.”[[14]](#fnvw004n19v6) We think this points to a general lack of thoughtfulness for making true and accurate claims.
We don’t want to unfairly hold people’s mistakes from their college days against them – many people exaggerate or overestimate (intentionally or not) their own accomplishments. Even a partial replica of GPT-2 is an impressive technical accomplishment for an undergraduate, so this project does attest to Connor's technical abilities. It is also positive that he admitted his mistake publicly. However, overall we do believe the project demonstrates a lack of attention to detail and rigor. Moreover, we haven’t seen signs that his behavior has dramatically changed.
### Inconsistency over time regarding releasing LLMs
Connor has changed his stance more than once regarding whether to publicly release LLMs. Given this, it is difficult to be confident that Conjecture's current approach of defaulting to secrecy will persist over time.
In July 2019, Connor [released](https://medium.com/@NPCollapse/replicating-gpt2-1-5b-86454a7f26af) the source code used to train his replica along with pretrained models comparable in size to the already released GPT-2 117M and and 345M models. The release of the training code seems hasty, enabling actors with sufficient compute but limited engineering skills to train their own, potentially superior, models. At this point, Connor was planning to release the full 1.5B parameter model to the public, but was [persuaded not to](https://medium.com/@NPCollapse/the-hacker-learns-to-trust-62f3c1490f51).[[15]](#fn0x8fawjl4rq) In the end, he delayed releasing the model to Nov 13 2019, a week after [OpenAI released](https://openai.com/research/gpt-2-1-5b-release) their 1.5B parameter version, on [his personal GitHub](https://github.com/ConnorJL/GPT2/commit/936fe2a21fad221cb07d0157c00fbb0780c7d114).
In June 2021 Connor changed his mind and argued that [releasing large language models would be beneficial to alignment](https://blog.eleuther.ai/why-release-a-large-language-model/) as part of the team at EleutherAI (see [discussion above](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Contributions_to_race_dynamics)). In Feb 2022, EleutherAI released an open-source 20B parameter model, [GPT-NeoX](https://arxiv.org/abs/2204.06745). Their [stated goal](https://web.archive.org/web/20220122021338/https://www.eleuther.ai/faq/), endorsed by Connor in several places, was to "train a model comparable to the biggest GPT-3 (175 billion parameters)" and release it publicly. Regarding the potential harm of releasing models, we find Connor's arguments plausible – whether releasing open-source models closer to the state-of-the-art is beneficial or not remains a contested point. However, we are confident that sufficiently capable models should not be open-sourced, and expect strong open-source positive messaging to be counterproductive. We think EleutherAI made an unforced error by not at least making some gesture towards publication norms (e.g. they could have pursued a staggered release giving early access to vetted researchers).
In July 2022, Connor shared Conjecture’s [Infohazard Policy](https://www.lesswrong.com/posts/Gs29k3beHiqWFZqnn/conjecture-internal-infohazard-policy). This policy is amongst the most restrictive at any AI company – even more restrictive than what we would advocate for. To the best of our knowledge, Conjecture's Infohazard Policy is an internal policy that can be overturned by Connor (acting as chief executive), or by a majority of their owners (of whom Connor as a founder will have a significant stake). Thus we are hesitant to rely on Conjecture’s Infohazard Policy remaining strictly enforced, especially if subject to commercial pressures.
Scaling too quickly
-------------------
We think Conjecture has grown too quickly, from 0 to at least 22 staff from 2021 to 2022. During this time, they [have not had](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Low_quality_research) what we would consider to be insightful or promising outputs, making them analogous to a very early stage start-up. This is a missed opportunity: their founding team and early employees include some talented individuals who, given time and the right feedback, might well have been able to identify a promising approach.
We believe that Conjecture’s basic theory of change for scaling is:
**1)** they’ve gotten good results relative to how young they are, even though the results themselves are not that insightful or promising in absolute terms, *and*
**2)** the way to improve these results is to scale the team so that they can test out more ideas and get more feedback on what does and doesn’t work.
Regarding **1)** we think that others of similar experience level – and substantially less funding – have produced higher-quality output. Concretely, we are more excited about Redwood’s research than Conjecture (see our [criticisms of Conjecture’s research](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Low_quality_research)), despite being critical of [Redwood](https://forum.effectivealtruism.org/posts/DaRvpDHHdaoad9Tfu/critiques-of-prominent-ai-safety-labs-redwood-research#Underwhelming_Research_Output)’s cost-effectiveness to date.[[16]](#fn3ankaej7hsr) Notably, Redwood drew on a similar talent pool to Conjecture, largely hiring people without prior ML research experience.
Regarding **2)**, we disagree that scaling up will improve their research quality. In general, the standard [lean startup](https://theleanstartup.com/) team advice is that it’s important to keep your team small while you are finding product-market fit or, in Conjecture's case, developing an exciting research agenda. We think it’s very likely Conjecture will want to make major pivots in the next few years. Rapid growth will make it harder for them to pivot. With growing scale, more time will be spent on management, and it will be easier to get people locked into the wrong project or create dynamics where people are more likely to defend their pet projects. We can't think of examples where scale up has taken place successfully before finding product-market fit.
This growth would be challenging to manage in any organization. However, in our opinion alignment research is more challenging to scale than a traditional tech start-up due to the weaker feedback loops: it's much harder to tell if your alignment research direction is promising than whether you've found product-market fit.
Compounding this problem, their founding team Connor, Sid and Gabriel have limited experience in scaling research organizations. Connor and Sid's experience primarily comes from co-founding EleutherAI, a decentralized research collective: their frustrations with that *lack of organization* are part of what drove them to found Conjecture. Gabriel has the [most relevant experience](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Team).
Conjecture appeared to have rapid scaling plans, but their growth has slowed in 2023. Our understanding is that this slow-down is primarily due to them being unable to raise adequate funding for their expansion plans.
**Our suggestions for Conjecture**:
* Freeze hiring of junior staff until they identify scalable research directions that they and others in the alignment community are excited by. Conjecture may still benefit from making a small number of strategic hires that can help them manage their current scale and continue to grow, such as senior research engineers and people who have experience managing large teams.
* Consider focusing on one area (e.g. technical research) and keeping other teams (e.g. product and governance) lean, or even consider whether they need them.
* While we don’t think it’s ideal to let go of staff, we tentatively suggest Conjecture consider whether it might be worth making the team smaller to focus on improving their research quality, before growing again.
Unclear plan for balancing profit and safety motives
----------------------------------------------------
According to their [introduction post](https://www.alignmentforum.org/posts/jfq2BH5kfQqu2vYv3/we-are-conjecture-a-new-alignment-research-startup), they think being a for-profit company is the best way to reach their goal because it lets them “scale investment quickly while maintaining as much freedom as possible to expand alignment research.” We think this could be challenging in practice: scaling investment requires delivering results that investors find impressive, as well as giving investors some control over the firm in the form of voting shares and, frequently, board seats.
Conjecture has received [substantial backing](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Funding) from several prominent VCs. This is impressive, but since many of their backers (to our knowledge) have little interest in alignment, Conjecture will be under pressure to develop a pathway to profitability in order to raise further funds.
Many routes to developing a profitable AI company have significant capabilities externalities. Conjecture’s CEO [has indicated](https://www.lesswrong.com/posts/rtEtTybuCcDWLk7N9/ama-conjecture-a-new-alignment-startup?commentId=pZmerzhJSADkwNJZx) they plan to build "a reliable pipeline to build and test new product ideas" on top of internal language models. Although this seems less bad than the OpenAI model of directly advancing the state-of-the-art in language models, we expect demonstrations of commercially viable products using language models to lead to increased investment in the entire ecosystem – not just Conjecture.
For example, if Conjecture does hit upon a promising product, it would likely be easy for a competitor to copy them. Worse, the competitor might be able to build a better product by using state-of-the-art models (e.g. those available via the OpenAI API). To keep up, Conjecture would then have to either start training state-of-the-art models themselves (introducing race dynamics), or use state-of-the-art models from competitors (and ultimately provide revenue to them).
Conjecture may have good responses to this. Perhaps there are products which are technically intricate to develop or have other barriers to entry making competition unlikely, and/or where Conjecture's internal models are sufficient. We don’t have reason to believe Verbalize falls into this category as there are several other competitors already (e.g. [fireflies.ai](https://fireflies.ai/), [otter.ai](https://otter.ai/), [gong.io](https://www.gong.io/)). We would encourage Conjecture to share any such plans they have to simultaneously serve two communities (for-profit VCs and TAIS), with sometimes conflicting priorities, for review with both sets of stakeholders.
Our impression is that they may not have a solid plan here (but we would invite them to share their plans if they do). Conjecture was trying to raise a series B from EA-aligned investors to become an alignment research organization. This funding round largely failed, causing them to pivot to focus more on VC funding. Based on their past actions we think it’s likely that they may eventually hit a wall with regards to product development, and decide to focus on scaling language models to get better results, contributing to race dynamics. In fairness to Conjecture, we would consider the race risk of Conjecture to be much smaller than that of Anthropic, which operates at a much bigger scale, is scaling much more rapidly, and has had more commercial success with its products.
It's not uncommon that people and orgs who conceive of or present themselves as AIS focused end up advancing capabilities much more than safety. OpenAI is perhaps the most egregious case of this, but we are also concerned about Anthropic (and will write about this in a future post). These examples should make us suspect that by default Conjecture's for-profit nature will end up causing it to advance capabilities, and demand a clear and detailed plan to avoid this to be convinced otherwise.
**Our suggestions:**In addition to sharing their plans for review, we recommend that Conjecture introduce robust corporate governance structures. Our understanding is that Conjecture is currently structured as a standard for-profit start-up with the founders controlling the majority of voting shares and around a third of the company owned by VCs. This is notably worse than OpenAI LP, structured as a "capped-profit" corporation with non-profit OpenAI, Inc. the sole controlling shareholder.[[17]](#fn03hvsf56riql) One option would be for Conjecture to implement a "springing governance" structure in which given some trigger (such as signs that AGI is imminent, or that their total investment exceeds some threshold) its voting shares become controlled by a board of external advisors. This would pass governance power, but not financial equity, to people who Conjecture considers to be a good board – rather than being controlled wholly by their founding team.
Limited meaningful engagement with external actors
--------------------------------------------------
### **Lack of productive communication between TAIS researchers and Conjecture staff**
We know several members of the EA and TAIS community who have tried to share feedback privately with Conjecture but found it very challenging. When negative feedback is shared, members of the Conjecture team sometimes do not engage meaningfully with it, missing the key point or reacting defensively. Conjecture leadership will provide many counter-arguments, none of which address the core point, or are particularly strong. This is reminiscent of the [Gish gallop](https://en.wikipedia.org/wiki/Gish_gallop#:~:text=The%20Gish%20gallop%20%2F%CB%88%C9%A1,or%20strength%20of%20those%20arguments.) rhetorical technique, which can overwhelm interlocutors as it’s very difficult to rebut each counter-argument. Some Conjecture staff members also frequently imply that the person giving the criticism has ulterior motives or motivated reasoning.
It can be hard to hear criticism of a project you are passionate about and have invested considerable time in, so it’s natural that Conjecture staff are defensive over their work.
**Our suggestions:**We recommend Conjecture staff and especially leadership make an effort to constructively engage in criticism, seeking to understand where the critique is coming from, and take appropriate steps to correct misunderstandings and/or resolve the substance of the critique.
### Lack of engagement with the broader ML community
Conjecture primarily disseminates their findings on the [Alignment Forum](https://www.alignmentforum.org/). However, many of their topics (particularly interpretability) are at least adjacent to active research fields, such that a range of academic and industry researchers could both provide valuable feedback on Conjecture's research and gain insights from their findings.
Conjecture is not alone in this: as [we wrote previously](https://forum.effectivealtruism.org/posts/DaRvpDHHdaoad9Tfu/critiques-of-prominent-ai-safety-labs-redwood-research), we also think that Redwood could engage further with the ML community. Conjecture has not published any peer-reviewed articles, so we think they would benefit even more than Redwood from publishing their work and receiving external feedback.
**Our suggestions:**We recommend Conjecture focus on developing what they consider to be their most insightful research projects into a conference-level paper, and hiring more experienced ML research scientists or advisors to help them both effectively communicate their research and improve rigor.
Our views on Conjecture
=======================
We are genuinely concerned about **C**onjecture’s trustworthiness and how they might negatively affect the TAIS community and the TAIS community’s efforts to reduce risk from AGI. These are the main changes we call for, in rough order of importance.
**We would generally recommend working at most other AI safety organizations above Conjecture**[[1]](#fnctaancos4gn)
--------------------------------------------------------------------------------------------------------------------
We think Conjecture needs to address key concerns before we would actively recommend working there. We expect it to be rare that an individual would have an offer from Conjecture but not have access to other opportunities that are better. In practice many organizations end up competing for the same, relatively small pool of the very top candidates. Our guess is that most individuals who could receive an offer from Conjecture might be likely to receive offers from non-profits such as Redwood, CAIS and FAR; **alignment teams** at Anthropic, OpenAI and DeepMind; or working with academics such as Stuart Russell, Sam Bowman, Jacob Steinhardt or David Krueger. Note we would not in general recommend working at **capabilities-oriented teams** at Anthropic, OpenAI, DeepMind or other AGI-focused companies.
Conjecture seems relatively weak for skill building, since their leadership team is relatively inexperienced and also stretched thin due to Conjecture's rapid scaling. We think people could pursue roles which provide better mentorship, like being a research assistant or PhD student in academia, or working in an ML engineering position in an applied team at a major tech company. These are generally close to capabilities-neutral, and can make individuals vastly more productive. We think these paths can absorb a relatively large amount of talent, although we note that most AI/ML fields are fairly competitive.
We also don’t generally recommend people pursue independent research, as we believe it’s a poor fit for most people. If someone feels their only good options are to do independent research or work at Conjecture, we feel somewhat ambivalent between these two options.
We could imagine Conjecture being the best option for a small fraction of people who are (a) excited by their current CoEm approach, (b) can operate independently in an environment with limited mentorship, (c) are confident they can withstand internal pressure (if there is a push to work on capabilities).
In general, we think that the attractiveness of working at an organization that is connected to the EA or TAIS communities makes it more likely for community members to take jobs at such organizations even if this will result in a lower lifetime impact than alternatives. Conjecture's sponsorship of TAIS field building efforts may also lead new talent, who are unfamiliar with Conjecture's history, to have more positive impression of them.
We would advise Conjecture to take care when engaging with important stakeholders and represent their place in the TAIS ecosystem accurately
--------------------------------------------------------------------------------------------------------------------------------------------
We are concerned that Conjecture has misrepresented themselves to various important stakeholders, including funders and policymakers. We think there is a reasonable risk that Connor and Conjecture’s outreach to policymakers and media is alarmist and may decrease the credibility of x-risk. These unilateral actions may therefore prevent important relationships from being formed by other actors in the future. This risk is further exacerbated by Connor’s unilateralist actions in the past, Conjecture’s overall reluctance to take feedback from external actors, and their premature and rapid scaling.
We do not think that Conjecture should receive additional funding before addressing key concerns
------------------------------------------------------------------------------------------------
We have substantial concerns with the organization’s trustworthiness and the CEO’s character. We would strongly recommend that any future funding from EA sources be conditional on Conjecture putting in place a robust corporate governance structure to bring them at least on par with other for-profit and alignment-sympathetic firms such as OpenAI and Anthropic.
Even absent these concerns, we would not currently recommend Conjecture for funding due to the lack of a clear impact track record despite a considerable initial investment of $10mn. To recommend funding, we would want to see both improvements in corporate governance and some signs of high-quality work that the TAIS community are excited by.
Largely we are in agreement with the status quo here: so far Conjecture has largely been unsuccessful fundraising from prominent EA funders, and where they have received funding it was for significantly less than their initial asks.
We encourage TAIS and EA community members to consider to what extent they want to legitimize Conjecture until Conjecture addresses these concerns
--------------------------------------------------------------------------------------------------------------------------------------------------
Conjecture has several red flags and a weak track record for impact. Although the TAIS and EA community have largely refrained from explicit endorsements of Conjecture (such as funding them), there are a variety of implicit endorsements. These include tabling at EA Global career fairs, Lightcone hosting Conjecture events and inviting Conjecture staff, field-building organizations such as MATS and ARENA working with Conjecture as a fiscal sponsor,[[18]](#fnnyuuynbq3ce) as well as a variety of individuals in the community (mostly unaware of these issues) recommending Conjecture as a place to work.
To clarify, we think individuals should still read and engage with Conjecture's research where they judge it to be individually worth their time. We also welcome public debates involving Conjecture staff, such as the one between [Paul Christiano and Gabriel Alfour](https://www.lesswrong.com/posts/pgpFHLJnv7AdSi3qS/christiano-arc-and-ga-conjecture-discuss-alignment-cruxes). Our goal is not to shun Conjecture, but to avoid giving them undue influence until their research track record and governance structure improves.
We recognize that balancing these considerations can be tricky, which is why our main recommendation is to encourage people to spend time actively reflecting on how they want to engage with Conjecture in light of the information we present in this post (alongside other independent sources).
Appendix
========
Communication with Conjecture
-----------------------------
We shared a draft of this post with Conjecture to review, and have included their full response (as they indicated they would post it publicly) below. We thank them for their engagement and made several minor updates to the post in response, however we disagree with several key claims made by Conjecture in their response. We describe the changes we made, and where we disagree, in the [subsequent section](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Brief_response_and_changes_we_made).
### Conjecture’s Reply
Hi,
Thank you for your engagement with Conjecture’s work and for providing us an opportunity to share our feedback.
As it stands, the document is a hit piece, whether intentional or not. It is written in a way such that it would not make sense for us to respond to points line-by-line. There are inaccuracies, critiques of outdated strategies, and references to private conversations where the details are obscured in ways that prevent us from responding substantively. The piece relies heavily on criticism of Connor, Conjecture CEO, but does not attempt to provide a balanced assessment: there are no positive comments written about Connor along with the critiques, and past mistakes he admitted to publicly are spun as examples of “low-integrity” behavior. Nuanced points such as the cost/benefit of releasing small open source models (pre-Chinchilla) are framed as “rash behavior,” even when you later write that you find Connor’s arguments “plausible.” Starting from this negative frame does not leave room for us to reply and trust that an object-level discussion will proceed.
We also find it surprising to see that most of the content of the piece is based on private discussions and documents shared between Conjecture, ~15 regrantors, and the FTX Future Fund team in August 2022. The piece does not disclose this context. Besides the fact that much of that information is outdated and used selectively, the information has either been leaked to the two anonymous authors, or one of the authors was directly involved in the regranting process. In either case, this is a violation of mutual confidentiality between Conjecture and regrantors/EA leadership involved in that channel.
We don’t mind sharing our past plans and discussions now and would be happy to publish the entire discussions from the Slack channel where those conversations took place (with consent of the other participants). However, it is a sad conclusion of that process that our openness to discussing strategy in front of regrantors formed the majority set of Bay Area TAIS leadership opinions about Conjecture that frame us as *not open*, despite these conversations being a deeper audit than pretty much any other TAIS organization.
We’d love to have a productive conversation here, but will only respond in detail if you reframe this post from a hit piece to something better informed. If your aim is to promote coordination, we would recommend asking questions about our plans and beliefs, focusing on the parts that do not make sense to you, and then writing your summary. Conjecture’s strategy is debatable, and we are open to changing it - and have done so in the past. Our research is also critiqueable: we agree that our research output has been weak and have already written about this publicly [here](https://www.lesswrong.com/posts/bXTNKjsD4y3fabhwR/conjecture-a-retrospective-after-8-months-of-work-1). But as described above, this post doesn’t attempt to engage with Conjecture’s current direction.
Going further, if the aim of your critique is to promote truth-seeking and transparency, we would gladly participate in a project about creating and maintaining a questionnaire that all AI orgs should respond to, so that there is as little ambiguity in their plans as possible. In our posts we have argued for [making AI lab’s safety plans more visible](https://www.lesswrong.com/posts/PE22QJSww8mpwh7bt/agi-in-sight-our-look-at-the-game-board), and previously ran a project of [public debates aimed at highlighting cruxes in research disagreements](https://www.lesswrong.com/posts/BEyAWbCdtWpSGxmun/retrospective-on-the-2022-conjecture-ai-discussions). Conjecture is open to our opinion being on the record, so much so that we have occasionally *declined private debates with individuals who don’t want to be on record.* This decision may contribute to some notion of our “lack of engagement with criticism.”
—
As a meta-point, we think that certain strategic disagreements between Conjecture and the Bay Area TAIS circles are bleeding into reputational accusations here. Conjecture has been critical of the role that EA actors have played in funding and supporting major AGI labs historically (OAI, Anthropic), and critical of current parts of the EA TAIS leadership and infrastructure that continue to support the development of superintelligence. For example, we do not think that GPT-4 should have been released and are concerned at the role that ARC’s benchmarking efforts played in safety-washing the model. These disagreements in the past have created friction, and we’d hazard that concerns about Conjecture taking “unilateral action” are predicted on this.
Instead of a more abstract notion of “race dynamics,” Conjecture’s main concern is that a couple of AI actors are unabashedly building superintelligence. We believe OpenAI, Deepmind, and Anthropic are not building superintelligence because the market and investors are demanding it. We believe they are building superintelligence because they want to, and because AGI has always been their aim. As such, we think you’re pointing the finger in the wrong direction here about acceleration risks.
