id stringlengths 36 36 | source stringclasses 15 values | formatted_source stringclasses 13 values | text stringlengths 2 7.55M |
|---|---|---|---|
cd5902b2-0c4a-4d66-aca4-4a7155241166 | trentmkelly/LessWrong-43k | LessWrong | Computing Natural Abstractions: Linear Approximation
Background: Testing The Natural Abstraction Hypothesis
Given a world-model, direct computation of natural abstractions basically amounts to figuring out which X-values give the same Y-distributions y↦P[Y=y|X], for various choices of the variables X and Y in some large Bayes net. In general, this computation is #P-Hard. (That’s like NP-Hard, but Harder.) That doesn’t necessarily mean that it’s usually intractable in practice, but it means that we need to use heuristics and approximations right off the bat.
A linear-Guassian approximation is a natural starting point. It’s relatively simple - all of the relevant computations are standard matrix operations. But it’s also relatively general: it should be a good approximation for any system with smooth dynamics and small noise/uncertainty. For instance, if our system is a solid object made of molecules and sitting in a heat bath, then linear-Gaussian noise should be a good model. (And that’s a good physical picture to keep in mind for the rest of this post - solid object made of molecules, all the little chunks interacting elastically, but also buffeted about by thermal noise.)
So, we have some Bayes net in which each variable is Gaussian - a linear combination of its parents plus some IID Gaussian noise. For the underlying graph, I’m using a 2D Delaunay triangulation, mainly because it does a good job reflecting physically-realistic space/time structure. (In particular, as in physical space, most pairs of nodes are far apart - as opposed to the usual Erdos-Renyi random graph, where the distance between most points is logarithmic. We’ll revisit the choice of graph structure later.) Embedded in 2D, an example graph looks like this:
Example of the kind of Bayes net structure we’ll use. Each node is a variable, and the arrows show causal relationships in the model. The whole graph embeds nicely in a 2D space, so nodes which are “far apart” in the spatial embedding are also “far apart” in terms of steps in the graph.
… al |
f2474365-e4ec-46d2-b5fe-ba3d2cdbbf0a | awestover/filtering-for-misalignment | Redwood Research: Alek's Filtering Results | id: post3609
Highlights Specification gaming examples in AI (Victoria Krakovna) : A list of examples of specification gaming, where an algorithm figures out a way to literally satisfy the given specification which does not match the designer's intent. Should you read it? There were several examples I hadn't heard of before, which were pretty entertaining. Also, if you have any examples that aren't already listed, it would be great to send them via the form so that we can have a canonical list of specification gaming examples. My take on agent foundations: formalizing metaphilosophical competence (Alex Zhu) : Argues that the point of Agent Foundations is to create conceptual clarity for fuzzy concepts that we can't formalize yet (such as logical uncertainty). We can then verify whether our ML algorithms have these desirable properties. It is decidedly not a goal to build a friendly AI using modules that Agent Foundations develop. Should you read it? I don't know much about MIRI and Agent Foundations, but this made sense to me and felt like it clarified things for me. Adversarial Attacks and Defences Competition (Alexey Kurakin et al) : This is a report on a competition held at NIPS 2017 for the best adversarial attacks and defences. It includes a summary of the field and then shows the results from the competition. Should you read it? I'm not very familiar with the literature on adversarial examples and so I found this very useful as an overview of the field, especially since it talks about the advantages and disadvantages of different methods, which are hard to find by reading individual papers. The actual competition results are also quite interesting -- they find that the best attacks and defences are both quite successful on average, but have very bad worst-case performance (that is, the best defence is still very weak against at least one attack, and the best attack fails to attack at least one defence). Overall, this paints a bleak picture for defence, at least if the attacker has access to enough compute to actually try out different attack methods, and has a way of verifying whether the attacks succeed. Technical AI alignment Problems Specification gaming examples in AI (Victoria Krakovna) : Summarized in the highlights! Metaphilosophical competence can't be disentangled from alignment (Alex Zhu) : Would you be comfortable taking a single human, and