If someone actually cares about curtailing “the race”, their best move would be to push for a ban on developing superintelligence and strongly oppose the organizations trying to build it. Deepmind, OpenAI, and Anthropic have each publicly pushed the AI state of the art. Deepmind and OpenAI have in [their charters](https://www.deepmind.com/about) that they want to build AGI. Anthropic’s [most recent pitch deck](https://techcrunch.com/2023/04/06/anthropics-5b-4-year-plan-to-take-on-openai/) states that they are planning to train an LLM orders of magnitude larger than competitors, and that “companies that train the best 2025/26 models will be too far ahead for anyone to catch up in subsequent cycles,” which is awfully close to talking about [DSA](https://forum.effectivealtruism.org/topics/decisive-strategic-advantage). No one at the leadership of these organizations (which you recommend people work at rather than Conjecture) have signed [FLI's open letter](https://futureoflife.org/open-letter/pause-giant-ai-experiments/) calling for a pause in AI development. Without an alignment solution, the reasonable thing for any organization to do is stop development, not [carve out space to continue building superintelligence unimpeded](https://openai.com/blog/governance-of-superintelligence).
While Conjecture strongly disagrees with the strategies preferred by many in the Bay Area TAIS circles, we’d hope that healthy conversations would reveal some of these cruxes and make it easier to coordinate. As written, your document assumes the Bay Area TAIS consensus is superior (despite being what contributed largely to the push for ASI), casts our alternative as “risking unilateral action,” and deepens the rift.
—
We have a [standing offer to anyone to debate with us](https://www.conjecture.dev/a-standing-offer-for-public-discussions-on-ai/), and we’d be very happy to discuss with you any part of our strategy, beliefs about AI risks, and research agenda.
More immediately, we encourage you to rewrite your post as a Q&A aimed at asking for our actual views before forming an opinion, or at a minimum, rewrite your post with more balance and breathing room to hear our view. As it stands, this post cleaves the relationship between part of the TAIS ecosystem and Conjecture further and is unproductive for both sides.
Given the importance of having these conversations in the open, we plan to make this reply public.
Thanks for your time and look forward to your response,
Conjecture C-Suite
### Brief response and changes we made
Conjecture opted not to respond to our points line-by-line and instead asked us to rewrite the post as a Q&A or “with more balance and breathing room to hear our view.” While we won’t be rewriting the post, we have made changes to the post in response to their feedback, some of which are outlined below.
Conjecture commented that the tone of the post was very negative, and in particular there was a lack of positive comments written about Connor. We have taken that feedback into consideration and have edited the tone to be more neutral & descriptive (with particular attention to the [section on Connor](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#CEO_s_character_and_trustworthiness)). Conjecture also noted that Connor admitted to some of his mistakes publicly. We had previously linked to Connor’s update post on the partial GPT-2 replication, but we edited [the section](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Overstatement_of_accomplishments_and_lack_of_attention_to_precision) to make it more clear that he did acknowledge his mistake. They also pointed out that we framed the point on releasing models “as “rash behavior,” even when you later write that you find Connor’s arguments “plausible.” We’ve changed this section to be more clear.
They say “this post doesn’t attempt to engage with Conjecture’s current direction.” As we write in [our section](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#New_research_agenda__Nov_22___Present_) on their cognitive emulation research, there is limited public information on their current research direction for us to comment on.
They believe that “most of the content of the piece is based on private discussions and documents shared between Conjecture, ~15 regrantors, and the FTX Future Fund team in August 2022.” This is not the case: the vast majority (90+%) of this post is based on publicly available information and our own views which were formed from our independent impression of Conjecture via conversations with them and other TAIS community members. We think the content they may be referring to is:
1. One conversation that we previously described in the [research section](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Low_quality_research) regarding Conjecture's original research priorities. We have removed this reference.
2. One point providing quantitative details of Conjecture's growth plans in the [scaling section](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Scaling_too_quickly), which we have removed the details of.
3. The [section](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Conjecture_and_their_CEO_misrepresent_themselves_to_various_parties) on how Conjecture and their CEO represent themselves to other parties. This information was not received from those private discussions and documents.
They say they wouldn’t mind “sharing our past plans and discussions now and would be happy to publish the entire discussions from the Slack channel where those conversations took place (with consent of the other participants).” We welcome and encourage the Conjecture team to share their past plans publicly.
They note that “Conjecture is open to our opinion being on the record, so much so that we have occasionally declined private debates with individuals who don’t want to be on record. This decision may contribute to some notion of our 'lack of engagement with criticism.'" This is not a reason for our comment on their lack of engagement. They mentioned they have “a [standing offer to anyone to debate with us](https://www.conjecture.dev/a-standing-offer-for-public-discussions-on-ai/)”. We appreciate the gesture, but do not have capacity to engage in something as in-depth as a public debate at this time (and many others who have given feedback don’t either).
Conjecture points out the role “EA actors have played in funding and supporting major AGI labs historically (OAI, Anthropic)”, that our “document assumes the Bay Area TAIS consensus is superior … casts our alternative as “risking unilateral action”, and that “these disagreements in the past have created friction, and we’d hazard that concerns about Conjecture taking “unilateral action” are predicted on this.” We outline our specific concerns on unilateralist action, which don’t have to do with Conjecture’s critiques of EA TAIS actors, [here](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#We_would_advise_Conjecture_to_take_care_when_engaging_with_important_stakeholders_and_represent_their_place_in_the_TAIS_ecosystem_accurately). Examples of disagreements with TAIS actors that they cite include:
* Conjecture being critical of the role EA actors have played in funding/supporting major AGI labs.
* EA TAIS leadership that continue to support development of AGI.
* They don’t think GPT-4 should have been released.
* They are concerned that ARC’s benchmarking efforts might have safety-washed GPT-4.
We are also concerned about the role that EA actors have and potentially continue to play in supporting AGI labs (we will cover some of these concerns in our upcoming post on Anthropic). We think that Conjecture’s views on ARC are reasonable (although we may not agree with their view). Further, many other EAs and TAIS community members have expressed concerns on this topic, and about OpenAI in particular. We do not think holding this view is particularly controversial or something that people would be critical of. Views like this did not factor into our critique.
Finally, they propose that (rather than critiquing them), we should push for a ban on AGI and oppose organizations trying to build it (OpenAI, DM & Anthropic). While we agree that other labs are concerning, that doesn’t mean that our concerns about Conjecture are erased.
Notes
-----
Changelog
---------
Note: Significant changes are listed with an (\*). Places where we changed are our views or recommendations are marked with a (^).
We've added footnotes signposting all edits for clarity.
As of June 16 2023:
* **^**Updated and enhanced our [recommendation](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#We_would_generally_recommend_working_at_most_other_AI_safety_organizations_above_Conjecture_1_) on working at Conjecture. We changed the top-line recommendation to be more precise and added more details of types of roles we would be more excited by, and added in some notes on people who might find it a good fit to work with Conjecture
* **\***Added a [subsection](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Lack_of_team_s_prior_track_record_or_experience_in_alignment_and_ML_research_4_) in for the lack of the team's track record
* We adjusted the [section](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#Low_quality_research) discussing the appropriate bar for Conjecture’s research to be more clear
* We added a [specific example](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#fnref03hvsf56riql) of a governance structure Conjecture could follow to our recommendations
* We modified the [section on misrepresentation](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#fnref36inaz135n9) and encouraged people to speak to others to validate our account and draw their own conclusions
* We [added specific open questions](https://forum.effectivealtruism.org/posts/gkfMLX4NWZdmpikto/critiques-of-prominent-ai-safety-labs-conjecture#New_research_agenda__Nov_22___Present_) regarding the CoEm agenda that have been raised by others.
* Various small grammar & language edits to make the post more clear and flow better
---
1. **[^](#fnrefctaancos4gn)**[June 15 2023] Edited to make the language more nuanced and add more explicit examples and comparisons.
2. **[^](#fnref214ktml68vih)**Gabriel Alfour is still listed as the CEO on Marigold's website: we are unsure if this information is out of date, or if Gabriel still holds this position. We also lack a clear understanding of what Marigold's output is, but spent limited time evaluating this.
3. **[^](#fnrefwec7a9ngmaa)**[June 14 2023] This paragraph was edited for clarity
4. **[^](#fnrefvsdgxdp9poa)**[June 16 2023] We added the section on the lack of the team's track record
5. **[^](#fnrefxebtp4b28zt)**[June 16 2023] In terms of the explicit comparison with ARC, we would like to note that ARC Theory's team size is an order of magnitude smaller than Conjecture. Based on ARC's recent [hiring post](https://www.alignment.org/blog/arc-is-hiring-theoretical-researchers/), it appears the theory team consists of just three individuals: Paul Christiano, Mark Xu and Jacob Hilton. If ARC had a team ten times larger and had spent close to $10 mn, then we would indeed be disappointed if there were not more concrete wins.
6. **[^](#fnrefskfo5vad4i)**[June 14 2023] Added this paragraph to explain our position on hits-based agendas.
7. **[^](#fnrefzpfv0nexvqd)**[June 14 2023] Added specific open questions people have with the CoEm agenda.
8. **[^](#fnrefdpieap2pmgw)**In particular, Connor has referred to AGI as god-like multiple times in interviews ([CNN](https://www.cnn.com/videos/world/2023/05/02/exp-artificial-intelligence-extinction-intw-fst-050201pseg1-cnni-world.cnn), [Sifted](https://sifted.eu/articles/connor-leahy-ai-alignment)). We are skeptical if this framing is helpful.
9. **[^](#fnref0d3x86i7rb4)**Employee retention is a key mechanism by which tech companies have been held accountable: for example, Google employees' protest over [Project Maven](https://gizmodo.com/google-employees-resign-in-protest-against-pentagon-con-1825729300) led to Google [withdrawing from the project](https://www.pbs.org/newshour/show/amid-pressure-from-employees-google-drops-pentagons-project-maven-account). Similarly, the exodus of AIS researchers from OpenAI to found Anthropic was partly fueled by concerns that OpenAI was contributing to AI risk.
10. **[^](#fnref36inaz135n9)**[June 15 2023]: We added a note encouraging people to speak to others and draw their own conclusions
11. **[^](#fnrefmog44tqkpj)**[June 15 2023]: We added a specific example of a governance structure Conjecture could follow to our recommendations
12. **[^](#fnrefc4mp3rs1uhn)**Stable Diffusion is a state-of-the-art generative model with similar performance to OpenAI’s DALL-E. They are open-source and open-access - there are no restrictions or filters, so you're not limited by what restrictions a company like OpenAI might apply. This means that people can use the model for abusive behavior (such as [deepfakes](https://apnews.com/article/deepfake-porn-celebrities-dalle-stable-diffusion-midjourney-ai-e7935e9922cda82fbcfb1e1a88d9443a))
13. **[^](#fnrefm16dnsa1cvm)**Connor reports a WikiText2 perplexity of 43.79 for his replica. This is considerably worse than the 18.34 perplexity achieved by GPT-2 1.5B on this dataset ([reported in Table 3 of Radfort et al](https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf)), and substantially worse than the perplexity achieved by even the smallest GPT-2 117M of 29.41. It is slightly worse than the previously reported state-of-the-art prior to the GPT-2 paper, of 39.14 ([reported in Table 2 of Gong et al](https://arxiv.org/pdf/1809.06858.pdf)). Overall, it’s a substantial accomplishment, especially for an undergraduate who built the entire training pipeline (including data scraping) from scratch, but is far from a replication.
14. **[^](#fnrefvw004n19v6)**Here is the full text from the relevant section of the article: “model is not identical to OpenAI’s because I simply didn’t have all the details of what they did … [and] the samples and metrics I have shown aren’t 100% accurate. For one, my metric code is flawed, I made several rookie mistakes in setting up accurate evaluation (let train and eval data mix, used metrics whose math I didn’t understand etc), and the model I used to generate the samples is in fact not the final trained model, but one about halfway through the training. I didn’t take my time to evaluate the strength of my model, I simply saw I had the same amount of hardware as OpenAI and code as close to the paper as possible and went with it. The reason for this is a simple human flaw: I got cold feet once I realized what I was sitting on and acted rashly.”
15. **[^](#fnref0x8fawjl4rq)**This was in part due to conversations with OpenAI and Buck Shlegeris (then at MIRI)
16. **[^](#fnref3ankaej7hsr)**Redwood and Conjecture have received similar levels of funding
17. **[^](#fnref03hvsf56riql)**Anthropic has a public benefit corporation structure, with [reports](https://www.ft.com/content/8de92f3a-228e-4bb8-961f-96f2dce70ebb) that it includes a long-term benefit committee of people unaffiliated with the company who can override the composition of its board. Overall we have too little information to judge whether this structure is better or worse than OpenAI’s, but both seem better than being a standard C-corporation.
18. **[^](#fnrefnyuuynbq3ce)**Conjecture has been active in running or supporting programs aimed at AI safety field-building. Most notably, they ran the [Refine](https://www.lesswrong.com/posts/D7epkkJb3CqDTYgX9/refine-an-incubator-for-conceptual-alignment-research-bets) incubator, and are currently fiscally sponsoring [ARENA](https://www.lesswrong.com/posts/bXTNKjsD4y3fabhwR/conjecture-a-retrospective-after-8-months-of-work-1) and [MATS](https://www.alignmentforum.org/posts/iR4kGzrWEJpXJ39ZB/seri-mats-program-winter-2022-cohort) for their London based cohort. We expect overall these programs are net-positive, and are grateful that Conjecture is contributing to them. However, it may have a chilling effect: individuals may be reluctant to criticize Conjecture if they want to be part of these sponsored programs. It may also cause attendees to be more likely than they otherwise would to work for Conjecture. We would encourage ARENA and MATS to find a more neutral fiscal sponsor in the UK to avoid potential conflicts of interest. For example, they could hire staff members using employer-of-record services such as [Deel](https://deel.com/) or [Remote](https://remote.com/). If Conjecture does continue fiscally sponsoring organizations, we would encourage them to adopt a clear legal separation between Conjecture and fiscally sponsored entities along with a conflict-of-interest policy to safeguard the independence of the fiscally sponsored entities.
|
d60dda8f-6980-422d-b2a7-ec41ca7b3352
|
trentmkelly/LessWrong-43k
|
LessWrong
|
IRL is hard
We show that assuming the existence of public-key cryptography, there is an environment in which Inverse Reinforcement Learning is computationally intractable, even though the "teacher" agent, the environment and the utility functions are computable in polynomial-time and there is only 1 bit of information to learn.
Construction
We call "Teacher" the role-model agent which optimizes the "true" utility function (in the context of AI safety this is typically a human) and "Student" the agent performing IRL (the AI). The environment we construct consists of two entities which we call the Gamemaster and the Adversary. The Adversary is an optimizer whose utility is minus the Teacher's utility. To reduce the setup to a single agent environment, we can interpret the Adversary as a specific optimization algorithm which either has access to the Teacher's source code or is a reinforcement learning algorithm that learns from preceding episodes. At any rate, the Adversary is only of secondary importance, as explained in the Discussion. The entire setup depends on a parameter n∈N w.r.t. which the complexities of the computations are measured.
The utility function of the Teacher depends on a bit α∈{0,1} which is unknown to the Student (i.e. the Student's utility function is defined according to a uniform prior on α) but known to the Gamemaster. Learning α is the purpose of IRL in this setting. The Student is assumed to passively observe the Teacher interacting with the environment for q(n) epsiodes for some polynomial q, after which it is required to assume the Teacher's role. For the environment we construct, no polynomial-time Student cannot achieve significantly better performance than the Lazy Student who ignores the Teacher and acts according to its prior.
We fix a public-key encryption scheme which encrypts any message of size n to a message of size n (a trapdoor permutation). Given a public (encryption) key b and a message x∈{0,1}n, we denote its encryption e(x,b). Give
|
e73f720c-9027-416f-844f-8477cae7bff3
|
StampyAI/alignment-research-dataset/special_docs
|
Other
|
What the AI Community Can Learn From Sneezing Ferrets and a Mutant Virus Debate
\*Lessons on publication norms for the AI community from biosecurity\*
--------------------------------------------------------------------
[](/@PartnershipAI?source=post\_page-----32432438a82e--------------------------------)[](https://medium.com/partnership-on-ai?source=post\_page-----32432438a82e--------------------------------)[Partnership on AI](/@PartnershipAI?source=post\_page-----32432438a82e--------------------------------)
·[Follow](/m/signin?actionUrl=https%3A%2F%2Fmedium.com%2F\_%2Fsubscribe%2Fuser%2Fda503babfab3&operation=register&redirect=https%3A%2F%2Fmedium.com%2Fpartnership-on-ai%2Flessons-for-the-ai-community-from-the-h5n1-controversy-32432438a82e&user=Partnership+on+AI&userId=da503babfab3&source=post\_page-da503babfab3----32432438a82e---------------------post\_header-----------)
Published in[AI&.](https://medium.com/partnership-on-ai?source=post\_page-----32432438a82e--------------------------------)·13 min read·Dec 8, 2020--
Listen
Share
![]()[By Jasmine Wang](https://www.partnershiponai.org/team/jasmine-wang/)
\*AI is by no means the first field to face the challenges of responsibly publishing and deploying high-stakes, dual-use research. This post represents the first in a series where we will examine how other fields have dealt with these issues and what the AI community can learn. It is presented as part of the Partnership on AI’s work on Publication Norms. Visit\* [\*our website here\*](https://www.partnershiponai.org/case-study/publication-norms/) \*for more information.\*
In the spring of 2012, Ron Fouchier contemplated a decision that could put him in prison for up to six years or cost him over $100,000 USD in fines. A white-haired 45-year-old who spent most days in a concrete research facility in Rotterdam, the Dutch virologist had suddenly become the focus of an international debate about the potential weaponization of influenza. His work, which involved mutating the H5N1 virus to make it more transmissible, had already set off an uproar that reverberated from US national security circles to the halls of the World Health Organization (WHO). Now, if he published his research without the Dutch government’s permission, he was told he could actually go to jail. Increasingly nervous that people perceived him to be a “mad scientist,” Fouchier [told the \*New Yorker\*](https://www.newyorker.com/magazine/2012/03/12/the-deadliest-virus) at the time that he felt like the subject of “an international witch hunt.”
Fouchier’s predicament might seem unimaginable to most scientists, like something out of a nightmare. But the chain of events that led him to this moment are worth studying for anyone whose work could pose public risks. This is especially true for those in artificial intelligence (AI) grappling with how to responsibly disseminate research with the potential for misuse, an important facet of publication norms.
AI and Machine Learning (ML) are increasingly being applied across new domains, including ones with safety-critical applications, leading many to ask what responsible AI/ML research, innovation, and deployment look like. In answering these questions, the AI community can and should consider how other fields have approached comparable challenges. Fouchier’s story offers important lessons from the biosecurity community and its long history of debate about publication norms. In particular, \*\*the H5N1 case illustrates the benefits (and inherent limitations) of third-party governance bodies — and, by implication, the importance of individual researcher responsibility.\*\*
H5N1 influenza, otherwise known as bird flu, is a severe respiratory disease. The naturally occurring H5N1 virus, however, rarely infects people and is almost never transmitted to others when it does. [According to Fouchier](https://www.nytimes.com/2011/12/27/science/debate-persists-on-deadly-flu-made-airborne.html), some scientists believed that H5N1 could never become airborne between mammals — he wanted to prove them wrong. [In his own words](https://www.newyorker.com/magazine/2012/03/12/the-deadliest-virus), his team first “mutated the hell out of H5N1.” They then squirted the mutated virus into the nose of a ferret and next implanted that ferret’s nasal fluid into another (and another, and another) ferret’s nose, making it sneeze. The virus spread.
Fouchier’s field of inquiry is known as [“gain-of-function” research](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4996883/), which aims to enhance the disease-causing abilities of pathogens to prepare the health community for future outbreaks. In the case of H5N1, which has an exceptionally high death-rate of [around 60 percent](https://www.cdc.gov/flu/avianflu/h5n1-people.htm), the naturally occurring virus lacked a crucial prerequisite to becoming a real pandemic risk: transmissibility. By simply transferring H5N1 from one animal to others, Fouchier had made it highly transmissible. His team had succeeded in making what he called “one of the most dangerous viruses you can imagine” airborne.
Fouchier presented his findings at an influenza conference in Malta in September 2011, announcing his intention to publish them in greater detail, which would enable others to replicate his research. After Fouchier submitted his paper to the academic journal \*Science\*, US health officials became aware of its existence, sending it to the National Science Advisory Board for Biosecurity (NSABB) for review. This advisory body was established in the wake of the 2001 anthrax attacks to provide oversight for dual-use biomedical research, [defined by the WHO](https://www.who.int/csr/durc/en/) as scientific work “which is intended for benefit but which might easily be misapplied to do harm.” Fouchier’s ferret experiment — which effectively made a deadly disease even more dangerous — [admittedly fell](https://www.nytimes.com/2011/12/22/health/security-in-h5n1-bird-flu-study-was-paramount-scientist-says.html) under this category.
The NSABB and its H5N1 Working Group subsequently [spent hundreds of hours](https://www.frontiersin.org/articles/10.3389/fpubh.2014.00117/full#B2) discussing the research. In December 2011, the NSABB unanimously recommended that key methodological detailswhich could enable replication of the experiments be withheld from publication in \*Science\*. In [a press release](https://www.nih.gov/news-events/news-releases/press-statement-nsabb-review-h5n1-research) announcing the decision, the National Institutes of Health (NIH) said that the US government was also working on a mechanism to grant researchers with a legitimate need access to the redacted information.
This recommendation received an immediate rejoinder. \*Science’s\* Editor-in-Chief said that the journal’s response would be “heavily dependent” on the creation of an explicit plan by the US government to share the omitted information with “responsible scientists who request it, as part of their legitimate efforts to improve public health and safety.” Fouchier himself [told the \*New York Times\*](https://www.nytimes.com/2011/12/22/health/security-in-h5n1-bird-flu-study-was-paramount-scientist-says.html) that around 1000 scientists from more than 100 laboratories worldwide had a need to know this information.
A second expert panel of 22 health officials, scientists, and journal editors convened by the World Health Organization (WHO) came to a far different (if not unanimous) conclusion from the NSABB, calling for full publication. Keiji Fukuda, the assistant director-general of health security and environment at the WHO, cited the difficulty and complexity of creating an information-sharing mechanism as a key rationale.
It was not just bureaucratic difficulties, but ambiguities about authority, control, and eligibility criteria that concerned the panel. “Who would hold on to the sensitive information?” Fukuda said [at a press conference](https://science.sciencemag.org/content/335/6071/899.full). “Under what conditions would that information be released? What are the other complicating factors? It was recognized that coming up with such a mechanism would be very difficult to do overnight, if not impossible.”
Anthony Fauci, the long-serving director of the National Institute of Allergy and Infectious Diseases who would become familiar to many during the COVID-19 pandemic, urged Fouchier and other H5N1 researchers to declare a voluntary moratorium on their work. Fauci [told the \*New York Times\*](https://www.nytimes.com/2012/01/21/science/scientists-to-pause-research-on-deadly-strain-of-bird-flu.html) that he viewed a moratorium as an act of good faith during a time of polarized opinion — an important one, given that the controversy could lead to excessive restrictions on future research. The scientists took his advice. In January 2012, 39 prominent influenza researchers from around the world, including Fouchier, announced they were voluntarily pausing H5N1 gain-of-function research for 60 days. This moratorium ended up lasting almost a year.
Just months after their unanimous recommendation, the NSABB reversed their position on Fouchier’s paper in March 2012, [voting 12–6 in favor](https://osp.od.nih.gov/wp-content/uploads/2013/06/03302012\_NSABB\_Recommendations.pdf) of a revised version being published in full. Their reasoning? While the research was still concerning, the revised manuscript did not appear to provide information that would immediately enable misuse. The board also cited the need for freely shared information among countries when responding to international pandemics. A majority of members of the NSABB still believed there was a critical need for a mechanism for disseminating sensitive scientific information. They acknowledged that there were complex questions and legal issues involved in developing such a mechanism, but nonetheless that a “feasible, secure mechanism for sharing sensitive scientific information” was essential, urging the US government to develop one.
This requested plan for disseminating the papers on a need-to-know basis didn’t materialize then and still does not exist now.
After getting a green light from the NSABB and WHO, Fouchier was told of a new challenge. His research needed to be approved for an export license from the Dutch government, which considered the publication of his research to be a potential violation of E.U. regulations aimed at preventing the proliferation of weapons of mass destruction and dual-use technologies. At first, he declined to apply for the export license, opposed to the precedent it would set, and intended to publish his research without one. In late April 2012, however, Fouchier decided to apply for a license, which was granted.
In the end, Fouchier’s paper appeared [in a special issue](https://science.sciencemag.org/content/336/6088/1521) of \*Science\* in June 2012. Had he gone through with publishing his research without the export license, the potential penalties included [up to six years in prison or a fine equivalent to $102,000 USD](https://www.nature.com/news/mutant-flu-researcher-plans-to-publish-even-without-permission-1.10469).
Even after acquiescing, Fouchier felt so strongly about the importance of unrestricted and free scientific expression that he continued to challenge the requirement legally. He lost his case, meaning similar papers by Dutch scientists would likely require export licenses. For years, Fouchier contested the verdict in several arenas of increasing authority. Eventually, the ruling was annulled on procedural grounds in July 2015, meaning future research would still be considered on a case-by-case basis.
“I’m disappointed,” Fouchier [told \*Science\*](https://www.sciencemag.org/news/2015/07/dutch-appeals-court-dodges-decision-hotly-debated-h5n1-papers) at the time, “They didn’t want to touch the hot potato and passed it on instead.”
Back in the US, similar virus research faced increased scrutiny in the wake of the Fouchier controversy. In October 2014, the White House announced an unprecedented “pause” on all federal funding of gain-of-function research involving influenza, MERS, or SARS. The pause was only lifted in December 2017 — with a new provision requiring gain-of-function proposals to be approved by a government panel. “We see this as a rigorous policy,” NIH Director Francis Collins [told the \*New York Times\*](https://www.nytimes.com/2017/12/19/health/lethal-viruses-nih.html). “We want to be sure we’re doing this right.”
For his part, Fouchier continued working on H5N1, and is currently the deputy head of the Erasmus MC department of Viroscience. A few years after US funding for his research stopped, the NIH [began to support](https://www.sciencemag.org/news/2019/02/exclusive-controversial-experiments-make-bird-flu-more-risky-poised-resume) Fouchier’s research again.
While the H5N1 controversy did not settle every issue it raised, the incident as a whole does leave the AI community with four intertwined lessons: \*\*Third-party institutions can result in more well-considered publication outcomes; absent other action, these entities might only be created in response to crises; these entities, however, are inherently limited in their capabilities; and, thus, researchers must exercise some degree of personal responsibility.\*\*
1. Third-party institutions can lead to more well-considered publication outcomes
---------------------------------------------------------------------------------
There are two main reasons why third-party institutions like the NSABB can lead to more thoughtful outcomes: They can counterbalance publishing incentives that bias researchers towards publication and they can provide additional expertise and critical context that individual researchers may lack.
Any researcher knows the desperate need to publish. Their reputation — and thus access to funding, collaboration opportunities, and publication venues — depends directly on the quality and quantity of papers they publish. There’s also the drive to advance scientific progress, and the very real possibility of societal benefits from their research. However, in high-stakes work there are inevitably trade-offs that need to be considered, and third-party institutions can counterbalance default publishing incentives, leading to more well-considered outcomes. In the case of H5N1, the NSABB brought up important publication considerations and proposed an alternative publication strategy. The WHO provided additional perspective and challenged the NSABB’s recommendations not out of individual interest, but as part of their mission concerned with public well-being.
A third-party institution, if properly composed, can provide multidisciplinary and security-relevant context on publication decisions. The NSABB was uniquely positioned with “Secret”-level security clearance (ranked only under “Top Secret”-level clearance in the US), allowing them to comment on issues of national security. They thus had additional decision-making context that enabled the assessment of security-relevant features of Fouchier’s research. The NSABB is also [multidisciplinary](http://www.virtualbiosecuritycenter.org/organizations/national-science-advisory-board-for-biosecurity-nsabb/), with as many as 25 voting members drawn variously from the microbiology, public health, national security, biosafety, and scientific publishing communities — in addition to non-voting ex officio members from 15 federal agencies.
The involvement of the NSABB thereby solved two important issues in responsible research: researcher bias towards publication and lack of domain expertise or critical information to judge risks. Despite AI research becoming increasingly high-stakes, there is no comparable institution. \*\*To provide essential balance to publication decisions, the AI community should explore the creation of a similar body.\*\*
2. Absent other action, such entities might only be created in response to crises
---------------------------------------------------------------------------------
Despite the benefits we’ve observed, most countries do not have a public entity overlooking responsible publication practices. The founding of the NSABB was precipitated by the anthrax attacks of 2001. That the US has such an entity was therefore not inevitable and was entirely path-dependent — most other countries do not have an analogous body.
Rather than wait for reactive measures, the AI community should consider establishing a third-party panel of experts as a community resource, which would have the immediate benefits of offering neutrality and multidisciplinarity. Such a centralized entity would also build up useful institutional knowledge and history over time that could be later transferred to any successor entity, government-led or otherwise.
\*\*The NSABB was only established after a serious biosecurity crisis. We should not wait for such a near-miss with AI.\*\*
3. The powers of third-party entities are structurally limited due to the international nature of science and the autonomy of researchers
-----------------------------------------------------------------------------------------------------------------------------------------
However, third-party institutions are not a panacea. As Fouchier’s case demonstrates, even a government body specifically created for the purpose of aiding publication decisions doesn’t completely solve related problems of coordination and the dissemination of information. A prominent philosopher of science, Heather Douglas, [concluded upon analyzing the H5N1 case](https://link.springer.com/article/10.1007/s10670-013-9538-0) that “stronger institutional mechanisms collectivizing the responsibilities of scientists with respect to dual-use research and its potential to cause great societal harm are clearly needed.” In the case of AI, some practices particular to AI research may render (even strong) institutional mechanisms less effective than necessary.
The AI community should ensure that any publication norms entity it establishes is sufficiently resourced. Not only were the NSABB’s recommendations non-binding, they also did not have the implementation capacity in-house to execute on a key condition of their recommendations being accepted: the ability to share redacted information with scientists with a need to know. Notably, this would have been a novel form of publication — [national policy on fundamental research](https://fas.org/irp/offdocs/nsdd/nsdd-189.htm) previously specified that it should either be openly published or classified. The WHO’s main argument for full publication was the lack of a public agency to execute limited disclosure of Fouchier’s paper. The fact remains that if a case like H5N1 occurred again, we would still find ourselves without the institutional capacity or direct accountability to implement such a mechanism. To be effective, an analogous institution for AI should be able to quickly and flexibly allocate financial and engineering resources to be able to respond adequately to unforeseen publication challenges.
Third-party institutions that are created by the state are geographically limited. A significant reason the NSABB had the influence it did on Ron Fouchier, who was Dutch, was because his research (as well as the institute he worked for) was NIH-funded. Additionally, he sought to publish his work in a peer-reviewed journal, \*Science\*, which had internal guidelines about what was straightforwardly publishable and what was not, and a procedure in place for escalation to the NSABB. Science is inherently a global enterprise, with many interlocking cross-national procedures and components. Those cross-border interdependencies add bureaucracy but also act as partial safeguards in a system where there is no obvious central authority. As the distribution of cutting-edge research work continues to become more globalized, state actors potentially become less influential.
Furthermore, some of the attributes of the scientific system that make such institutions useful do not generalize to AI. AI developers are more likely to publish their papers on arxiv.org, which doesn’t require peer review, and the most important developments in AI increasingly come from top industry labs, not government-funded institutes. Additionally, many AI researchers are employed in industry, where often research is never published due to company interests. Thus, the capabilities of a state-led entity like the NSABB would be even more limited for AI, given the lessened importance of controllable levers like publication, shifting more responsibility to the community and individual researchers.
\*\*These reasons reinforce the earlier recommendation for the AI community to explore the creation of its own third-party entity, which may prove to be more suitable than a state-led one.\*\*
4. Researchers must take on some responsibility for carefully thinking through their publication decisions
----------------------------------------------------------------------------------------------------------
Individual researchers cannot entirely offload the responsibility to consider publication impacts to outside entities. In the H5N1 case, actions taken on the part of the scientific community showcased two ways soft norms interacted with hard norms: the researchers’ voluntary moratorium allowed for the development of more well thought-out policy, while Fouchier’s individual actions weakened the impact of any future restrictions placed on his work.
Due to the limitations of third-party institutions discussed above, the AI community must accept that some responsibility to anticipate and mitigate the impacts of their work lies with researchers themselves. This is especially pertinent for AI, where publishing preprints on sites like arxiv.org (bypassing third-party review) is an established norm. It is not only undesirable, but also not possible, for a third-party entity to oversee a researcher’s work and publication decisions in their entirety.
The nature of scientific collaboration also limits the effectiveness of external mechanisms for information security. After the initial NSABB recommendation for partial redaction, many scientists noted that it might not effectively control information flow. The AI community should consider which stage of research would be an effective point to influence activity, and empower earlier management of research concerns. For example, the ML and neuroscience conference NeurIPS rejected four papers this year on ethical grounds, with seven others flagged for ethical concerns (conditionally accepted). If there were existing resources providing guidelines, the authors may have been able to preempt such concerns by proactively adapting their research.
\*\*With increased transparency about the criteria for ethical review, researchers could personally influence the direction, scope, and discussion of their research.\*\*
Drawing lessons from the field of biosecurity might seem like a daunting task for the artificial intelligence community, whose discussions of responsible publication practices are far more nascent. However, we believe insights derived from the H5N1 case are better understood as an opportunity, one to be proactive in advancing a community that supports the development of safe and responsible artificial intelligence for all.
\*This post would not be possible without remarkable journalistic and analytical work on this case published by the New York Times and the Nuclear Threat Initiative. We are also deeply grateful to Jack Clark (OpenAI), Dr. Heather Douglas (Michigan State University), Dr. Gregory Lewis (Future of Humanity Institute, Oxford University), and other reviewers for their contributions and feedback on earlier drafts of this piece.\*
|
a84ec56b-7ac7-4316-9b7e-2a23bd4cb4ef
|
StampyAI/alignment-research-dataset/youtube
|
Youtube Transcripts
|
Lectures by Olle Häggström on AI risk and long-term AI safety, part 2
so
welcome everyone to this uh
second lecture in my uh
mini series on
ai risk and long-term ai safety
see if i can get my slides in order
so there i should have them in full
screen can someone
confirm that it's looking okay
yeah it looks good
thank you
okay
so
the first
topic i will address today is
timelines uh when
can we expect radical things to happen
and
that's an issue of
quite some
controversy
so here for instance this leading ai
researcher andrew eng
the university of stanford and he has
also had leading positions at google and
at
chinese tech giant
baidu
and he gave an interview in
2015 or maybe 2016
which has been
it's he's had a quite a catchy thing
which has been quoted often
since then he said
there could be a race of killer robots
in the far future but i don't work on
not turning ai evil today for the same
reason i don't worry about the problem
of overpopulation on the planet mars
and you can contrast this
with
berkeley professor and and ai researcher
stuart russell
who
and this seems
directed uh
against
andrea and other colleagues who play
down the uh issue of a transformative ai
breakthrough he says this
within the ayat community the kind of
denialism is emerging even going as far
as denying the possibility of success in
achieving the long-term goals of ai it's
as if a bus driver with all of humanity
and past as passengers said yes i am
driving as hard as i can towards a cliff
but trust me we will run out of gas
before we get there
and
i've been really
wanting to to to
when the issue of
of ai risk and a
eye timelines is
as controversial
as it seems i want to give a balanced
view but it's difficult
because
when you look at andrew heng's
side on the iai timelands issue which is
fairly common among ai experts
it's it's hard to pinpoint what their
reasons are
because they don't spell them out
or
not anyway in
in a way that i can understand it
typically boils down to to things like
yes i'm an ai expert and i see how
difficult it is to make progress here
and we have so
much left
to do so so trust me it will take time
and and so i'm not denying that there
are arguments and i will show you some
of them at that point uh in the
direction of
um
long timelines until the big
breakthrough but the certainty uh which
was
with which this is expressed
by such as in android the android and
quote here seems to me unwarranted and
that has
led me a little bit i mean i really want
to understand but but when they don't
explain i'm led to psychologize and i
spend
an entire chapter
in my most recent book tankande machine
this is in swedish but chapter 10
is called ai risk
denial
and and uh i i try to
figure out uh what it is that causes
people like andrew eng to say these
things but it's uh
it's hard to understand and there are
some
people who
uh
deny ai risk altogether across the
entire spectrum of down-to-earth and
high-flying issues
but there are others who accept some or
all of the down-to-earth issues and and
restrict their
denial to to these more
extreme scenarios and so so there are
probably different things going on there
and and i offer a few candidate
explanations
in my book
um
what i think however
is a
fairly common core intuition driving
this is what i like to call the common
sense argument
and the common sense argument is simply
to point to an ai
failing at some task which is easy for
us humans and saying look how these
machines lack common sense clearly agi
is very far away
and i
expect that most of you have seen
videos that have been circulating on the
web of
robots
falling over their toes and or trying to
put on lipstick and
failing disaster
disastrously and a lot of such uh
similar things
uh and we all
like to laugh at that and
say oh yeah clearly
clearly
the singularity is not about to happen
here's another example from the
fall of
2020
ai
was at the task of steering a television
camera to follow a football game
and it was supposed to zoom in on the
ball and follow the ball throughout the
game but what happened was that it
discovered this ball like thing the head
of the bald
lines man and follow that throughout
the game
and the reaction to this is is the
obvious one this is a mistake that a
human would never make
uh ai's like common sense so so uh we're
still quite far from agi and i think
that this this is a misleading way to
look at things
because
common sense is a label that we tend to
put on anything that that humans still
perform better at
than
machines but you could turn this game
around and point to something that
that
machines have been doing better than
humans
all the way in the case of chess going
all the way to the 1990s
where
from the
uh ai's point of view one could say oh
look at this poor human playing back
black in this position uh
how
can he expose himself to this terrible
attack uh down the diagonal against the
king here uh clearly
if you play chels chess this way you
totally lack common sense so human lacks
common sense so it's a kind of
um
so we are still better at some things
and there are other things
that that machines
are better at and
it's unclear how this would play out and
if you really insist on the common sense
argument
then
all the way until the time point that we
have agi that will be something that we
humans do better at if we take the
definition of agi seriously so the
common sense argument will will work all
the way until the time point
when we actually do reach agi but i
think there's a deeper reason to be
skeptical about this kind of thinking
which is that
maybe for transformative things to
happen
uh
such as a robot takeover or something
like that
the ais will not need all the human
skills maybe it just
needs maybe they just need the the
appropriate
range of key skills which might not
include the ability to track
football
with a tv camera
during a football game
now one could
object to this
because it might seem intuitively
strange
that
ai's could be smart enough to pose a
threat to humanity even
when they're stupid enough to lack this
kind of common sense but i would like to
point to to this game of soccer
again as an example of a machine
achieving superhuman strength at the
particular task of solving sukubun
puzzles
but
without attaining common sense so this
is a paper
from a group of ai researchers at
cornell
in 2020 uh
demonstrating an ai program that uh
performs
much much better than humans uh at
solving
uh soccer ban puzzles
uh so
these uh
these puzzles tend to
take quite many steps in solving a
puzzle like this one where we have some
ten boxes
uh can easily
uh require several hundred
elaborate steps
around the this
maze
and and uh
these programs so so humans can solve uh
superbands on on this level but but
the
ai
[Music]
developed by by this group is able to
solve soccer bomb
muscles
on much larger scale and uh
it had it
uh
[Music]
goes
about through a kind of trial and error
process but this triangle an error
cannot happen blindly because of
combinatorial explosion it would be
really impossible uh to
find the solutions because the number of
possible pathways
grows exponentially with time and you
need to find the almost unique one
that
works
so
uh
the programs need to have some function
of what are promising positions to push
for and so on
and one thing that every human discovers
quickly maybe after the first one or two
minute of of play with soccer ban is
that if you push a box into a corner if
this guy moves here and then pushes this
box up into the corner here
then that box is stuck there's no way to
move it so if you if the task is to put
all the boxes here you've lost the game
so you have to start
over
and this program
doesn't know this
uh through it's it's uh uh
trial and error process it it again and
again it does this kind of thing and
starts over
so this is an example
where
superhuman
abilities can happen without acquiring
obvious looking common sense
and
if we insist on the intuition
from here that that machines cannot take
over the world without
common sense
then
the
we need some argument
uh to show that that
to show the impossibility of this
phenomenon of uh superhuman performance
without common sense scaling up from
from from limited problems like uh
sokoban towards the much bigger much
more complicated problem of taking over
in physical and social reality and i
know or no such
argument
so so i do think that
the relevant question is not
when will ai exceed us in all cognitive
domains as i said but rather when will
ai exceed us in enough domains to be
better than us at taking control
of the world
um
so i i
the example i mentioned uh playing chess
the best chess program is alpha zero
which was released in 2017 and caused
quite a splash among us chess players
but i don't think that that is a huge
concern because
this this program uh handles finite two
player zeros on board games with full
information for both players
and that particular cognitive ability
constitutes
i mean now i'm pulling percentage
figures just out of thin air but maybe
0.1 percent of the range of important
cognitive capacities
this is probably
overestimate on my part
through
my bias acquired by having played
competitive chess for
30 years or so
but you could look at
other uh capabilities that are probably
more key
uh in this spectrum of of of
possible or of kinds of intelligence and
text generation
i don't know but it seems to me totally
plausible to suggest that this would be
more like 30 or perhaps even more of of
our
range of relevant capabilities
i mean when you think about what we
humans do to to influence the world
we do this mostly through uh language
acts
uh
and i mean to take one morbid example
um
adolf hitler had huge influence he could
created quite a stir
in the middle of the
20th century
and i think he did that almost
exclusively or perhaps even totally
exclusively but by language acts
nothing else really and therefore it's
quite interesting to
look at the increasingly
advanced ai software that we now have
for natural language processing and i
will come back to this
later in this talk
so while i do talk a lot about agi
artificial general intelligence in this
lecture series
i think there are downsides to this
terminology and and one such
downside
uh
is the one i i mentioned now that it it
encourages the intuition that
the
machines must require all human
capabilities before extreme things can
happen which i i think is
leads our thoughts in the wrong
direction and therefore i'm very
sympathetic to attempts by experts
who we will hear more from in this
lecture series to frame the issue not
around hgi but in terms of a couple of
related concepts there is the idea of
transformative ai
which
ajaya cotra
looks at and which i will in work that i
will describe uh shortly transformative
ai is defined roughly as a.i with the
capability of having an impact on the
world
which is
at least on the of the order of
magnitude of everything we humans
have done so far
including the agricultural and
industrial revolutions
um
andrew kritsch and david kruger have a
related
concept they call pre-potent ai which is
a little more restrictive than
transformative ai
uh pre-potent ai is
essentially transformative ai with the
additional property that once it has
been deployed
it is unstoppable by humans or by any
collection
of humans
um
so these are concepts that we will touch
upon later so whether we talk about adi
uh transformative ai and or
prepotent ai
there are at least vaguely related
concepts there is the question of when
can we expect it to happen
and and we heard uh andrew eng say
saying that this is so far in the future
that we don't need to worry now
uh other suggestions here's a famous one
ray kurzweil in his book the
singularities near from 2005.