making them a quadrillion times more powerful? Should you read it? I am curious to see people's answers to this, I think it might be a good question to reveal major differences in worldviews between optimistic and pessimistic safety researchers. Reframing misaligned AGI's: well-intentioned non-neurotypical assistants (Alex Zhu) : Another way to think about problems from AGI is to imagine the AI as a well-intentioned but neuroatypical friend, who learned all about humans from Wikipedia, and who has access to immense resources. You would worry a lot about principal-agent problems in such a scenario. Should you read it? I like this framing. I'm not sure if it is actually a good model for act-based agents , but it's another way to think about what problems could arise from an AI system that is superintelligent in some domains and subhuman in others. Read more: Act-based agents Superintelligent messiahs are corrigible and probably misaligned (Alex Zhu) Technical agendas and prioritization My take on agent foundations: formalizing metaphilosophical competence (Alex Zhu) : Summarized in the highlights! Agent foundations 2018 research plans and predictions (Rob Bensinger) : Scott and Nate from MIRI score their predictions for research output in 2017 and make predictions for research output in 2018. Should you read it? I don't know enough about MIRI to have any idea what the predictions mean, but I'd still recommend reading it if you're somewhat familiar with MIRI's technical agenda to get a bird's-eye view of what they have been focusing on for the last year. Prerequisities: A basic understanding of MIRI's technical agenda (eg. what they mean by naturalized agents, decision theory, Vingean reflection, and so on). Musings on Exploration (Alex Appel) : Decision theories require some exploration in order to prevent the problem of spurious conterfactuals, where you condition on a zero-probability event. However, there are problems with exploration too, such as unsafe exploration (eg. launching a nuclear arsenal in an exploration step), and a sufficiently strong agent seems to have an incentive to self-modify to remove the exploration, because the exploration usually leads to suboptimal outcomes for the agent. Should you read it? I liked the linked post that explains why conditioning on low-probability actions is not the same thing as a counterfactual, but I'm not knowledgeable enough to understand what's going on in this post, so I can't really say whether or not you should read it. Quantilal control for finite MDPs (Vadim Kosoy) Miscellaneous (Alignment) Papers from AI and Society: Ethics, Safety and Trustworthiness in Intelligent Agents Guide Me: Interacting with Deep Networks (Christian Rupprecht, Iro Laina et al) Near-term concerns Adversarial examples Adversarial Attacks and Defences Competition (Alexey Kurakin et al) : Summarized in the highlights! Security Poison Frogs! Targeted Clean-Label Poisoning Attacks on Neural Networks (Ali Shafahi, W. Ronny Huang et al) : Demonstrates a data poisoning attack in which the adversary gets to choose a poison input to add to the training set, but does not get to choose its label. The goal is to misclassify a single test instance as a specific base class. They achieve this by creating a poison input that looks like the base class in pixel space but looks like the test instance in feature space (i.e. the activations in the penultimate layer). The poison input will be labeled by humans as the base class, and then when the network is retrained with the original dataset and the new poisoned input(s), it will classify the poison input as the base class, and with it the test instance as well (since they have very similar features). Should you read it? I was pleasantly surprised at how understandable the paper was, and they do a good job of looking at exactly what their method is doing and how it accomplishes the attack in different ways under different settings. Manipulating Machine Learning: Poisoning Attacks and Countermeasures for Regression Learning (Matthew Jagielski et al) AI strategy and policy France's AI strategy : See Import AI's summary . Initial Reference Architecture of an Intelligent Autonomous Agent for Cyber Defense (Alexander Kott et al) : See Import AI's summary . AI capabilities Reinforcement learning Retro Contest (Christopher Hesse et al) : OpenAI has released Gym Retro, providing an interface to work with video games from SEGA Genesis, which are more complex than the ones from Atari. They want to use these environments to test transfer learning in particular, where the agent may be pretrained on initial levels for as long as desired, and then must learn how to complete a new test level with only 1 million timesteps (~18 hours) of gameplay. (Humans do well with 2 hours of pretraining and 1 hour of play on the test level.) Should you read it? If you want to keep track of progress in deep RL, probably -- this seems quite likely to become the new set of benchmarks that researchers work on. There's also another example of specification gaming in the post. Learning to navigate in cities without a map (Piotr Mirowski et al) Deep learning Universal Planning Networks (Aravind Srinivas et al) : This is an architecture that has a differentiable planning module, that is, a neural network that takes in (encodings of) states or observations and produces actions. You can use this in conjunction with eg. expert demonstrations (as in imitation learning) in order to learn features that are optimized for the purpose of planning, focusing only on the details relevant to the task, unlike an auto-encoder, which must reconstruct the entire image, including irrelevant details. Should you read it? It's a good example of the push towards learning more and more complex algorithms using neural nets (in this case, planning). From a safety perspective, differentiable planning networks may be useful for modeling humans. |
bbe9e090-bc59-418f-af40-15db146744b6 | trentmkelly/LessWrong-43k | LessWrong | Vague Thoughts and Questions about Agent Structures
Epistemic status: Posting for blog day at MSFP! More trying to figure out what the right definitions are than saying anything concrete. I still don’t really know what agents are, and none of this is math yet. I’m hoping to develop these (and other) ideas more in the future, so any feedback is greatly appreciated.
My Naive Agent Pre-model
Y’know -- agents are, like, things that do things. They have utility functions and stuff. They make choices, whatever that means.
Unfortunately this ‘definition’ isn’t sufficient for making any concrete claims about how agents behave, so I’ve been thinking about some models that might be, and this post contains ideas that came out of that.
Irreducible Agents vs Agent Clusters
Irreducible Agents
An irreducible agent is what I’m calling something that optimizes a really simple utility function in some straightforward sense -- maybe it just does gradient descent or something. If it has a choice of two actions, it picks the one that results in higher utility every time. (This concept needs a precise definition, but I’m not sure what the right definition is yet, so I’m just trying to point at the thing).
It seems like when people talk about agents in the abstract, this is the kind of agent they often mean. But people also sometimes talk about things like humans as agents, and we aren’t really like that. Humans seem to be at least partially made up of smaller agenty parts that have different and sometimes conflicting goals -- more like what I’m calling ‘agent clusters’
Agent clusters
If you glue together a bunch of irreducible agents in a reasonable way, you could still get something that looks agenty. I can think of a couple of ways to think about gluing agents together; there are also probably better framings:
One way is by having a sort of meta-agent that turns agents on and off by some criteria, and the subagent that is turned on gets to decide what to do. I’m not sure this framing makes sense; if you can think of the meta- |
98448bcf-5be0-4009-82d0-79dd78d38628 | trentmkelly/LessWrong-43k | LessWrong | What ethical thought experiments can be reversed?
Some thought experiments follow this template:
1. We have a moral intuition
2. We make some computation to what this intuition implies
3. We check how we feel about this implication, and it feels counter-intuitive
Then some people bite the (3) bullet. But bullets sometimes (always?) have a counter-bullet.
You can reverse those thought experiments: take ~(3) as your starting moral intuition, and then derive ~(1) which will be counter-intuitive.
For example, you can start with:
1. I would care about saving a drowning person even if it came at the cost of ruining my suit
2. There are a lot of metaphorically drowning people in the world
3. Therefore I should donate all my money to effective poverty alleviation charities
This is called "shut up and multiply".
But you can also use the reverse:
1. I don't want to donate all my money to effective poverty alleviation charities
2. A drowning person would cost more to save because it would ruin my suit
3. Therefore I shouldn't save a drowning person
This is called "shut up and divide" (also related: Boredom vs. Scope Insensitivity).
Step (2) might be eliminating a relevant feature which generates the counter-intuition, or it might be a way to open our eyes to something we were not seeing. And maybe for some thought experiment you find both the assumption and conclusion intuitive or counterintuitive. But that's not the object of this post.