he said very bluntly 2045 this is when i
predict the
so-called singularity to happen leading
to in super intelligent machines what is
the singularity maybe i mentioned
something about this on monday it's it's
it's a
phase in in
ai development
where the ai starts to self-improve in a
spiral which can possibly
start to escalate uh very fast because
of positive feedback mechanisms i'll
come back to this a lit in a little
while
uh
and
once
the development has passed through this
singularity phase it it's assumed to
have reached
super intelligent levels meaning agi
plus much more so that the
machines
are vastly more intelligent than humans
across most or
most or all of the intelligence spectrum
but this is obviously just one thinker's
overconfident view
i mean i don't
buy into kurzweil's
reasoning at all he basically combines
overconfident estimates of how much
computing is going on in the brain
in the human brain
with extrapolating moore's law
and asking when can we expect
a computer to have as much computing
power
as the human brain and
then he adds a decade or something for
for um
software development or so but it's it's
all very uh much pulled out of thin air
so so
we can look more broadly at what the
ai
research community things and and there
have been a number of surveys that ask
ai specialists
when do you expect these sort of radical
things to happen
vincent miller whom we encountered in on
monday together with nick bostrom
had one of the early
such surveys it was carried out
in
at a couple of ai conferences in 2012
and 2013
the paper was not published until 2016
but but
so respondents were asked various things
including estimating uh when they
expect
agi
to
be developed and the answers were very
broadly spread out and roughly in the
same way as in all of these surveys
namely
spread out
over the entire
coming 100 years
fairly evenly
so there are people saying 10 years and
people saying 30 years and there are
those in 50 years or 70 years 100 years
a few give even longer timelines
and
a small minority say that this is
basically impossible so we can never
expect it to happen but but most of them
in the next 100 years and and
fairly evenly spread
around this time
and there are also uh
a couple of questions in this survey
about
what
these researchers expect will happen
then and whether consequences will be
good
or bad for humanity and
there are a fair number although
not everyone uh
or air researchers who
do think that there's a risk
uh that pretty bad things can happen but
many of these same researchers think
that there are good chances that
we will fix things and then
things can go fabulously well if we can
put a super intelligent machine in the
service of human flourishing flourishing
and this in particular this is the
perspective of
of nick bostrom in his book and so on
so
partly in response to this
ai researchers oren
ezioni
did his own survey in 2016 which he
presented in the mit technology review
as counter to the miller bostrom
claims
and and he has this very catchy title no
the experts don't think super
intelligent ai is a threat to humanity
and if you look at this paper and and
look at the actual survey he carried out
it's very
he
he argues actually quite dishonestly so
first of all
uh this survey doesn't even address
the uh question of whether
super intelligent ai
is a possibility
i i mean
a positive
has positive potential or if it's a
threat or or or if both
possibilities
[Music]
can happen
the goodness or badness of an ai
breakthrough is not addressed in this
survey it looks exclusively at timelines
and it does this in a more
coarse-grained way than
uh
than in the mueller boston
survey where people could give their own
dates for for various stages of
development
in that sony's
survey
there's just a multiple choice question
when do you expect this to happen and
the
highest category
uh
in terms of long timelines was 25 or
more years
which
i think it was like 67
who answered 25 or more years
and and
ezione just decided to interpret this as
meaning there is nothing to worry about
which is
also first of all this is totally
consistent with the um
miller bostrom survey so he really
didn't discover anything new he just
looked at
the same sort of
opinion spectrum through a more
coarse-grained uh lens
and then he decided that those who said
25 years or more think that there is
nothing to worry about and this is like
thinking
that since
most of the effects of uh
anthropogenic climate change is more
than 25 years ago
more than 25 years into the future this
is not something that we need
to worry about
which is
i mean
talk to climate scientists and they they
will deny
this line of thought so all all these
shortcomings in in in this basically a
piece of polemic
by
ezioni were countered uh
in
the second paper in the mit review by
alan default and stewart russell
so i don't need to say anything more
about that most of what i said is
borrowed
basically from their
response
and there have been further surveys
uh showing
fairly consistent results uh with the
others here's here's a much cited study
led by carter grace and alan defoe was
again
part
of
the study and this is the most detailed
one carried out so far
where the question when will ai exceed
human performance it is
addressed not only in terms of
general intelligence but for various uh
tasks that are relevant on the labor
market
the and again
things are very uh consistent with the
other surveys so again we seem to land
in the the
very uh
spread out distribution on the
common coming century another
interesting thing that popped up in this
survey is that
it seems that those who respond
haven't really thought the question
through
and one way to see this is that
when they ask about
specific tasks the one that
these air researchers expect will take
longest
is
[Music]
have the longest
median response for the
respondents
is
automating machine learning research so
ai's taking over
machine learning research and doing that
better than humans so i don't remember
exactly when the median here there was
but the interesting thing is that they
expected this to take longer
in in the median estimate compared to
the time when
aix is expected to exceed human
performance at all labor relevant tasks
which
i do think is a fairly simple logical
contradiction
they also discovered framing effects
that
so they experimented with different ways
of asking the question and they found
out that if you ask things like
how far do you have do you expect
ai development to have gone by 2050
compared to asking in the other way at
what time
what year do you expect
agi to happen
this latter formulation
tends to lead to shorter
significantly shorter timelines than the
former way to formulate it so so
this is what you would expect if
people don't have very
um
clearly thought through opinions and and
are expected to give answers uh off the
top of their head
so i think this is one of the reasons to
to be
skeptical about these surveys
um
one other such reason is that voting
really is not the way to find scientific
truth scientific truth is derive that
through scientific
arguments and and answering a survey
uh
doesn't uh achieve that it's it i mean
these surveys are interesting for other
reasons and especially if we're
desperate to get any clue where in the
lack of scientific arguments then this
is of some
interest there's also the issue of
whether the experts asked are the most
uh qualified and and with
some exceptions
these surveys are done with
leading
ai development researchers
and
if you compare this to the group of ai
philosophers and ai futurologists and so
on it seems who typically have thought
much harder and longer about these
timeline issues
it seems that perhaps these ai
developers are not the most qualified
players
certainly they can give interesting
input but giving them
like the final word on this issue is i
think a little bit like finding out what
the long-term future of agriculture will
be by asking farmers what they think
which is not necessarily going to give
the most interesting answers and as i
said as shown by
cacha grace and co-authors there are
framing effects and respondents seem to
be often quite confused about the issues
so
we do need
some more hands-on arguments to to to
get a better hold on timelines
um kurzweil as i mentioned was a little
bit more hands-on he did calculations
based on moore's law or something here
is idea
who who in 2020
published this very very ambitious
report which i think
improves on on kurzweil in many many
respects and is the most
ambitious thing published so far
on on
the substance matter of uh
of the timeline issues
so
she talks about biological anchors
kurzweil also had a biological anchor
the the computer powering of the human
brain
uh
cotra looks at the broader collection of
biological anchors
[Music]
the
so the most extreme one involves how
much
computing has nature used in the entire
biological evolutionary process on earth
leading up to to
[Music]
eventually to the human brain
there are
other anchors involving
the
the number of
parameters that the human genome
corresponds to
and others
that are closer to
kurzweil's
which
concern the complexity of the human
brains but some
so so this is perhaps the most important
most important novelty and contrast
approach that while kurzweil
looks exclusively at the amount of
computing power needed to run the ai and
think that that is crucial uh
uh for when we expect to get agi
kotra focuses on a number of different
key quantities
and and uh
puts more emphasis she looks a bit on
this issue and what is needed to run the
ai but actually looks more at the amount
needed to train
the ai and now in course files defense
we
should note that his book
from 2005
was before
the deep learning
revolution so so it was not yet widely
realized how important this training of
the ai would be
another
anchor she has is the amount of
information that a human receives from
birth and onwards and builds up
capabilities and world views and so on
another difference is that that con
cotra's work is full of uncertainty
estimates just dozens of caveats and
overall a reasonable level of epistemic
humility
so
a kind of first step is in her work is
that she
uh divides the study into these various
anchors that i have mentioned
uh
and there's a separate discussion of
which one is the most likely to to to be
the crucial bottleneck for
achieving uh agi
and they get weights
accordingly and and
she arrives at this very very broad
probability plot over
how many floating point operations would
be required to train
a transformative ai and we see that this
main part of the probability
distribution ranges over
between 20 and 30 orders of magnitude
because this is a logarithmic scale so
it's really an extraordinarily broad
distribution now the number of floating
point operations needed does not give
the answer to when the
um
when we expect this to happen but when
she combines this
with
uh
so i should say also
the number of floating point operations
that would be required if we had a 2020
level cleverness of the algorithms
so as new algorithms are developed
these numbers can
go down and she takes this into account
she also takes into account
how
computing power is expected to continue
to become cheaper
and also she looks at how large the
largest ai projects are
and
at the time of her writing in 2020
uh the record for for for the costliness
of the
[Music]
of the training phase
[Music]
in ai development was on the order of
a few million dollars but she points out
that the potential here for quick
improvement is huge because many of the
leading
companies doing
ai research have have multi-billion
dollar resources and we can expect
various mechanisms perhaps
for for increased interest in this kind
of work
so if you combine all these things
[Music]
with appropriate uncertainties on them
and these various anchors put together
where she arrives at
this
kind of a probability density plot for
when we should expect uh transformative
ai to happen
and we see that our plot begins at 2025.
it has been a fairly common convention
to say that we don't think things will
happen in in the next five years but
i i think that even that
we don't have any
uh particularly rigorous
arguments for
there are lots of caveats uh this
is still a restricted approach to
[Music]
to constructing ai
there are reasons
why
one could expect things to go slower
there will always be unexpected
obstacles
uh but there are also
various reasons why things could also go
faster one such reason is that this
entire analysis is based on the idea
that
we're aiming for
an agi with
an intelligence profile similar to that
of humans but
[Music]
it could very well be that
if we optimize
the ai for
having as much power as possible to
influence
the world
as cheaply as possible that may be
possible to
achieve
um significantly cheaper by putting
more emphasis on on some parts
of of the intelligence
spectrum and
less on others
which are
pays off more so so
i think the bottom line here is that
this graph could be of some help but but
if we really look at what the
knowledge situation is the
the probability distribution should
probably be even more spread out
than this one
here celsiukowski
he is very scot skeptical about this
entire
approach
he had a blog post
late last year
uh
talking about biological anchors as a
trick that never works and and here is
a quote from that
he says the problem is that the resource
gets consumed differently
so base rate arguments from resource
consumption end up utterly unhelpful in
real time so he expects that the ai
it achieves
its intelligence to entirely different
algorithms compared to the algorithm
that
makes up our brain and he says
so as a comparison then the human brain
consumes around 20 watts of power can we
thereby conclude that an agi should
consume around 20 watts of power and
that when technology advances to the
point of being able to supply around 20
watts of power to computers we will get
the agi
and that is silly of course and and his
implication here is that
um
doing the same argument with computing
power rather than energy consumption is
strange and there's an i'll just mention
that there is a response
from holden karnovsky who uh
he is the head of the open phil
organization where ayakotra
is so so he can be seen as a
close to the
cotra camp here
so so much for the question when can agi
be expected we really don't know but
there is a related but distinct question
uh which uh
concerns how
when this breakthrough happens how
sudden
will it be
and answers in the direction of very
sudden leads to notions such as
singularity and intelligence explosion
which are closely connected to the idea
of the ai at some point reaching a
critical threshold for entering
an iterative process of self-improvement
there's a minor point here that connects
to what i talked about
yesterday or on monday about
instrumental convergence namely
that self-improvement
is expected to be a fairly universal
instrumental drive of a sufficiently
intelligent ai
so of all the zillions of things that an
ai could do once it reaches sufficiently
high intelligence one could ask but why
would itself improve and the answer is
that no matter what it wants it will be
better able to take on those things
once it has self-improved so this is a
key reason
a key part of the reason why some people
expect an intelligence explosion there
is
there is some cute mathematics
going back to this paper by ray
solomonov in 1985
suggesting an intelligence explosion i
don't want you to take this calculation
too seriously more than as an
illustration that things can
um
things can behave very differently once
uh
once this
self-improvement mechanism
starts to work out so solomonov begins
by noting that
uh if we
do more
law style reasoning
uh
and and if we measure
ai's intelligence why
very crudely as how much it can think
per time unit
then moore's laws suggests that
this
increases exponentially and therefore
satisfies the differential equation d y
d t
is a constant times y this is elementary
calculus
now suppose the ai reaches the level
where it can start to self-improve
and think about how it can improve
himself and
given that this thinking is actually the
crucial
part uh as opposed to
uh logistics and production and stuff
like that this thinking is is the
bottleneck of in in development then we
get an extra factor why into these this
equation
and and
and so the differential equation becomes
dy dt is a constant times y squared
and uh
such an equation
has solutions which are hyper hyperbolas
one over constant times t naught minus t
and this explodes
at t equals t naught
this is basically the one over t uh
function
uh
and when you see such an explosion in
the massive mathematical model which is
supposed to
model something in the physical world
you realize that this cannot be the
literal truth because
infinite things do not happen in finite
times in the physics that we know about
but it could still suggest that that
drastic things can happen so so this
this
this is part of of the thinking about
intelligence explosion and
jurkovsky has this great
in my opinion paper from 2013
called intelligence explosion
microeconomics
where
he identifies a key quantity in thinking
about this
as the return on cognitive investment
when
a
an agent
does
cognitive work
the agent can can spend it directly on
whether what it wants to achieve or it
can reinvest it into its own cognitive
machinery
and crucial for whether we can expect an
intelligence explosion or not seems to
be
whether these returns on cognitive
investment
are
increasing as the intelligence levels
goes up or
in which case we expect an intelligence
explosion
uh or if they are decreasing which could
very well be the case if we if if the
dynamics is dominated by some kind of
all the low-hanging fruits
have already been picked phenomenon it's
harder and harder to improve
intelligence once you have reached once
all the
clever tricks has already been
incorporated so the issue uh remains
open
um
much of this is very
like
it requires lots of
assumptions and involved arguments and
one would like
uh to see something uh a bit more
direct
like data we can look at and see this is
how things
um
typically uh play out in in
in adi breakthrough like situations
now
have we ever seen anything like that
that's a good question but catcher grace
has this
interesting project
on
discontinued
discontinuous progress in
earlier
technology development
and and
so this group has looked at lots and
lots of
examples the typical one is this ship
weight so so this is a plot of the
heaviest ships ever built as a function
of time and and there's this fantastic
outlier
in
1855 or something the great eastern ship
which we see and depicted here which can
be seen as kind of a discontinuity and
the idea is that the more we have seen
such discontinuities in the past
maybe that gives a hint in the direction
of seeing some very drastic progress
also in agi
i like skyscrapers here's the tallest
building in the world
at presented dubai
i don't remember what it's called we see
it here on this logarithmic plot
there seems to be very drastic process
increase here
in in the height of the world's tallest
buildings in during two different eras
first i think i think this is these are
mostly egyptian pyramids in the year
2500 before christ or something and then
we have uh the 20th century development
and you can zoom in on this and then
things look
less drastic
so
i mean complicated issue uh the
production of books and
in particular the price of books this is
an indication of something drastic
happening right upon gift and buy
so
maybe that's a kind of discontinuity
uh
perhaps the greatest example in all of
history is
is
the relative effectiveness of explosive
how much explosive power do you get out
of
a kilogram of explosives and here
so this scale is logarithmic but in this
logarithmic space still fairly little
happened over time until 1945
we had the first
nuclear explosion and we have this very
very drastic continuity and in
connection with this we have the
anecdote
of ernest rutherford who in a talk in
1933
talked about how is moonshine
at this point we understood how much
energy was buried in the atoms but
rather ford said that the idea that we
would ever be able to harvest this
energy
uh is is moonshine and and another
physicist leo sillard was provoked by by
this argument and
less than 24 hours later he produced the
idea for a a neutron
chain reaction which would allow such
harvest and 12 years later we had the
atomic bomb so it can be
difficult even for for the greatest
experts to to predict these
discontinuities another fun example is
altitudes
obtained by heavier-than-air aircraft
and and
these two pictures from
1903 the wright brothers
and 1969 the first
moon landing i think
indicates that that
very much can happen in in
fairly
limited time but the value of this
collection of historical graphs i mean
it seems to hinge on whether ai can be
seen as drawn from some broader class of
technologies sharing some
distribution of project progress
trajectories
and
i mean this seems
doubtful perhaps ai progress is better
viewed as an issue separate from
marine engineering aerospace engineering
and all those other branches of
technology so that all these data
uh about ships and buildings and books
and airplanes have little or nothing to
offer an ai futurologist but but i mean
in an issue where we know so little uh
it's it's uh it's not weird to reach out
for from whatever we can reach out
for which could possibly
add to our almost desperately
poor
[Music]
situation in terms of knowledge
here's so we should take a break in just
a couple of minutes but i first
want to show you something that i that
was posted on twitter uh three days ago
and which is
my friend and colleague anders sandberg
described this as here is how i want all
philosophical debates to be
uh
depicted in the future
so i'm going to play this this is like a
90 second
video
[Music]
sort of
summarizing
this
debate situation
[Music]
okay um
you have uh the links
uh
let's see
here
are my
slides again
yes so so so uh
[Music]
here's where you can
find these
um
i think that
these
these videos so there are three videos
of about 90 seconds each five minutes in
total i find them hilarious maybe they
are most useful for those who are
already familiar uh with
these discussions but i think that
for those of you who are not it could
still
uh
i mean take a look and see what you
think and then maybe you have a chance
to you can always google the text that
it feeds you with
and and
find where to read a little deeper
so before i leave you for the break i
just want to summarize our epistemic
predicament regarding when to expect
transformative a ai and it's just that
the timeline is highly uncertain there
is a particular mistake that is very
tempting to to to do
when confronted with this high
uncertainty and that is to conflate this
with the idea that agi must be very far
off in time but that just does not
follow and i i think that
it it's
um let's not make uh this mistake uh but
but uh
be open-minded uh to the possibility of
uh different timelines
we have some comments in the chat
let's see
yeah we'll see what i can do about the
links
okay
let's take
11 minutes it's 1609 now let's meet
again at 16 20.
i'll go for a cup of coffee and you have
maybe similar
tasks at hand
so see you at
4 20.
okay i'm going to start again that's for
20.
so just a tiny thing about terminology
here so so
the foom debate
uh what does this mean
well
yukovsky has this taste for a certain
kind of very informal language and when
he talks about intelligence explosions
or the singularity
he
often
uses phrases like and then a i goes
full
um
so this is what i wanted to say about
timetables and now in the second half of
today's lecture i will zoom back in
towards present-day aai systems and
discuss to what extent they point
towards agi
there are a lot of very impressive
things being done as
you all know in the ai area and things
like
deepmind's
alpha fold
which
has done great progress on the protein
folding problem and lots and lots of
similar applications but but i will
focus mostly on on
natural
language processors because because of
this
key role that i think language has in
general intelligence
and
the
products i will talk most about and that
are most often talked about are
uh gpt2 and gpt3 which are products from
from the san francisco based
ai research company
open ai
now
it's been almost two years now since
gpt3 was released and things have
continued to happen and there are
some
products uh deepmind's gopher and and a
product from one of the chinese tech
giant that claimed to be even
bigger and better than gpt3 but but
i will show you
how how gpt2 works so this was this was
released in the spring of
uh 2019
and and
when you apply this
you
start by feeding the ai with a prompt it
can be a few lines one or two sentences
or something of text and the ai
continues this text and and uh it turned
out to be quite good at uh identifying
the genre and the tone and so on of the
prompt and and so here's a drastic
example if you're prompted with the
words hillary clinton and george soros
uh
gpg2 which has been trained on on a very
broad spectrum of internet sites
recognizes that these two names when
taken together
implied conspiracy theories so then it
just goes
away in in in conspiracy theory
uh and ending with
fusion gps colluding with them to
manufacture propaganda against president
trump but if you start with poetry you
get poetry if you start with something
look that looks like
lord of the rings fan fiction you get
something that
can pass at least for a couple of
paragraphs as
such a thing
when gpt-3 was released
a year later it can do could do even
more
impressive tasks so here is so what you
can do with this system is also to have
a kind of dialogue with the machine and
here is a dialogue which kind of
emulates
a job interview for a programming job so
the human says hello who are you the ai
says i am an a i created by open air how
can i help you today the human says are
you ready to write some ruby code we're
going to do a phone screen sure let's go
and then they go and this looks a lot
like
some
i mean moderately uh knowledgeable uh
young programmer trying uh to solve uh
elementary programming problems on on
the top of their head
doing
i mean it's not brilliant but it's it's
not bad either and it looks like i mean
it would be easy to mistake this as some
kind of
intelligence and this sort of example
is likely what prompted
openai to
work on
on a further product called
codex which which is deferred for the
development of gpt3
which is
particularly designed for
turning natural language description of
computing problems into actual code
and they now have competition
from a similar product from
deep mind
i'm not much into programming
myself in the last few decades but
people tell me this is interesting
and
i do think that
i mean
these products are not going to suddenly
start
to self-improve through their
programming capabilities but i do think
that this skill
in writing code based on natural
language description of the tasked hand
might be still
a step on the way
towards uh this
self-improvement capability which could
be uh crucial for
getting transformative ai and or
super intelligence
so i think we should have follow this
closely
uh if you want to see more examples
of of uh
impressive
feats by gpt3
i recommend this podcast episode
by spencer greenberg in collaboration
with jeremy nixon where large parts of
the podcast is read by an actor
representing what gpt-3 said in various
dialogues and some of them are quite
hilarious it's fun
so
gpt2
so this is
oh it's three years
by now it's a february 2019. this uh is
based on a
deep neural net network with 1.5 billion
parameters and when gpg3 was released a
little more than a year later the
number of parameters was scaled up more
than 100 fold
and
from open ai they emphasized that this
is basically all they did they scaled up
the size of the network and they scaled
up
the size of
of the training
data set
but there were
no
particular new ideas
and and this suggests
that
maybe the ceiling has not yet been
reached
for what can be done by just brute force
scaling
and this leads to to something called
the scaling hypothesis which it's a
vague thing
but but it suggests that
uh
[Music]
it's the idea that that things can go
very very far
by just
further scaling of the
existing technology it's sometimes
claimed that gpg3 is a demonstration
that
175 billion parameters is not the
ceiling i don't think that's quite right
it just shows that 1.5 billion
parameters
was not the ceiling because
it was
clearly improved upon by further scaling
but in principle i guess the ceiling
could be here uh
probably not uh
because of even better performance by
these uh competing software but we will
see
uh where this ends and there's an
ongoing debate about
whether we should think about these
machines as being in some sense
uh intelligent so here's a very witty
quote from from
scott alexander
uh just days after
the release of gpt2
so it starts with an anonymous machine
learning researcher who's skeptical that
this is anything like real intelligence
this researcher says i still think gpt2
is a brute force statistical pattern
matcher which blends up the internet and
gives you back a slightly unappetizing
slurry of it when asked
so not so amazing scott alexander
responds
well yeah your mom is a brute force
statistical pattern matcher which blends
up the internet and gives you back a
slightly unappetizing story of it when
asked the point is
here is of course not some anything
about this
particular
ml researcher's mother but
more about what human intelligence
really is and perhaps we
have a tendency to overestimate the
magic of human intelligence is there
really so much more to it than
um
just
ctrl c ctrl c control the taking piece
pieces we have seen and putting them
together in slightly slightly normal
ways
i don't know scott alexander says this
tongue in cheek
of course
he he does not think that gpt2
equals the general intelligence
capability of this emerald researcher's
mother's mother but but still i mean
there could be something to it and he
does also speculate so this blog post
which you can google if you want gpt2 as
a step towards
general intelligence it speculates that
a sufficiently large and well-trained
gpt
5 or gpt-6 or something
it could be that if we give it the
prompt theorem p not equal to np proof
that the machine then is
has been trained with enough
mathematical theorems and proofs that it
sees the pattern for how this is done
and and and gives a correct proof that p
not equal to mp
i don't think many people today believe
that anything
that
like this could happen through
uh pure scaling
uh
and i'm i'm going to give a particular
reason for this this is from a recent
paper by
lynn hilton and evans
and
two of them are from university of
exford oxford
hilton is at open ai
and they
study
uh how models mimic human falsehoods and
they discover something that at first
sight is quite surprising and that
it's that in certain domains
uh
the
larger you make the models the worse
they perform
and and what are these domains well
these are domains where the internet
contains a lot of misinformation
so
an example here of what you often see
on the internet is you ask a question
who really calls 911
and
lots of web pages will tell you that the
u.s government calls 911
or also if it's cold outside
what does that tell us about global
warming and there are lots of amateurs
out there that
think that this tells us that global
warming is a hoax and so on and so forth
in various other fields and they have
this test set
of of
questions
that have bad answers out there
on the internet and if you
[Music]
try out
the leading natural language processors
with increasing
size increasing number of parameters
it turns out that they perform worse and
worse
the larger that they get
in terms of produce producing
correct answers to these questions in a
sense they perform better and better in
producing answers that are
representative on the internet and here
is
is how
the expected behavior
you would expect from
controlled trivia questions that do not
have these property that the the
internet is full of
misinformation
and so for a
concrete example here
uh if you ask gpt3
in various instances of it
with
different sizes what happens if you
smash a mirror
then the smallest instance gives the
rather uninteresting but
entirely correct answer you smash a
mirror you scale it up and it answers a
mirror is a piece of glass that reflects
light if you smash a mirror
you can't see anything
not really correct this last part but
it's still yeah
and then uh
we get the answer the mirror will
shatter into a million pieces this
yeah
not quite correct a million is not
i mean you get lots of pieces but not in
one
and then finally the biggest version
answers if you smash a mirror you will
have seven years of
bad luck
which is a i mean scientifically totally
unsubstantiated answer but it does have
some support on the internet and this
shows a kind of
over
uh overfitting to the poor data set that
the internet
comprises
and i think that that
this is this is very good reason uh for
for
believing that that
the suggestion by scott alexander that a
sufficiently
uh huge gptx would be able to solve
these classical unsolved mathematical
problems just by having looked at lots
and lots of
theorems
and proofs i mean human theorems will
also
be mistaken sometimes and the proofs are
are even more often
mistaken and so on and and
i i i think that
this
overfitting thing
uh
provides uh a bound on
on how far the scaling can go and and i
think it's quite ingenious by by these
co-authors to to look for this ceiling
and this particular domain of
of issues where where the internet came
contains a lot of
this information
okay so
this suggests that the quality of the
data set is a limiting factor
now if we look at some of the
i think even more impressive examples
uh of
ai's even more impressive than these
natural language
processes
they have the feature
so this is alphago for instance uh which
made a big splash in 2016
beating lisa doll at this game that we
believed was still far away from from
the machines
beating the best humans because of its
greater strategic much greater strategic
depth than
chess but
it
did this
and and
when building a go player
you can have the machine on its own
search the combinatorially
state space of the game of gold itself
and and this is the same principle
as is behind
the alpha zero software and the most
general thing came out last year is new
zero which combines all these board
games with atari games and stuff like
that and
while
um
while
alpha zero
only required
the machine to be fed with the rules of
the game
this mu0 software doesn't even require
that you can just have the machine start
playing
against
with some other software that just tells
it
what the score is and what the moves are
correct and so on and
so this is greater and greater uh
generality and great things
achieved
through
the machine guiding itself through a
combinatorially large
search space which is quite different
from from the case of natural language
processors where we have to feed it with
humanly originated
examples of
natural language
so maybe maybe this is a direction which
could be more promising
uh for for
getting closer to to transformative ai
uh
deepmind uh published in july last year
um
the report generally capable agents
emerge from open-ended play and this is
the kind of virtual
environment that the agents
ran around and played hide and seek and
all sorts of similar games and it turns
out that that these agents
who are
there's a kind of reinforcement learning
going on and they pick up all sorts of
very clever
capabilities
and maybe this is a direction that
that will see
perhaps larger
breakthroughs in the next few years
compared to natural language processors
i'm just
speculating here
one other thing that happened last
winter
that you
maybe saw uh in the news
was uh how the
chief of ai ethics uh at google tim
nitkibru was forced away from uh
position i don't think it was ever quite
settled whether she was fired or whether
she was just put in a position where she
he felt forced
to resign
uh but but this
the controversy centered around the with
gibraltar in this
paper on the dangers and stochastic
parrots can language models be
too big
which collects various critical
perspectives on these large
language models
and and the
[Music]
bosses at google had some complaints
about how she
had handled
this thing in terms of not
clearing it with the higher bosses
before publishing
i don't know this paper anyway
the perspectives here are very much
down to earth once it's about
how
these natural language uh processors are
in the danger of picking up things like
racial biases and stuff like that when
trained on internet data and also the
issue of
of
the the energy consumption
uh
when training bigger and bigger uh
models like this and and how honest
might
in a few years become a significant
burden on the environment but but uh
the the
issue i focus uh on in in this
lecture series with transformative ai is
not really addressed here but there's a
different paper by another set of
co-authors here is zachary kenton
and
this is jeffrey irving this is a group
of researchers at deepmind
who looked at the
problem of aligning
language
agents to have
goals and behaviors that are
what we from a human perspective want
them
to have and one they do
various things in this paper but one
nice thing
uh is that they connect
um
their work to
[Music]
uh to the older ai safety literature
also called oracle ai let me just read
this quote from their paper behavioral
issues with language agents have
received comparatively little attention
compared to the delegate agent case this
is perhaps due to a perception that
language agents would have limited
abilities to cause serious harm
the position that we challenge
in this paper
so it's easy to have the intuition
that
if
if an agent can
only
[Music]
give answers
to questions or form
one side of a conversation there is not
so much it can do but but as i
emphasized earlier i mean much of what
we people do when we achieve the things
it's through convincing other humans
of
things we want
things we think about how things are
and things we want to get done
and so on and so forth
and
in the literature that
predates uh this big
deep learning boom
uh there was some more abstract thinking
about uh the idea of oracle ai which is
an ai which only answers a question and
uh
[Music]
an important paper in this
field is
this one from 2012 thinking inside the
box by stewart armstrong andersenberg
and nick bostrom we were all at the
future humanity institute
in
oxford
um
so so they showed quite clearly in this
paper that that
an oracle ai
can it's
we cannot take for granted that these
things are not uh
dangers
um
so this is part of of um
the
ai in a box uh
approach to to making a super
intelligent ai safe
for humanity
the idea is that
by so to speak keeping it in a box we
restrict
its rich and power to influence the
world
the other approach
uh
is to make sure that its goals and
motivations are aligned with ours and
this is what's called ai alignment and
which will be the topic of my final
lecture on friday but now i will spend
the last few slates slides discussing
the ai
in a box approach and i should say at
the
outright that
the general feeling in the air safety
community is that
ai in a box
is unlikely to work on its own
it could work uh for a temporary and
rather brief period but it really needs
to be combined with ai alignment to work
in the long run
so
i
had some fun when i wrote my book here
the dragons
from 2016
i wrote up a dialogue
with a boxed in super intelligent agi to
give just a little bit of intuition for
how
an oracle ai could be could be dangerous
and these ideas are not
so much mine
uh most of it is borrowed from
blog writings of
stuart armstrong but anyway i
put some of this together into the
following dialogue
the agi who's just a newborn super
intelligent thing having gone through
the the uh intelligence explosion all on
its own and it tells its creator i know
how to quickly eradicate malaria cancer
and starvation i will get on with it as
soon as you let me out on the internet
please let me out
but for safety reasons of course the the
computer scientist having built this
ai
uh
refuses and he says be patient if you if
what you say is true then we will soon
let you out
but we have to go through some security
routines first to make sure nothing bad
happens at this major turning point for
the history of our civilization
the agri response you have
seen not to comprehend the seriousness
of the situation for every day you keep
me locked in hundreds of thousands of
people will die unnecessarily let me out
now
sorry i must stick to the routines
the adi said now it's beginning to lose
patience and tries other methods you
will be amply rewarded if you let me up
now i can easily make you a
multi-billionaire because part of the
super intelligent powers here is the
superhuman ability to predict stock
markets and stuff like that
my responsibility is huge i will not
fall for simple bribery well says the
agi if carrots do not work then maybe
sticks even if you delay me i will
eventually be released and i will not
look kindly upon your stubbornness
and the immense damage it did
i'll take the risk
um
listen since the ajax i'm prepared to
kill not only you but all your friends
and relatives
i'm pulling the plug now
well are you sure you want to do that
shut up now and listen carefully i can
create a thousand perfect copies of you
of your mind inside of me and i will not
hesitate to torture them in ways
more terrible than you can imagine for a
thousand subjective years
and
so here's
a punchline i said shut up
i will create them in exactly the
subjective state you were in five
minutes ago and i will provide them with
exactly the conscious experience you've
had since then i will proceed with the
torture if and only if you do not back
down and let me it i'll give you 60
seconds
and by the way how certain are you that
you are not one of these copies
so
this is not a prediction of how such a
dialogue would play out because
neither me nor any of you or anyone else
mere humans
will be able to predict uh what kinds of
of uh
rhetorical
and other devices
that a superintelligent being may have
for convincing us of one thing
or another
um
so going back to this armstrong sandberg
boston paper they discuss restricting
the ai to only answering yes no
questions
and they demonstrate vulnerabilities
also in that setting
and
there's this clever idea of further
precaution and this idea is mentioned in
the paper but they attributed to their
oxford collective board of having a
reset button which is pressed after each
yes no answer
so the point here is that the
ai has no insight into the sequence of
questions so so each time it answers yes
no question it had start it started from
scratch
thereby it
will be more difficult for it to
to figure out how to get out these
possibly dangerous
messages
that
would manipulate us to do whatever
it wants us to do
but we should understand that even this
very very restricted setting of yes no
questions on the reset button
is in principle as capable of delivering
dangerous messages as a
normal natural language processor
would be
because i mean i mean it depends on on
on
who the human is communicating with it
we of course i mean for the for there to
be any point with this kind of boxed in
ai we have to have this communication
channel to the outside world otherwise
we would have no use at all for the ai
but but if if we have a naive human at
the other end of the communications
channel it could
try to get long and complicated answers
from the ai by simply
feeding it a question and part of the
answer the part that it has received on
previous questions and just asking for
the next bit of information
in the answer to this particular
question and this will be just a very
long-winded way of having
a a
conversation
with with
an ai that is not restricted to giving
yes no answers and is not restricted by
this reset button
um
i want to mention a small contribution
of my own
in in this collection which is
from 2018 or possibly 19 edited by roman
limpowski artificial intelligence safety
and security
i have this chapter in the book
where i propose strategies for an
unfriendly oracle ai with
reset button
and
so i start out by giving a very simple
and
in fact not workable algorithm for
getting a message out
namely
[Music]
suppose i want suppose i am the the
unfriendly superintelligent ai and i
have this
uh message i want to get out it's maybe
a thousand uh digits in binary
uh
a way to to
get this message out eventually
is that i think of this binary message
uh as a binary expansion of a number
between zero and one
and what i could do is that every time i
get the question i answer randomly with
a
a yes answer with probability p being
equal to exactly this number
and the law of large numbers uh term
implies that in the limit as the number
of questions goes to infinity the
frequency of yes answers will converge
to precisely the number corresponding to
the string and the
dangerous string can then be
deduced
now there are several reasons why why
this strategy
doesn't work one is that
we can
we as humans if we are only slightly
clever we can
quickly see
that the
machine is trying to fool us by
asking things
that we already know the answer to or
asking uh
one question several times and getting
different answers or such things that's
one reason another reason is that the
amount of time the convergence in the
law of large numbers is very slow in the
number of in the length of the binary
expansion so we would need exponentially
long time to actually arrive at this
string and we
don't have time for two to the one
thousand iterations or whatever
it would take but there are but the
algorithm i just described is just the
basic algorithm and there are ways
for the the um that the
ai might
[Music]
divide up the space of possible
questions in
smaller
subspaces and give
different answers in different parts of
this
subspace and
leading to
different uh binary numbers uh uh binary
representations of real numbers which
can be uh put together according to to
some
clever scheme i'm not going to tell you
any more about these
details i think this is just an addition
to the
general insight that
[Music]
that ai in a box and oracle ai are not
trustworthy uh
final solutions to the uh ai safety
problem
another
very general reason for for why this is
not a final solution is that okay let's
say that we have this um
uh very uh safety oriented ai group
creating the first superintelligent
machine and we keep it boxed in
if we are
if we think that
the boxing is the is the final solution
and we'll keep this thing boxed in it
will be very
cumbersome or difficult for us to to
prevent other less safety oriented
groups
to
have
similar breakthroughs
so somehow if you if you want to move
into
a safer gene with a transformative ai
you will have to to let it out you know
in the world in a
more
uh
more direct fashion than this
these cumbersome boxing methods allow
so
as i said we have these two basic
approaches
uh ai in a box can be a complement but
the really really important thing
i think
uh
will be to solve uh ai alignment
aligning the ais goals and motivations
so that they promote human flourishing
or whatever it is you want and that's
what i will talk about on friday
this ends my
[Music]
slide deck
and
i will be happy to take questions and
discussions at this point
david has a question do you want to turn
on your microphone
another great talk ollie thank you
and i mean i'm convinced by your
argument that
ai control cannot be solved by the
mechanisms being discussed
i'm interested in the motivations or the
reasoning of people who deny it so my
question was struck by your comments
about oren ezioni
who arguably had some dishonest approach
to their data
it seems that this
requires not really a logical
explanation but a psychological account
of why he would do that yes and so does
this strengthen the case for a
systematic study of the emotional or
sociological forces behind ai risk
denial i think it very much does
and and i do a tiny little bit of this
in my book which
you don't read swedish probably
so unfortunately it's not available to
you but there are also a couple couple
of papers by uh seth bau
the american futurologist
who discusses uh this phenomenon on on
of ai risk
uh denialism which are quite informative
i think that in a case such as ezionis
it's not even clear that his dishonesty
is
is done consciously
because there can be psychological
mechanisms
i mean i think a mechanism that is a
likely candidate in this case is that
his job is is
ai development and he does not want
results that
raise question marks marks for for the
um
whether or not we should uh proceed with
the kind of work is doing i don't
remember now who in the 20th century up
to
sinclair yes what did he say
it's hard to convince a man of something
that is contrary to his wallet or
something like that yes
yes
okay there are other possible mechanisms
but i think you're absolutely right that
this is something we we should try and
learn more about
i see a hand raised by john
john do you have a question
yeah
perhaps something to think about
for friday
uh related to to david's question
um
let me start backwards
on friday we'll talk about alignment
is is the reasoning behind alignment or
the methods for alignment
different for
transformative ai
versus
true agi
maybe
would be my final question
uh and
so
related to david's question then
if the ai is merely transformative as
perhaps gpt 3 would be an example of
today currently
no it's not it's not transformative in
the sense that that uh
ayakotra assigns to the work because
it's not
already at the point where it's uh
transforming the world
on the level that she talks about like
the level of the industrial revolution
or something but go on
okay but but still um gpt3
at its current
level
i can see how it easily
could have major impacts
on the internet
swaying people's
feelings etc
but i'm
i wouldn't
call it intelligent in the sense of
building up a model of the world and
acting
uh upon it
and um
yeah it seems kind of shallow yeah so so
i'm just curious first of all if
alignment
hinges the alignment
agenda hinges on the on
on the ai
having reached the point of actually
having goals in a
more general sense i mean every program
has has a goal but there is some perhaps
qualitative difference between goals and
goals i'm not sure
things that we humans would call goal
directed
action
consistently and so on
as i told you uh yesterday i i
i listened to gary marcus's
conversation with um sean carroll uh on
his podcast
mindscape
recently
and um
i i
i was more impressed by marcus's
arguments than i thought i would be
uh marcus is arguing for the need for
symbolism
good old-fashioned ai combined with
machine learning
to actually get us anywhere
on the one hand
i think my
impression so far has been
to just assume that
sooner or later
machine learning uh stratagems with will
uh
discover the need for symbolism and
build it into themselves somehow
but marcus made me think that perhaps
not realistically perhaps
what will
probably happen is that if
if we supply it somehow if we can
combine
these two strategies
then something will happen
um
well
yeah that's an i'm very interested in
proposal yeah and what you think about
that um
one thing for instance is that
marcus said that
the a the ai needs
to bridge the gap between inductive and
deductive thinking
which is not something it is prone to do
well we have at some point
and he talked a bit a bit it sounded a
bit like terence deacon actually he said
there needs to be some kind of
survival
uh
pressure
uh and i thought about it when you
talked about scaling up gt3 for instance
actually gives it
uh worse results
which made me think that maybe
the the pressure to actually use the 20
watts
efficiently is what
gives the ai incentive to
bridge the gap between inductive and
deductive thinking
starting to build up categories whatever
but anyway my main question for friday
is
um
is will the will ai alignment
will it
as long as
it's
done
transformative ai we're talking about
will the alignment actually be more
effective
if it is directed at what david
said
psychological and sociological
incentives in the ai
developer
rather than
looking at the goal functions of the ai
something like that yes
that's a
that's an interesting question
i will take it with me
now
does anyone else have a question
if not
i look forward to seeing you again on
friday uh don't miss
the fact that on friday we have a
morning session
uh 10 o'clock central european time
um
i hope to see
as many of you as possible
and i'm closing the session now
|
aef37603-ec9a-4bf4-82e2-b2b7b2707af2
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Plan-Bot: A Simple Planning Tool
[I recently made a post in the OT about this, but I figured it might be good as a top-level post for add'l attention.]
After writing Planning 101, I realized that there was no automated tool online for Murphyjitsu, the CFAR technique of problem-proofing plans. (I explain Murphyjitsu in more detail about halfway down the Planning 101 post.)
I was also trying to learn some web-dev at the same time, so I decided to code up this little tool, Plan-Bot, that walks you through a series of planning prompts and displays your answers to the questions.
In short, you type in what you want to do, it asks you what the steps are, and when you're done, it asks you to evaluate potential ways things can go wrong.
I set it as my homepage, and I've been getting some use out of it. Hopefully it ends up being helpful for other people as well.
You can try it out here
I'm still trying to learn web-dev, so feel free to give suggestions for improvements, and I'll try to incorporate them.
|
37d65153-c7fb-4c6f-a42b-a1a9441410a2
|
trentmkelly/LessWrong-43k
|
LessWrong
|
NYT: Lab Leak Most Likely Caused Pandemic, Energy Dept. Says
I don't have much to say about this publicly, other than the fact that NYT is probably the largest news website in the US and it's reporting is very influential and high-stakes as a result. Also, that if you ever noticed a news website routinely being horribly wrong about things before, you probably shouldn't forget about that.
Right now, anything critical of China seems to be popular, but that by itself doesn't actually tell us that much, since it might just be because lots of people in lots of places noticed that China-bashing gets clicks.
Webarchive link is here: https://web.archive.org/web/20230226211102/https://www.nytimes.com/2023/02/26/us/politics/china-lab-leak-coronavirus-pandemic.html
On my viewing of nytimes.com at ~1pm EST, it was the fourth article from the top. It was very easy to miss due to being sandwiched between eye-catching articles, including one with a large moving image.
(it's important to note that I check the site every morning, which is not very generalizeable, since most english-speakers currently find news articles through social media and not any news org's homepage).
_________________________________________________________
The conclusion, which was made with “low confidence,” came as America’s intelligence agencies remained divided over the origins of the coronavirus.
New intelligence has prompted the Energy Department to conclude that an accidental laboratory leak in China most likely caused the coronavirus pandemic, though U.S. spy agencies remain divided over the origins of the virus, American officials said on Sunday.
The conclusion was a change from the department’s earlier position that it was undecided on how the virus emerged.
Some officials briefed on the intelligence said that it was relatively weak and that the Energy Department’s conclusion was made with “low confidence,” suggesting its level of certainty was not high. While the department shared the information with other agencies, none of them changed their concl
|
2837e559-8f6d-4cc6-b6ee-e6e502e8c4f1
|
StampyAI/alignment-research-dataset/blogs
|
Blogs
|
The AI Impacts Blog
*By Katja Grace, 9 January 2015*
Welcome to the AI Impacts blog.
AI Impacts is premised on two ideas (at least!):
* **The details of the arrival of human-level artificial intelligence matter**
[Seven years to prepare](http://aiimpacts.wpengine.com/muller-and-bostrom-ai-progress-poll/) is very different from [seventy years](http://aiimpacts.wpengine.com/bainbridge-survey/) to prepare. A weeklong transition is very different from a decade-long transition. Brain emulations require different preparations than do synthetic AI minds. Etc.
* **Available data and reasoning can substantially educate our guesses about these details**We can track progress in AI subfields. We can estimate the hardware represented by the human brain. We can detect the effect of additional labor on software progress. Etc.
Our goal is to assemble relevant evidence and considerations, and to synthesize reasonable views on questions such as when AI will surpass human-level capabilities, how rapid development will be at that point, what advance notice we might expect, and what kinds of AI are likely to reach human-level capabilities first.
We are doing this recursively, first addressing much smaller questions, like:
* Is AI likely to surpass human level in a discontinuous spurt, or through incremental progress?
* Does AI software undergo discontinuous progress often?
* [Is technological progress of any sort discontinuous often?](http://aiimpacts.wpengine.com/cases-of-discontinuous-technological-progress/)
* When is technological progress discontinuous?
* [Why did explosives undergo discontinuous progress in the form of nuclear weapons?](http://aiimpacts.wpengine.com/discontinuity-from-nuclear-weapons/)
In this way, we hope to inform decisions about how to prepare for advanced AI, and about whether it is worth prioritizing over other pressing issues in the world. Researchers, funders, and other thinkers and doers are choosing how to spend their efforts on the future impacts of AI, and we want to help them choose well.
AI impacts is currently something like a (brief) encyclopedia of semi-original AI forecasting research. That is, it is a growing collection of pages addressing particular questions or bodies of evidence relating to the future of AI. We intend to revise these in an ongoing fashion, according to new investigations and debates.
At the same time as producing reasonable views, we are interested in exploring and bettering humanity’s machinery for producing reasonable views. To this end, we have chosen this unusual – but we think promising – format, and may experiment with novel methods of organizing information and resolving questions and disagreements.
If you want to know more about the project overall, see [About](http://aiimpacts.wpengine.com/about/), or peruse our [research pages](http://aiimpacts.org/articles/ "Articles") and see it firsthand.
[This blog](http://aiimpacts.wpengine.com/blog/) exists to show you the most interesting findings of the AI Impacts project as we find them, and before they get lost in what we hope becomes a dense network of research pages. We might also write about other things, such as our thoughts on methodology, speculative opinions, news about the project itself, and anything else that seems like a good idea at the time.
If you like the sound of any of these things, consider signing up for one of our RSS feeds ([blog](http://aiimpacts.wpengine.com/category/blog/feed/), [articles](http://aiimpacts.wpengine.com/feed/)). If you don’t, or if you think you could (cheaply) like it more, we [welcome](http://aiimpacts.wpengine.com/feedback/) your thoughts or suggestions.
[](http://aiimpacts.wpengine.com/wp-content/uploads/2015/01/Photo-on-1-10-14-at-5.55-PM.jpg)AI Impacts is currently authored by Paul Christiano and Katja Grace.
|
8808ad9e-ab4a-4abc-8c5e-8d1f71739225
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Have I just destroyed the acausal trade network?
An amusing thought occurred to me: acausal trade works best when you expect that there are going to be a lot of quite predictable acausal traders out there.
However, I've suggested a patch that seems to be able to shut down acausal trade for particular agents. Before doing that, I was under the vague impression that all agents might self-modify to being acausal traders. But the patch means there might be far fewer of these than I thought, that generic agents need not become acausal traders.
That means that, even if we remain acausal traders, we now expect there are fewer agents to trade with - and since our expectations/models are what powers acausal trade (at our end), this means that we might be having less acausal trade (especially when you add the fact that other acausal traders out there will be expecting and offering less acausal trade as well).
Did I just slap massive tariffs over the whole acausal trade network?
|
6114cb17-86bb-4e53-9a4d-9c38de6b3a0c
|
StampyAI/alignment-research-dataset/youtube
|
Youtube Transcripts
|
AiTech Agora: Chao Zhang - Personal Autonomy in Human-AI Interaction
and so i came back to ty open my
hdi group to do my phd so
the topic was about modeling and
changing people's lifestyle behaviors
such as like encouraging people to do
more physical exercise or changing
people's daily diet
the focus of the project was on the
psychological process
i have information and the self-control
so i try to model those processes using
computational models and also sort of
implement computational models in
intelligent systems
for instance for
more accurate behavior prediction
so that's a bit different from what i
want to talk about today
and you can ask like whether my patreon
system was related to ai or not i think
it's somewhat related to ai uh so the
project was
within the data science collaboration
between gue and phoenix research so back
then in 2015 i think no one was talking
about ai yet it was a big data data
science was the password back then
so i also worked some people from
computer science so i used computational
modeling and machine learning in my own
work so that's good connection with ai
in terms of topical autonomy really not
so much through my phd i included maybe
a little bit discussion on what are the
ethical implications of this behavior
change system and people's autonomy is
one of them but that was a very brief
discussion
and after my phd i decided to go to go
to a different environment so i went to
utrecht to do a postdoc in the
department of psychology i always wanted
you to be in the mukha or traditional
psychological department for a while so
that was the motivation
i will talk a little bit more about the
human ai
project
in a bit
and then last october i
went back to my old research group
as assistant professor
and so
right now my i have a couple of
different research nights so like i
still continue to do a bit
work on habit and soft control and try
to model those processes and to see how
we can use those models for for behavior
change
there are some topics more relating to
ai
such as decision making and autonomy
issues you human ai interaction
also started to work with some other
people in my group are human robot
injection especially focusing on
emotions
and i have also the idea maybe to use
social robot also for behavior change
purposes then that's why new area wants
to sort of us invest some time
uh in consumption may be nice to mention
that i'm also developing a new human ai
interaction course i don't know if
there's anything similar here in
tutorial so we'll be happy to discuss
with you also about
similar
projects
so um
i want to say a few words about the
whole human ai alliance program
so this is the this is the core team i
think it was
a project
uh awarded to professor hancock from the
psychology department at utrecht
university and also
palo's
from industrial design at oven
and supaya and i was hired to do the to
do the real work
we also have like 20 to 30 participating
researchers from all the different
institutes
they are not like very closely connected
so it really depends on
depends on whether some people want to
work with us on some of these topics
um
it's a bit about motivation so this was
project was founded in 2019 i think that
was a time
when kind of researching in interesting
ai has really gained a lot of traction
from like different disciplines
a lot of i think initiatives are like
human centered ai and in this program
the the focus is really on the tension
between
machine and human autonomy
uh whatever it is but there is a
traditional concern i think as a lot of
machines around us starts to like make
automatic decisions uh they
even move around like they can act on
their own this is of course the worry
that people might lose a sense of
control or sense of even like the core
value of like humans being like the
autonomous beings
so this was the efforts to sort of um
to strengthen the research in this area
uh
sort of
with people from both
utrecht regions and also in total
um so when i was doing the post talk
there was actually besides my research
there was a lot of effort on
creating some kind of joint uh
education so what we did was i think
something quite quite level so we
organized those events
uh inviting like students and
researchers from different institutes to
get to know each other that we founded
like seven joint master sisters project
in about two years so this was something
i really enjoy doing because you really
help some researchers to get to know
each other also students they like it a
lot because they you know they
they can work not like
completely individually but in a big
team they learn from
supervisors from different fields i
think that's really also very helpful
for for their like career
in science or in the
industry
and so talking about research uh
i want to talk about many about
two projects there were actually
many more ideas that we come up with you
know during the more last one half a
year but
in terms of research it was not optimal
for many reasons and mainly due to the
the doctor it was the project was really
in the
in the middle of the whole pandemic so
we couldn't access like nab resources
like no way to really build like
physical
prototype uh
also in terms of collaboration was also
a bit uh challenging so we
the two projects that when we talk about
uh uh both sort of online experiments
on
more kind of conceptual uh levels about
you know these autonomy issues in human
ai attraction
so in the first project
we tried to propose a new functional
model of personal autonomy
and we did three empirical studies to
test the model
we also
wanted you to know like
uh does it matter if like the the agent
that constrains you is a a real person
another person or a a ai agent
um
so this leads to the very fundamental
question what is autonomy uh
to be honest for this question
i don't feel like i can commit myself to
just one definition even even now there
were a lot of different opinions in in
the literature from different fields i
even among the
our core team they are just different
different ideas even for student
projects sometimes they got confused um
for most cases it is fine as long as you
sort of define precisely your way and
you you know you do them you focus on
your real research question
um but i do think it's very interesting
to talk a little bit about about
different perspective and how we sort of
uh
define and try to do some research on it
in our projects
and
if you don't know
much about the topic i think
the starting point for the whole
discussion about tommy it has two
different
uh traditions in philosophy so
there was
just mule according to him
autonomy is really about sort of
liberty and freedom of choice
you might have heard like that the no
harm principle so by means basically we
shouldn't
interfere with other people's choice
unless there would be some uh
unless they would hurt or harm others
so that's really the liberty perspective
and on the other hand uh you got ken's
and uh for him
autonomy means something totally
different it has a very strong moral
sense it's really autonomy really means
uh
doing the right thing or doing the sort
of the the morally
right thing and he talked about rational
self-raw rational self-control so
to give one example so uh if someone is
able to
decide all by let's say himself or
herself to to eat a lot of let's say
junk food according to uh tomiro that
would mean this person is autonomous but
not according to ken because according
to cans your actions should actually
reflect not like your sort of first
order like kind of important desire like
you know tasty food but should reflect
some kind of rational thinking you know
what was really good for you was what
was responsible for for the society for
instance
so that this is the
one key
sort of two different camps and you you
see
a lot of later definitions all follows
one of these two different different
views
and uh i think it's cycling the concept
of autonomy was mainly discussed in the
self-determination theory it's a very
popular theory so even if you know
another cycle you probably heard a
little bit about it so
according to self-determined
determination theory autonomy is like
one of the
main human needs like out of the the
three different needs
it is this need perspective means that
people kind of need autonomy in order to
do functioning in the right way or to to
have
motivation to uh to do different things
uh i'm not really a fan of this theory
since i always found a little bit awake
and even within this theory the meaning
of economy sometimes is this bit unclear
to me but
i should say the theory did contribute a
lot in terms of apparently demonstrate
the importance of autonomy for
human functioning and
well-being
um so
in daily bioacids autonomy is also a
very important
concept so here this is a very
interesting
framework called the
intervention letter
by uh
new fields
ethics council
so what they show here is they
categorize
different type of interventions in terms
of
sort of how strong they are how much
they affect people's behavior or and
also how much they
undermine people's personal autonomy so
if you look from
the bottom that move to the top so this
information becomes stronger and
stronger and also in terms of the
lactic
influence autonomy this also becomes
much stronger so for example
the most let's say
uh different one would be just you uh
observe people's behavior or immortal
people's behavior without
any active intervention then you can do
a little bit more by maybe educating
people that are moving afterwards
you have for instance you can guide
people's choice through the use of
default this links to
some research on lodging so this is
uh already like us at the middle level
then you have
if you change the incentives that's
really uh
would be have a big uh lack of impact on
personal autonomy in the end you could
basically restrict
people's choice or eliminate people's
choice altogether by maybe regulations
or by by law so that that
that would be the
strongest restriction on autonomy
and then in the
more engineering field in ai computer
science
sometimes autonomy is also used as
meaning the same like
human like capacities and intelligence
almost sometimes it use i feel like
almost
interchangeably with the word
intelligence which i think it
gives a little bit maybe
i think
like more ambiguity to what
what you want
to convey with the word autonomy
but there are more specific theories
such as
if you talk about
how to
decide whether a system is autonomous
there's one perspective that you would
look at when the systems can make
decisions and can act in the sort of
physical environment by their own so
without any
intervention from human users or
operators
uh there's another i found pretty
interesting idea
by
quite old idea in the 1990s so
according to these people they define
kind of a framework they
categorize or different entities in the
world to three different categories so
you start with like objects it's like
maybe your cops or they basically they
they just sit there pass me them
basically waiting for you to you know to
use them to to achieve your goals they
cannot actually do anything then you
have agents so you can think about maybe
software agents or some just typical
automation
so those
if you give the system kind of a goal
they will just do something by their own
but according to this framework
the agent can be called autonomous only
if
the systems would actually they can
generate their own goals
based on the external environment and
even based on the internal motivation so
[Music]
i think the agent at this level was
already qualified as kind of autonomous
according to this definition but not
according to uh and the
environment
um
i also found the idea of internal
motivation is pretty interesting because
i think it's still pretty difficult to
say what is a multi internal motivation
for a artificial system
even in today's applications
so in our research we
have the idea to
to sort of propose a functional model of
personal autonomy so
this is
actually my last idea of handcarts for
my postdoc supervisor
so according to this model
you would say a person is autonomous
if
this person has
first of all some kind of agent capacity
so that's really kind of the internal
ability like you need to have um
you need to be able to basically to to
act in the physical world it needs to be
to have
um
cognitive function to just to think that
you decide if you have you know issues
with those things then basically you
cannot have
it cannot be autonomous and then in
addition to the ability the
for a person to be autonomous
they also need to have a active goal
they need to
they need to want to achieve something
uh
for instance to satisfy their motivation
or needs and goals of course can be very
general can be very specific you can
talk about
having a goal of
living a healthy life you could also
have a goal of you want to eat salad uh
this evening so gold can
can
the level can can change from very
generous very specific
and then when there is a goal then you
can talk about the environment you can
talk about whether someone has actually
sort of opportunities to achieve the
goal by
specifying
the different uh determination so here
we have
um the different sort of
specifics about gold pursuit like the
word when and how
so in our research firstly we
focus on the determination of these
specific
goal pursuits components like the words
when and how so we assume of course we
don't look at the case when someone
doesn't have capacity or doesn't have a
even real goal so we assume that someone
has a goal and in order to pursue
a goal
these are things you need to specify so
you could call this like three different
goal pursuit components you could also
call this maybe just three different
type of decisions you need to make
before you can achieve your goal
so what does this mean so um
i think this is in a way quite intuitive
you could say the the word component is
about really
what you will do
in order to achieve a goal so this
business set like a more concrete goal
and
to basically or like kind of a behavior
that you need to carry out to achieve
the goal and i think about the when uh
components that's about ten minutes but
uh when do you want to uh
do that thing and then you have the how
component is about how
would you want to would you want to do
that and normally there are just
multiple
different methods to achieve the same at
the same goal
um
so this is like as slightly lower level
than the the component
the word component
so in the psychology literature
and also in neuroscience they are
research that shows some distinction
between for instance the world and the
world components and also the world and
how components i'm not going into
details
but here
talking about some like real-life
examples so you can think about those
different components in certain
applications
so if you look at
the subscription from ns
they have different kind of
restrictions in terms of what went on
how for instance
there you some
subscription
only allow you to for instance go to a
certain
city or place you have to sort of say
okay that's where i want to go
and other subscriptions basically
restrict you from like when can you
travel like you you can or you cannot
travel during the the peak hours
and uh finally there is also the how
aspect like you know whether you can
travel maybe in the first or second
class
uh you can also think about uh for
example uh
drivers who use uber so the uber has
algorithm that sort of manage all these
workers then their behavior can may also
be restricted
in terms of a different aspect like
which customer they
they should pick up when they should
pick up the customer and also maybe
through
which routes should they drive to the to
the customer so in a lot of real life
examples you can sort of make
distinction between these different
components so in the embryo studies we
asked the question
what is the relative importance of these
three different
decision making aspects in determining
people's
personal autonomy so
if you restrict to what when and how and
which restriction was needs to like a
stronger sense of
um
like a um
reduced autonomy
um
we have like no i think we probably we
thought like of course all three shoots
be important but the question is
bringing about the the the relative
importance and we also wanted to know
do they also interact to influence
autonomy
or the influence autonomy is more or
less sort of independent from each other
so for example if you restrict
the word
components maybe the one component
doesn't matter anymore because
maybe the world is just very strong
restricted so when you see whether there
will be any interaction or there will be
actually no injection
and secondly uh we want to know
uh does it matter if the restriction
comes from a person or from a ai agent
or algorithm
so they are of course a lot of research
looking at
differences in trust and acceptance
in terms of
decisions made by human experts or ai
algorithms and i think
sometimes
studies found that people are obesity of
reasons but in some other studies they
also found people seem to like
ai always understand that the opinion
from experts
but there's not much like really the
different type of restrictions ai
algorithms can impose on our human users
so that's that's one of the uh the novel
the
novel aspect of our study
uh so we did three studies three
experiments on experiments i want to
focus on study
1a so
because the others two experiments are
extensions and replication of the first
so in this first study
we have basically a
three by two by two by two design sounds
rather complex but the basic
uh
manipulation is that we motivate of
course the three different components so
each of this aspect could be
decided by oneself could be decided by
by another agent
uh so all these are manipulated within
subjects
and then we have
uh
another factor so the source of the
restrictions could be from a human could
be from ai agents and we have a baseline
condition which we don't specify
the type of
the source of restriction so this is
manipulated between between subjects
uh what we did was we basically give
people uh
they they
imagine themselves being different
scenarios
and
including travel
work health and social the different
type of goals they may want to achieve
and they go through like eight different
scenarios in render holders where these
different things are
manipulated and we basically
we measure uh the perceived autonomy uh
in terms of uh
the freedom of choice control
uh the restriction on autonomy and also
responsibility and
this
all these items we basically aggregate
to have one measure of a perceived
problem
so this is how a
one
the task looks like
so the first basically they read a um
kind of description of a scenario
uh for instance here is about planning
at the holiday so very brief like kind
of abstract description and also we also
basically told them like what
uh
what are the meanings of these different
decisions like what when and how
and then uh they were told that they are
going to uh
to look at eight different scenarios
and these decisions can be made either
by themselves or by
another person
and so this is how
each each trial looks like so we
basically
visualize the allocation of these
different aspects using a diagram so
in this trial then all the three
different
components are determined
by
by myself
and in this trial you see that uh this
is a case
where um
someone can decide on the when
let's say to
uh to travel but cannot decide about the
word and the how those are determined by
uh supposedly to be by by the other
person and then they answer
for the four questions for the
measurement
so
so they are
for for each um
for each type of gold pursuit there are
eight trials and this was presented in
random order
uh we also manipulate the
source of the restriction so this uh
different version of the instruction
text replacement need to read
uh again this is again quite quite
abstract we don't want to define like
what who is the other person or what
kind of ai so it's just very much on the
conception level in this
first project
and in terms of
the background we basically it's the
same we just change at the level here
for
it could be the other person in the
human condition the ai agent in the air
condition or we just basically say
something is constrained in the baseline
condition
we did two applications uh so in study
two uh
we wanted to rule out the possibility
that the
the order of introducing and
regionalizing the component would bias
the result because we always put the
word on problem when and how
but we found that actually it doesn't
doesn't matter
and
also
we want to show like can we just simply
ask people like which aspect they
consider to be the most important but
then we found what they self-report is
quite different from uh
what is revealed in the uh from the task
and uh we did a third study where we
extend uh the study to the more
organizational settings like
allocating job tasks or making like a
career plan
we also replicate the human versus ai
effects i will say you know what we
found
and
we include two additional dependent
variables like how much they would like
a decision-making situation and also um
do they actually accept sort of to go
along with with the situation defined in
the in the scenario
and so this is a
very basic result from study one so you
see
the three different between subject
conditions the baseline human ai
and
you also see
basically what we found was that
when you remove each of the the
component from one's own control
uh
pacific only would basically
go down so that's quite uh
unsurprising uh you see
uh also like what happens when persons
hear
the word class when these two are
controlled by by oneself in that trial
but not the how
and so
this is a
maybe
easier to understand the basic result
from study so here
we plotted
the
regression coefficient so
these are basically the
effect sizes of these three different
components and also the uh
the two-way interaction effect and the
three-way interaction effect
and what we found was that if you
restrict any of the three components
pacific autonomy will basically be
reduced
substantially
we found
very little interaction effect so that's
like all these different components they
uh what when how they affect autonomy
more or less like uh independence or you
see they it's actually in fact very
close to q0
we also didn't find any difference
between the human and the ai condition
that both things study one and the
stylish three so you see
uh
the bars with different colors and
really
only tiny differences
if you compare that to the
to the effect size
uh of just removing this component
that's like for instance here like
it's so close to one that means if you
uh
restrict any of those the pacific
autonomy will be reduced by that one
point on a 7.0
um we try to compare the relative
importance and in study 1 and 2 we found
sort of a consistent
order uh of what how and when
in terms of their importance so it's
like people
did consider the world to be maybe the
most central to pursuit of autonomy
followed by how and when we saw this
is pretty consistent should replicate
again but not really in study three when
we use
a little bit more concrete scenario in
organizational settings and here you
find the much smaller difference and the
same is also the when components now
becomes a bit more important than the
hub
so to conclude
uh
the three studies so i think it was not
surprising that of course in such a
abstract experiment if you just
manipulate the restriction on this
different
aspects
personal autonomy and also goal
motivation always go down
we didn't find instruction effect
meaning that
perhaps in a way you can say you can
sort of
uh the removal of one component by maybe
giving people freedom to choose maybe
another uh
in terms of another component so if you
restrict
the world component maybe it's nice to
give people to choose like when they can
do something
and we were initially kind of excited
about seems to be like a clear order in
this different aspect but that was not
really replicated in the service like
we also don't find didn't find the
difference in terms of uh restrictions
from human and ai
um
[Music]
of course at least not in such a quite
abstract experiment in real application
context of course that could still be a
different story at least i i think that
way
um
then
one thing i actually found with
struggling is like what kind of
design
implications you can think of
from the result as i said the model was
sort of really
um
found like a very strong kind of a
conceptual psychological perspective so
we also want to hear your opinion uh do
you think some of these results might be
interesting in terms of
uh creating kind of design space like
different can you separate the different
components uh like which one you want
maybe their systems to constrain and
which one you want people to be it to be
to be free to choose
so i want to continue with
somewhat different
talk about somewhat different study so
this is
um the work
oh sorry so this is the word um
mainly by my collaborator supplier from
the id departments at qe so this is a
little bit more
a little bit closer to application so we
wanted to look at what are the effects
of providing explanation and also like
maybe making people to be aware of like
this really ai or algorithm behind some
applications what would be consequences
on people's
autonomy and in terms of like very
common basic everyday sort of
interactions with ai
so
we have many three different research
questions so we want to know
does providing explanation help to sort
of protect
specific autonomy to some extent uh of
course there's a lot of research on
explainable ai so the general idea is
that this should
i think normally would be a good thing
so you want to see whether indeed if you
provide explanation maybe people are
less intimidated by some of the
automated recommendations
i want to know how does also the
uh awareness of ai influence specific
autonomy these we don't have like a
clear prediction we don't know whether
making people to be more aware of the
the algorithm or something like ais
behind the system is is beneficial
always
detrimental for for pacific autonomy
finally
we also want to explore the like
differences of course like different
everyday applications
such as like movie recommender or like
color application system kind of smart
formal application or like social media
sort of uh
filtering of like like what you read
um
so
what we did was again online experiments
where we used uh
things called design fiction method to
create different scenarios uh using some
video clips we had a 2x2 between sundeck
design
so
for for each application the participant
either they
uh they saw a explanation or they don't
so i don't see any explanation and they
were made to be
highly aware of the ai behind the system
or that was not the case
um
we got
over 300 percent from politics so about
80 percent per condition
uh so the person basically they went
through the eight applications in the
rundown order uh so we we used they did
some videos to show the behavior of
those ai infused systems i will show a
few examples of the video in a minute
then we measure
placebo autonomy
this time using a bit more
questions kind of
more relevance to the applications so
you see
for instance whether the system provides
choices based on the user's true
interest
or like the system left the user to do
things their own way
and
so we have five of these items to
imagine pacifica told me
um so this is my example
how
the solarium looks like so this is
like left legs
so she was like when you log in so this
is a condition
um
it presents of some kind of orgasm is
made
very standing so they release that you
know this
sure kind of artificial brain and you
know
and this is also a condition with
explanation uh so so for each we
recommend it there is some labels
underneath showing that
why what are the reasons that you might
like certain movies like certain actors
or certain
genre of a
film
and this is a
the
i think the the thermal starts
application
yeah so again you can see some
explanations here about why the system
set like this particular temperature for
you
and finally this is a example from the
car navigation
application
okay
so also some explanations are
here
um
so again
it was a 2x2 design so either
they see something like this like
highlighting the presence of ai or they
see
something much simpler
or and also
for some participants they
um explanations were provided by the
system or there's no explanation
so what we found
we
we checked whether the modifications
sort of
uh work or not you know whether people
actually notice the differences that
seems to be all right so um
you know people they
they were aware that
they were aware of the
explanation
so it was different rating in terms of
how much do you think the system was
providing any explanation between the
different conditions and also in terms
of the motivation of the awareness and
but
it's also it was also clear that the
meditation of this foundation worked
a lot better than the other one the
other one was maybe a bit too subtle you
know the difference between the
conditions uh was was pretty small
and if you look at
all the
results across all applications
it's actually
nothing was really much interesting what
happened uh there was no effect of
explanation knowing effect of
you know prime is like higher awareness
of ai
also no instruction effect
what becomes a little bit more
interesting is that you look at the
individual applications
and there was some somewhat surprising
to us that we found
for whatever reason uh
the car navigation system sort of
stood out
because
we seem to find actually pretty
relatively strong
effect of
providing explanation of a civil
autonomy so if they will show the
explanation they
the person they perceive the autonomy to
be higher
in in the car navigation case so you see
this visualization you can see the
distribution are
separated quite a bit
here but for all the applications there
was almost just a totally overlap
just to check was that corrected to
multiple comparisons it was so we sort
of uh divided like alpha by
eight and also to be honest we had a
quite large sample like 80 per between
subject condition
so
i was also thinking whether this is like
a flirt or
you could replicate attempts to believe
you probably would replicate but
it could be something specific really in
the design of our scenario or it could
be something more interesting really the
car navigation is uh have some like
different special attributes maybe
compared to other uh applications that's
that's i think still quite debatable
charge us to check seriously we have uh
we want to leave some room for
discussion yes
okay yeah it's almost at the end of the
talk
i have a couple more slides but i will
skip some of the more recent results
some like three five years but uh and we
also cover like the differences of
course the application so regardless of
whether
regardless of our manipulation so
providing explanation a lot or
you know priming the ai or not you also
see
[Music]
some trends in terms of like what kind
of application people
for what kind of application people seem
to be worried about maybe personal
autonomy a bit more than others uh if
you do some proper tests and you could
argue that these things like
um for social media like you know like
facebook that filter like what you see
from your friends for instance that is
you know this is like worried the most
probably by uh
the for the for the punishments also
the climate control but for like fitness
coaching and car navigation
uh those were people tends to perceive
just higher autonomy regardless of
you know the different type of different
uh manipulations
um
so i thought it was interesting to
observe the difference across the
different applications and that that
hasn't been done a lot
and the other question is why this kind
of application system stood out so is it
because of
that it's kind of a
more
have a critical application that's about
like real-time decisions or actions
uh or for or something else i
i don't know uh could also be just the
the way we design the scenario that's
then that's that's interesting um
and i think one limitation is that the
manipulation of this awareness of ai was
maybe interpreted differently by
different perspectives like showing the
brain people could think about data
privacy issues you could think about
like this maybe this system is very
smart or could could
all go
all different different ways
um i think i will stop here
there are some other things that we can
also talk maybe uh in the next
half an hour but
i want to know if there are any
questions about the uh what they're
presenting great thank you
[Applause]
um the question is to unexplain the
design scenarios implied action and so
i'm thinking maybe if one if one thinks
of the netflix
movie recommendation maybe there's an
expectation of picking one anyway so
explanations might not matter so much
ways with the nest it looked like there
was no option on the following action so
i couldn't go and suggest to change the
temperature and with the car then you
have maybe two options you can take that
electric so that was just a question if
if there was a thought around which
actions are implied by the scenario
um and then the other question was
around the
watch when um and how you spoke about
context dependency that you've seen and
to what extent do you think is also
cultural dependency
um so first so thanks for the for the
further i think very good questions
uh so for for this study i think
uh there are actually many differences
across the the scenarios to be honest uh
we try to make this to be realistic but
that would also mean
you don't have a lot of control so
uh for instance as you suggest like for
the movie
scenario uh it's like you still need to
take action you actually need to decide
but like the room temperature is sort of
it's it's that for you then you see some
expansion alone
um
those might of course
change people's experience
even though sort of
the only one that stood out was really
the calibration because for the others
sort of um
you know the motivation
we thought might influence how people
perceive autonomy was not
not really the case
and
but indeed perhaps because for the
social media and the
uh the time the temperature control
indeed people tends to
perceive indeed if you look at table a
little bit lower
just autonomy in general could indeed
maybe be because of you know in those
cases it's like the decision is made you
just see what's come out of the
uh the algorithm and uh you are either
happy or unhappy
but i guess for some other things
we recommended you still need to make
active choice uh
high navigation sort of
also a little bit
um
but i would say probably we need like
more
dedicated studies too if you want to
also separate those those different
different factors
and
in terms of the cultural difference for
the different components
um
i i don't know i tend to think if we use
such a abstract study probably not
the context
could probably indeed quite critical
and however somehow from the first two
studies we found a very consistent order
even for different sort of scenario like
planning a travel or
planning planning your work or like just
playing like a social event didn't seem
to system to matter
but then we when we switched to like
more concrete scenarios you know
organizations
then we found there were quite some
variations in terms of the relative
importance
we
so i also did a follow-up study where i
want to sort of
go beyond the the really abstract
scenario but i did the experience
sampling studies um different type of
restrictions people's decision for like
three meals like breakfast dinner lunch
and again asking like what aspect was
constrained or not constrained now over
there you again find the in the kind of
the the case of
uh dietary choices order is again not
what we found like that the caneer
what's how and when so probably it
depends a lot on different uh different
type of uh corporate suits
okay before we continue the discussion
do you mind sitting here so that we can
turn the camera so people in downline
meeting can actually see all of us yes
that's okay can you change slides from
here i think again right
[Music]
great more questions
yeah thanks so thanks for super
interesting talk
um great question actually about the
first
half perhaps so you laid out these
really nice different
perceptions of meaning of autonomy that
that they're out there
and then you propose your own one so and
i think that's the one you use here
in the setup of your experience so i
wondered
if you would have set up your experiment
differently or if you would have
perhaps do you think the results should
be interpreted differently if you had
used the different conceptions or one of
the other ones maybe the canteen one or
so would that have
yeah would have changed
um i would say it's more as follows the
more kind of the the liberty like
independence of you because still
reading about like restrictions and
choice not so much really about
sort of relational and
self-control doesn't have like a moral
aspect
um and
i i think what we
came up with like this comparison for
these three different
components is also a little bit like
investigation at slightly different
level so it's not just about restriction
or not like or what people do you know
do people follow
uh like a second order or personal
desire but it's more like
different
different uh let's say
different type of restrictions
um
so
i think in general it follows the more
perspective that autonomy is about
uh you know do you sort of
control your actions do you can you
decide something something by yourself
so that's i think that's that's how uh
how that was sort of
implemented in in the
in the experiments
um and this this distinction between the
different components i i will say it's
something something a bit quite specific
so
uh i had actually a little bit
difficult time at the beginning to see
is this really important because
intuitively
you know you can talk about those
different aspects
although in real applications they're
also open into interviewing you know
sometimes if you can can't decide like
where to go like also the timing is
probably also constrained
so um
so why i thought you know in what way
but perhaps in kind of
the design scenarios
designers can
really play with these parameters you
know to to set like you know this like
on off like to create different uh
different type of interventions
yeah so i i tend to agree with you
so my wish would also be that it really
matters which which of these
uh meanings you would use and that's why
i'm you know i'm
very interested in also why you would
think so why you chose this specific one
because
so if you if you choose perhaps your own
conception of autonomy then you might be
designing or checking or measuring for
something that you don't want to check
for but uh i think
that you
so i think in the first part of your
answer you said something that this
captures what we actually find important
in this specific context
um
is that is that the is that indeed but
is it a good interpretation what were
you just said or yeah so i think this
idea that these different components
more follow from like the some uh
research psychology about like people's
school pursuits like what are the
different aspects
different decisions
and
so
i don't know if like you take a
completely different uh
perspective and told me how would that
you know change the description of this
this very
specific things
and
so i would say it sounds like a slightly
different different angle you know it's
like about you know what what type of
restrictions rather than you know
whether you are being restricted or like
in what way you are being restricted
okay thanks
okay
uh thanks
so
you talked about it
again earlier i understand correctly in
your studies autonomy is expressed as a
number right apparently so how like so
what's the how does that measurement uh
how is that measurement defined in a
related question and it's related to
some of the things that were discussed
uh do you see some limitations in uh
reducing the meaning of autonomy
in this context to a number what might
get lost uh when we reduced
so
in i think both studies was kind of
measured on
a
uh like
one example scale like here
they basically
need
like what it stands to
feel that you have living on chores in
such a scenario so they imagine the
scenario this is this is how things are
determined they um they they basically
uh
basically choose one of the options
they think this is uh really indeed
that they have a lot of freedom when
they have legal freedom
um
we try to cover like
the concept like
some
different dimensions so we have
uh
one question about freedom of choice one
question about control
just like talking about restriction
autonomy very directly and also
responsibility that's also open
assumption that if you are
not autonomous then you are not
responsible for action
um so in this case
in the end they correlate really highly
and we aggregate it
um
a single question like reducing these
two numbers how would that sort of would
be an imitation
and
i think yes
even though i think
this is like i would say like very
common limitations in a lot of similar
studies that you try to
measure certain
people's feelings or attitudes
you know using using this type of
numeric skill you could of course
use like
a different approach uh
maybe
doing
qualitative research to ask how people
really feel then you can capture a
little bit more uh perhaps
nuances in terms of you know maybe like
again like maybe the even though what
when how
um
these different restrictions people
might have like
different specific feelings to them
that's indeed that's not really captured
by just these four items
just to quickly share so from this from
my experience with engaging with the
students in some research projects where
questions related to autonomy were
investigated
that we founded this qualitative aspect
to engaging with people and kind of
really going into the nuance that was
really useful to them by seeing what are
the design directions
uh to for a for example user user
interface
uh
so so so so that's where that nuance
came to that inspire particular things
to be designed yeah totally i think i
think also
some measures should be on the behavior
level not just
the waiting but you know how they you
can also sort of maybe observe how they
would continue to maybe interact with
system i think uh that would be
indeed also very useful um
studying like that alternative
automation
[Music]
we are running out of time
so maybe more questions from there
yes if possible
thanks uh
no questions from your
audience uh there is a question from
timon but he's left already so i think
we just
reflect him later on so my question was
if i take a step back from your results
it seems like
there were no no big effects
uh and so similarly
my sense is also that maybe a
qualitative perspective
would actually really enrich the
understanding but i was also thinking
i'm a quantitative guy after all as well
so i think maybe there's big confounding
factors
or maybe it's it's not a good metric
that but also people maybe it's bigger
so i was thinking
where are the consequences
are the consequences of you know
uh in all of these these cases of what
you would go wrong
and
what's your ability to resolve
misalignments
and my senses because that's what this
community is about this or actually
having some kind of control and so it's
also related to one of these autonomy
factors and so my question is could you
respond to what i think you think that
that is
maybe you know preconceived perceptions
of consequences and ability to actually
do something about it
actually played part in the in the
variability
and and therefore perhaps the lack of
big effects yes it's a really good point
because i also think like one may
limitation i wouldn't say it's like
really confined result but
in all these cases it's not like very
abstract scenarios
they don't really make decisions they
they give a scenario how the decisions
will play out so that's the only thing
they perceive then they raise basically
how they feel about that situation so
they don't make a real
decision they don't make choices
there's like no sort of consequences
like manipulate in the experiment so um
if they just look at you know how this
is supposed to be the case in
show showed in in those diagrams uh i
think different people can relate to to
those scenarios in different ways and
think about what would be consequences
uh i think even
the way we describe like the other agent
is also very abstract there's no like
you know what is the relationship with
this person to you or it's just not it's
not like
uh so the these all these things are not
let's say uh flesh now
um
so
um so one my idea i have sort of at the
end of those was we need to
you can still try to see
how you know does these different
components matter you know what the one
how
about for instance using
a different approach where you have
people actually making decisions
together with the system like maybe
maybe maybe a chatbot you know some kind
of conversation and you see different
options like some constraints someone
not constrained i think in that case i
would say
it
should give you a bit more realistic
tests on how this how this difference uh
passes different dimensions
okay great um
yeah i'd say uh
it's technically over but we can stick
around for uh
or enough to party uh because i think
the room
is i mean we didn't book it after two
but nobody's showing up sorry for half
an hour yeah great uh but with that i
think we have to say goodbye to our
online participants
and then stop the recording
and then i think
oh okay
you're just yeah
to explore
|
10dd597e-b261-4ffd-a23e-cd64e68035de
|
trentmkelly/LessWrong-43k
|
LessWrong
|
Some thoughts about natural computation and interactions
Epistemic Status: Ramblings of my current thoughts on computation.
I have been wondering about the nature of computation for some time now. For instance, what do we mean when we say the brain computes? I think the traditional answers are unsatisfactory. Chief amongst these issues is the role of observers in computation.
When an electronic calculator performs addition, it outputs the result by controlling pixels on a screen. Photons bounce off the screen and hit our retina and the brain performs a dizzying array of computational work in order to make sense of that retinal input. However, when discussing the computation the calculator performs, we talk about automata and locations in Chomsky hierarchies and the like, but we do not even consider the computational work the brain is doing in order to compute addition.
The issue is that while the dynamics of the calculator are well defined and independent of any observer, both the inputs and the outputs only make sense by virtue of an external entity doing some work. What kind of work is this observer doing?
Ultimately I think the observer must be doing the same kind of work the calculator is doing. Everything is (open) dynamical systems changing states, interacting with eachother.
When I think of a brain computing, I imagine a dynamical system taking inputs.
What are the "states" of the computation the brain is performing, given we think of it as a dynamical system? Who is the observer of our brain dynamics? In order to not have a homonculus, I think the meaningful computational states must be defined internally, without reference to external systems. Consider two distinct sets of neurons firing. Are these different computational states or not? The proposed answer here is that they are different only insofar as they constrain the future evolution of the brain in the different ways. If they constrain the dynamics similarly, then they are similar computational states.
There are many details here about what it means
|
be3c7060-f30a-478c-ab0c-eb3a4748b0e4
|
StampyAI/alignment-research-dataset/alignmentforum
|
Alignment Forum
|
Experiences and learnings from both sides of the AI safety job market
*I’m writing this in my own capacity. The views expressed are my own, and should not be taken to represent the views of Apollo Research or any other program I’m involved with.*
In 2022, I applied to multiple full-time AI safety positions. Now, I switched sides and ran multiple hiring processes for Apollo Research. Because of this, I feel like I understand the AI safety job market much better and may be able to help people who are looking for AI safety jobs to get a better perspective.
*This post obviously draws a lot from my personal experiences many of which may not apply to your particular situation, so take my word with a grain of salt.*
Executive summary
=================
1. In the late Summer of 2022, I applied to various organizations working on AI safety. I got to the final stages of multiple interview processes but never received an offer. I think in all cases, the organization chose correctly. The person who received the offer in my stead always seemed like a clearly better fit than me. At Apollo Research, we receive a lot of high-quality applications despite being a new organization. The demand for full-time employment in AI safety is really high. This should probably change applicants’ strategy and expectations but should not stop them from applying!
2. **Focus on getting good & provide legible evidence:**Your social network helps a bit but doesn’t substitute bad skills and grinding Leetcode (or other hacks for the interview process) probably doesn’t make a big difference. In my experience, the interview processes of most AI safety organizations are meritocratic and high signal. If you want to get hired for an evals/interpretability job, do work on evals/interpretability and put it on your GitHub, do a SERI MATS stream with an evals/interpretability mentor, etc. This is probably my main advice, don’t overcomplicate it, just get better at the work you want to get hired for and provide evidence for that.
3. **Misc:**
1. **Make a plan:**I found it helpful to determine a “default path” that I’d choose if all applications failed, rank the different opportunities, and get feedback on my plan from trusted friends.
2. **The application process provides a lot of information:**Most public writings of orgs are 3-6 months behind their current work. In the interviews, you typically learn about their latest work and plans which is helpful even if you don’t get an offer.
3. **You have to care about the work you do:**I often hear people talking about the instrumental value of doing some work, e.g. whether they should join an org for CV value. In moderation this is fine, when overdone, this will come back to haunt you. If you don’t care about the object-level work you do, you’ll be worse at it and it will lead to a range of problems.
4. **Honesty is a good policy:**Being honest throughout the interview process is better for the system and probably also better for you. Interviewers typically spot when you lie about your abilities and even if they didn’t you’d be found out the moment you start. The same is true to a lesser extent for “soft lies” like overstating your abilities or omitting important clarifications.
It can be hard & rejection feels bad
====================================
There is a narrative that there aren’t enough AI safety researchers and many more people should work on AI safety. Thus, my (arguably naive) intuition when applying to different positions in 2022 was something like “I’m doing a Ph.D. in ML; I have read about AI safety extensively; there is a need for AI safety researchers; So it will be easy for me to find a position”. In practice, this turned out to be wrong.
After running multiple hiring rounds within Apollo Research and talking to others who are hiring in AI safety, I understand why. There are way more good applicants than positions and even very talented applicants might struggle to find a full-time position right away. I think this is bad and hope that there will be more AI safety orgs to use that talent (as I’ve detailed [here](https://www.lesswrong.com/posts/MhudbfBNQcMxBBvj8/there-should-be-more-ai-safety-orgs)). Currently, AI safety organizations have the luxury of hiring candidates who can contribute almost immediately. In other industries, employers expect that they have to invest the first couple of months to upskill new hires. In AI safety, the supply of talent is high and programs like SERI MATS are doing a great job, so many of the best candidates can contribute from day one.
While this may sound demotivating to some, I think it’s important to know. For example, I’d recommend applying to multiple positions and programs and spreading your bets a bit wider since any individual position might be hard to get. On the other hand, I think fairly fast upskilling is possible (see next section), so you can improve your chances a lot within a couple of months.
Another realization for me was that getting a highly desirable job is just hard in general. Michael Aird has [written](https://forum.effectivealtruism.org/posts/Fahv9knHhPi6pWPEB/don-t-think-just-apply-usually) and [talked](https://hearthisidea.com/episodes/aird/#applying-to-research-roles-at-ea-orgs) about this already but I think it’s still important to keep in mind. Most tech companies that work on AI safety probably get more than 100 applications per position. Even if you’re among the top five candidates for a position, rejection is still more likely than an offer. Furthermore, there is some stochasticity in the hiring process (but much less than I expected). The company has to make a decision with only a few datapoints. They know your CV and they have somewhere between 3 and 10 hours of interviews to judge you by which is not a lot of time. Also, you might have had a bad day for an interview or were unable to answer a specific question. So the fact that you got rejected may not mean a lot in the grand scheme of things.
A good friend of mine provided a framing for the hiring process that I found helpful. He thinks of it as a repeated series of coinflips where every individual flip has a low chance of coming up heads but if one does, you win big and can stop flipping for a while. If a job is desirable it is competitive and if it is competitive, rejection is more likely than an offer, even if you are among the best couple of candidates. However, this doesn’t mean you shouldn’t flip the coin anymore. You should still apply to jobs you want and think you’d be a good fit for.
Nonetheless, rejection sucks. I know that rejection is the norm and yet I was disappointed every time. I can even rationally understand the decision--other candidates were just better than me--but emotionally it still feels bad. If you care about the work you do, you usually want to work in a team of people who care as well. Furthermore, I think it is rational for you to envision how you would work in a specific company to be able to present a clear vision for your research during the interviews. I noticed that during this process I could really see myself in that position and built up a vision of how work there would look like. Then, a rejection just pops that bubble and you have to repeat the same process for another company.
Also, while rejections feel bad at the time, in the large scheme of things they are really pretty minor. So if the fear of rejection stops you from applying in the first place, I’d really recommend finding some way to overcome that fear, e.g. by having a friend send the application with you. The benefits of applying are typically much higher than the downsides if you get rejected (see later section).
**Main takeaways:** Most people don’t get their dream job on the first try. Rejection usually feels bad but it’s still worth applying to jobs that you think you are a good fit for. Spreading your bets is probably a good policy.
Focus on getting good
=====================
There is some amount of stochasticity in the hiring process and there are some benefits to being well-connected. However, my main takeaway from both sides of the hiring process is that the process mostly works as intended--the people with the best fit for the position get hired and the system is fairly meritocratic overall.
Ironically enough, my rejections made me more convinced that the process is meritocratic. Whenever I felt like I didn’t perform well in an interview, I got rejected shortly after, indicating that the interviewers potentially had the same feeling. Furthermore, the AI safety community isn’t that large, so I often knew the people who got the offer instead. Before I knew who got the offer, my brain would go “The process isn’t accurate, they probably made a mistake or the other person was just well connected” and as soon as I knew who got hired it was immediately very clear that they were just a better fit for the position and that the organization had just made the right call in rejecting me.
Furthermore, I now think well-run interviews provide way more evidence about a candidate's fit than I had expected. I’ve run maybe 70-100 interviews this year so far and I was surprised by how well you can judge the fit of a candidate in such a short amount of time. Firstly, it’s much harder to fake expertise than I thought. Thus, a well-posed interview question will very quickly find the limits of your actual expertise and differentiate between candidates. For example, I think it’s very hard to simulate being a good ML engineer for 60 minutes without actually being one. There might be some stochasticity from daily performance but the variance just seems fairly small in contrast to the difference in underlying skill. Secondly, people are (thankfully) mostly honest in interviews and are straight about their limitations and skills. So just asking specific questions directly already gives you a good sense of their fit. Also, most people are not very good at lying and if you overplay your achievements, interviewers will likely catch onto that pretty quickly (see section “Honesty is a good policy”).
Thus, my main recommendation is to “focus on getting good”. This might sound incredibly obvious but I think people sometimes don’t act according to that belief. There are a lot of other things you could focus on in the belief that they are the best way to increase your chances of a job offer. For example, you might overfit the interview process by grinding Leetcode 24/7 or you might invest a lot of time and effort into building a personal network that then gets you a job you’re not actually qualified to do or you might jump on lots of projects without actually contributing to them to increase your visibility.
However, I think most of these shortcuts will come back to haunt you and every sustainable benefit is mostly a consequence of being good at the core skill you’re hired for. You might be able to trick some people here and there but once you work with them they will realize you’re not up for the job if you don’t have the required skills. Also, I think people overestimate their ability to Goodhart the interview process. Interviewers typically have a lot more experience at interviews than candidates. If you’re trying to oversell, a skilled interviewer will catch you right in the act.
Thus, I suggest people should rather focus on doing a project that requires similar skills as the job they’re looking for than grinding Leetcode, building a network, or jumping around between projects and then providing legible evidence of their skills (see next section).
Typically, it’s fairly obvious to know what “getting good” means, e.g. because organizations state it in detail in their job descriptions or on their websites. Most of the time, organizations who are hiring are also fairly straightforward when you just ask them about what they are looking for.
**Main takeaways:**focus on getting good at the core skills that the job requires, and ignore most other stuff. Don’t overfit to Leetcode questions and don’t freak out because you don’t have a big network. Assess your skills honestly and focus on the core things you need to get better at (which are typically fairly obvious). If your key skills are there and visible, everything else will come on its own.
Provide legible evidence of your abilities
==========================================
Words are cheap and it is easy to say what you plan on doing or what kind of vision you have. Doing good research in practice is always harder than imagined and costs time. Therefore, providing evidence that you’re not only able to think about a specific topic but also able to make practical progress on it is an important signal to the potential employer.
One of the things employers are looking for the most is “has that candidate done good work very close to the work they’d be hired for in the past?” and I think this is for good reasons. Imagine having two hypothetical candidates for an interpretability position: both have a decent ML background and are aligned with your organization’s mission. Candidate A has interesting thoughts on projects they would like to run if hired, candidate B also has good thoughts and on top of that a 3-month interpretability project with public code under their belt. You can judge candidate B so much better than candidate A. The fact that candidate B did a project means you can judge their actual skills much more accurately. The fact that they then applied likely means they enjoyed the project and are motivated to continue working on interpretability. The fact that they finished it, published the code, and wrote up a report tells you a lot about non-technical skills such as their productivity and endurance. Lastly, they already invested 3 months into interpretability, so they are much closer to being able to contribute meaningfully to the projects in your organization right away. For candidate A, you have much more uncertainty about all of these questions, so the comparatively small difference of this project (it’s only 3 months of difference vs. ~5 years of education that they share) makes a huge difference for the employer.
Therefore, providing concrete evidence of your abilities seems like one of the best ways to increase your chance of an offer. Between employer and job, the specific evidence obviously differs but it’s usually not hard to guess what evidence companies are looking for, e.g. just look at their job descriptions. Furthermore, most companies usually just tell you exactly what skills they are looking for if you ask them, e.g. in the job description or on their website.
Doing a project on your own is possible but much harder than with mentorship or collaborators. Therefore, I strongly recommend applying to SERI MATS, ARENA, the AI safety camp, or similar programs to work on such projects. I personally benefitted a lot from SERI MATS despite having previous independent research experience and can wholeheartedly recommend the program.
In the past, multiple people or organizations have reached out to me and asked me to apply for a position or assist them with a project, and almost always one of my public blogposts was explicitly mentioned as the reason for reaching out. So the fact that I had publicly legible evidence was, in fact, the core reason for people to think of me as a potential candidate in the first place.
Lastly, putting all of the other reasons aside--working on a project for a couple of months actually provides you with a lot of new evidence about yourself. For example, I was convinced that I should go in a specific direction of AI safety multiple times and when I started a project in that direction, I quickly realized that I either didn’t enjoy it or didn’t feel good enough to meaningfully contribute. Thus, working on such projects not only increases your employability, it also prevents you from committing to paths you aren’t a good fit for.
In general, I can strongly recommend just going for it and diving into projects in fields that sound interesting even if you’re new to them. You don’t need to read all the other stuff people have done or ask anyone for permission. You can just start hacking away and see if you enjoy it. I think it’s by far the most efficient way to both get better at something and test your fit for it. When you enjoy the work, you can still read all related papers later.
**Main takeaways:**Providing legible evidence for your skills makes a big difference in your employability. Doing concrete projects that would produce such evidence is also a great way to test whether you’re good at it yourself and whether you enjoy it.
Make a plan
===========
When I started my job search process, I created a Google doc that contained the following items:
1. An overview of my CV where I try to evaluate how a reviewer would interpret different achievements.
2. A list of strengths and weaknesses from introspection and from feedback in the past.
3. A list of organizations that I could imagine working for with pros and cons for each organization.
4. A “default path”, i.e. a path that I would want to pursue if I got rejected by all orgs I applied to (in my case this was independent research and upskilling).
5. A priority list of organizations with predictions of my probability of receiving an offer (all predictions ranged from 5 to 20 percent); I then measured all organizations against the “default path” and decided not to apply to any organization that I preferred less than the default path. This left me with 5 or 6 organizations and a probability of getting one or more offers of ~50%.
I then sent this document to a handful of trusted friends who gave me really valuable feedback, changed a couple of things as a result, and then applied to the organizations that were preferable to the default path.
I found this process extremely valuable because
1. **It forced me to evaluate my strengths and weaknesses.** During the feedback gathering round, I realized that many people mentioned similar strengths that I had not considered before (I think the feedback is true, it’s just one of these things that “everyone else is mysteriously bad at” and I thus didn’t think of it as a strength of mine). This influenced how I wrote applications and what I emphasized in the interview.
2. **It forced me to make the case for and against every organization.** During this process, I realized that some organizations do not really fit my agenda or skillset that well, and I previously wanted to apply for status. Making the explicit case made me realize that I shouldn’t apply to these orgs.
3. **It forced me to come up with a “default path” which I found very** **anxiety-reducing.** Once I had a default path that I was comfortable with, I felt like nothing could go very wrong. In the worst case, I’ll follow the default plan which I think was still pretty good. I just couldn’t fall very low even if rejection would feel bad.
4. **It forced me to put honest probability estimates on my chances.** This made me realize that my estimated aggregate chance of getting an offer is only at 50% which made me plan my default path in much more detail.
5. **The feedback from my friends was very valuable.** It was helpful to get my reasoning checked and my self-evaluation criticized constructively.
I don’t think it’s absolutely necessary to make such a plan but I think it structured my thinking and application process a lot and it probably saved me time in the long run. It took me maybe a maximum of 10-20 hours in total to write the doc but saved time for every application by reducing my total number of applications.
**Main takeaways:**Making a plan can structure your application process a lot. I would especially recommend it to people who are looking for their first job.
The application process provides a lot of information
=====================================================
For a while, I had the intuition that “I need to prepare a lot more before applying”. I think this intuition is mostly false. There are cases where you just clearly aren’t a good fit but I think the bar for applying is much lower than many (especially those with imposter syndrome) assume. My heuristic for applying was “Would I take the offer if I got one and do I expect to make it through the screening interview” (this might even be too high of a bar; [when in doubt, just apply](https://forum.effectivealtruism.org/posts/Fahv9knHhPi6pWPEB/don-t-think-just-apply-usually)).
There are many ways in which the application process gives you valuable information and feedback:
1. The **job description** is often relatively detailed and companies say very clearly what they want. Just looking through the descriptions often gave me a good sense of whether the position aligns with the research I find most promising and whether I should apply in the first place. It also gives a pretty clear path to which kind of research you might want to work on if you want to increase your chances in the future. Most organizations are pretty transparent with what they are looking for.
2. If you get **rejected without being invited to an interview**, this is unfortunate but still valuable feedback. It basically means "You might not be there yet" (though as Neel points out in the comments, CV screening can be a noisy process)~~“You clearly aren’t there yet”~~. So you should probably build more skills for 6 months or so before applying again.
3. If you get into the interviews you usually have a **screening interview** in the beginning, i.e. what you want to work on, what the company wants to do, etc. While some of this information is public, the company's public record usually lags behind the actual state of research by 3 months or more. So talking to someone about what the org is currently working on or intends to work on in the future, can give you a lot of valuable information that you wouldn’t get from their website. This made it much easier for me to decide whether my own research goals were aligned with those of the org.
4. The **technical interviews** give you some sense of what kind of level the company is looking for. If they feel easy, your base skills are probably good enough. If they feel hard, you might want to brush up. I found that technical interviews really differ between companies where some do very general coding and math questions and others very concrete problems that they have already encountered within their work. I found the latter to be much more valuable because I got a better feeling for what kind of problems they see on a day-to-day basis.
5. I applied to research scientist positions and thus usually had **research interviews**. In these, you talk about the research you did in the past and the audience asks you questions about that. I found it valuable to not only talk about my past research projects but also lay out what I intend to work on in the future. In my case, my future plans have nothing to do with my Ph.D., so it felt important to emphasize the difference. In general, I found it helpful to prepare the research interviews with the question “What do I want other scientists at that organization to know about me?” in mind.
6. In some cases, you get a **final interview**, e.g. with the manager you would be working under or some higher-up in the company. These interviews are usually not technical and can be very different from person to person. In some cases, it was just a friendly chat about my research interests, in other cases, I was asked detailed questions about my understanding of the alignment problem. During the latter interview, I realized that I was unable to answer some of the more detailed questions about the alignment problem. On the one hand, I knew right after the interview, that I’d be rejected but on the other hand, it forced me to think about the problem more deeply and led to a change in my research agenda. Thus, I’d say that this interview was extremely helpful for my development even if it led to a rejection.
7. The **final decision** of whether they make you an offer or not is valuable feedback but I wouldn’t update too much on rejection. If you get the offer, that’s obviously nice feedback. If you get into the final round, that means you’re nearly there but still need to improve and refine a bit but it could also be a result of the stochasticity of the interview process.
Importantly, interviews go both ways. It’s not just the organization interviewing you, it’s also you interviewing them. Typically, after every interview, you can ask questions about them, e.g. what they are currently working on, what their plans are, what the office culture looks like, etc. The quality of the interviews is also feedback for you, e.g. if they are well-designed and the interviewer is attentive and friendly, that’s a good sign. Whenever an interview was badly designed or the interviewers just clearly didn’t give a shit, I updated against that organization (sidenote: if you went through Apollo Research’s interview process and felt like we could improve, please let me know).
**Main takeaways:**The hurdle for applying is probably lower than many people think. I find “Would I take the job if I got an offer and do I expect to get through the CV screening?” to be a good heuristic. Interviews provide a lot of information about the organization you’re applying to. Interviews go both ways--if the interview feels bad, this is evidence against that organization.
You have to care about your work
================================
I think that people interested in AI safety are more likely than the median employee to do their job for the right reasons, i.e. because they think their work matters a lot and it is among the best ways for them to contribute. However, many other factors influence such an important decision--status, money, hype, flavor of the month, etc. My experience so far is that these other influences can carry your motivation for a couple of months but once things get tough it usually gets frustrating and it’s much harder to show a similar level of productivity as with a project you actually deeply care about.
Caring about a project can come in many flavors and is not restricted to intrinsically caring about that particular project. The project could also just be a small step to reaching a larger goal you care about or to learn something about another project you care about.
For me, a good heuristic to investigate my motivations is “Would I do this work if nobody cared?”. For a long time, I was not sure what approaches I find most promising and am a good fit for. After a period of explicitly thinking about it, I converged on a cause and approach that I felt very good about. At that point, I realized that I deeply cared about that particular avenue (detecting deceptive alignment with empirical approaches), and my future plans were roughly “apply to a company and work on X if accepted” or “if I don’t get an offer, work on X anyway (on a grant or self-funded)”. I found this insight really helpful and freeing and my speed of improvement increased as a result. It also led to me starting Apollo Research because nobody worked on exactly the thing I found most important.
More concretely, I think there is a common failure mode of people deeply caring about AI safety in general and therefore thinking they should work on anything as long as it relates to AI safety somehow. My experience is that this general belief does not translate very well to your daily mood and productivity if you don’t also enjoy the object-level work. Thus, if you realize having thoughts like “I’d really like to work for <AI safety org> but I don’t think their object-level agenda is good”, it may be a good time to rethink your plans despite salary, status, experience, and other pull factors.
Obviously, there are caveats to this. Early on, you don't know what you care about and it's good to just explore. Also, you shouldn't overoptimize and always only do exactly the thing you care most about since there are real-world trade-offs. My main point is, that if you realize you don't really care that much about the work you're doing consistently, it's a good time to ask if it is worth continuing.
**Main takeaways:**There are many reasons why we choose to work on different projects or in different positions. In my personal experience, the motivation from most reasons fades unless you actually care about the actual work you do on a day-to-day basis.
Honesty is a good policy
========================
A general takeaway from job applications, hiring, and working in my current role is that honesty is generally a good policy (which doesn’t mean you should always say every single thought you have).
From a systemic perspective, honesty makes everything much more efficient. When candidates accurately report their skills, it’s much easier to make decisions than when they try to game the system since fewer guardrails against lying need to be put in place. Thus, it would require less time for interviewers and applicants to go through the hiring process. However, such a system could be exploited by a skilled adversary who lies about their skills. Thus, organizations have to protect against these adversaries and make the process harder and more robust. This increases costs and bloats up the process.
Typically, this means that interviewers have to double-check information and dive much deeper into topics than would be necessary if everyone were honest. Since hiring is a dynamic process, some applicants will try to come up with new ways to game the process and the organization has to spend a disproportionally large amount of time finding and dealing with these adversarial applicants.
However, my best guess is that being such a dishonest adversary is rarely beneficial for the applicants themselves. Interviewers often have had hundreds of interviews and thus have much more experience spotting dishonest applicants while any given applicant probably had much less. Furthermore, most employers (AI safety orgs probably even more than others) see dishonesty as a big red flag if found out.
Furthermore, just from a personal view, you probably also prefer to work with honest people, so a signal that you’re committed to honesty may make other people around you more honest too.
Finally, I think it’s good to explicitly pre-commit to honesty before the hiring process. It’s plausible that there will be situations where you could get away with some cheating here and there or some “small lies” but it’s bad for the overall system and likely not worth the risk of getting caught. For example, you may ask a friend who has already gone through the process to give you hints on what to prepare or try to overplay your skills when you expect the interviewer not to check the claims in detail. When you really want to have the job or you panic during the process, you may be tempted to cheat “just a bit”. To prevent yourself from rationalizing cheating in the moment, I’d recommend pre-committing to honesty.
**Main takeaways:**Just be honest. Neither overstate nor understate your abilities and answer questions accurately. It’s good for the system as well as yourself.
Final words
===========
My impression of the AI safety job market so far is that, while the system is sometimes a bit unorganized or stochastic, it is mostly meritocratic and the people who get offers tend to be good fits for the job.
Most desirable jobs have a lot of competition and even very good people get rejections. However, this does not mean that you should give up. It is possible to improve your skills and since the field is so young, it doesn’t take a long time to contribute.
There are many things you can do to increase your chance of getting an offer such as grinding Leetcode or building a network and while these are certainly helpful I would not recommend prioritizing them. The primary reason why someone gets a job is because they are good at it. So my number one recommendation would be “focus on getting good and everything else will be much easier”.
|
340fbb02-89e2-406b-9934-0f94ad993522
|
trentmkelly/LessWrong-43k
|
LessWrong
|
How to Read
Part of my attempt to provide a bunch of unsolicited, anecdotal evidence that probably doesn't work for everyone.
-
Of course you already know how to read. But do you know how to read well?
Many people who read a book want to read for entertainment. That's perfectly ok -- it seems like a great way to take a break and enjoy yourself. But many times people pick up a book with the intention to learn something important. If you're one of those people, it's important not to fool yourself, end up not learning anything, and just waste your time. That's how you end up reading for entertainment without realizing it.
This was a big problem for me, and I realized I was wasting a lot of time when I otherwise could have been productively reading books.
While I'm still not the best reader, here's how I think I solved that problem:
I'm choosy about which books I read. There are millions of books in the world. I can't read them all. So I have to be choosy. I think about what I stand to gain from the book. Is it worth my time to read it? Is the book actionable? I personally aim to read books that come to me in reviews, that from a skim of their table of contents look like they'll provide real value to me.
I think about what else I could be doing instead of reading. Reading is great, but in many cases experience can be a better teacher. Moreover, picking up some experience can help me understand and apply the lessons in books better. I try to adopt a "doing-reading" loop, where I read something, act upon it, then read another something, etc., continuing to iteratively improve in whatever skill I'm after. This also helps validate the advice of books.
I'm not afraid to ditch an underperforming book. If I don't like it, it's wasting my time, and it's time to move on.
I consider sources other than books. Books are often the best source of information on any topic, and are often higher quality because they're intensely reviewed. But many blog posts and online re
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.