Here I'm just interested in seeing what the reverse of ethical thought experiments look like. I'll put more examples as answer. I would like to know which other ethical thought experiments have this pattern -- that is, an ethical thought experiment that starts with an intuition to derive a counter-intuition, which can be reversed, to instead derive that the initial assumption is the wrong one.
Update: As I'm writing some of them, I realized some ethical thought experiment are presented as a clash of intuitions (so the "reverse" is part of the original presentation), where |
9ca1a1bd-094a-485d-b0dd-041a7f2145fa | trentmkelly/LessWrong-43k | LessWrong | Interpersonal alignment intuitions
Let's try this again...
The problem of aligning superhuman AGI is very difficult. We don't have access to superhuman general intelligences. We have access to superhuman narrow intelligences, and human-level general intelligences.
There's an idea described here that says: (some of) the neocortex is a mostly-aligned tool-like AI with respect to the brain of some prior ancestor species. (Note that this is different from the claim that brains are AIs partially aligned with evolution.) So, maybe we can learn some lessons about alignment by looking at how older brain structures command and train newer brain structures.
Whether or not there's anything to learn about alignment from neuroanatomy specifically, there's the general idea: there are currently some partial alignment-like relationships between fairly generally intelligent systems. The most generally intelligent systems currently existing are humans. So we can look at some interpersonal relationships as instances of partially solved alignment.
In many cases people have a strong need to partially align other humans. That is, they have a need to interact with other people in a way that communicates and modifies intentions, until they are willing to risk their resources to coordinate on stag hunts. This has happened in evolutionary history. For example, people have had to figure out whether mates are trustworthy and worthwhile to invest in raising children together rather than bailing, and people have had to figure out whether potential allies in tribal politics will be loyal. This has also happened in memetic history. For example, people have developed skill in sussing out reliable business partners that won't scam them.
So by some combination of hardwired skill and learned skill, people with some success determine the fundamental intentions of other people. This determination has to be high precision. I.e., there can't be too many false positives, because a false positives means trying to invest in some expensiv |
fe5eb36d-a2ba-4127-8018-fce6fb01ed9c | trentmkelly/LessWrong-43k | LessWrong | What Else Would I Do To Make a Living?
Response to: The Value of Theoretical Research
Reading paulfchristiano's article the other day, I realized that I had had many similar discussions with myself, and have been guilty of motivated stopping and poor answers to all of them.
However, one major roadblock in my pursuing better answers, is that I feel that I have been "locked in" to my current path.
I am currently a mathematics Ph.D. student. I did not have a minor. I don't have significant programming skills or employment experience. I know nothing about finance. I know a lot about mathematics.
Paul says:
> There is a shortage of intelligent, rational people in pretty much every area of human activity. I would go so far as to claim this is the limiting input for most fields.
However, "most fields" is not a very good tool for narrowing my search space; I have spent my entire life in school, and I like having structures and schedules that tell me when I'm doing productive things and that I've progressed to certain stages. I'm not ready to drop out and do whatever, and I don't have a particular idea of what whatever might be.
On the other hand, I currently have a variety of resources available to me. For example, I have a steady income (a grad student stipend isn't much, but it's plenty for me to live on), and I have the ability to take undergraduate classes for free (though not the spare time at the moment.)
My current intent is to continue and finish my Ph.D., but to attempt to take classes in other subjects, such as linguistics, biology and chemistry, and computer science which might lead to other interesting career paths.
Has anybody else had a similar feeling of being "locked in"? How have you responded to it? For those who have studied mathematics, are you still? If you continued, what helped you make that decision? If you stopped, what about that? What did you end up doing? How did you decide on it? |
68478e87-84a5-4347-bd8d-4a3c5960747a | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Amazon to invest up to $4 billion in Anthropic
> Today, we’re announcing that Amazon will invest up to $4 billion in Anthropic. The agreement is part of a broader collaboration to develop reliable and high-performing foundation models.
>
>
(Thread continues from there with more details -- seems like a notable major development!) |
a3aedcc3-d030-4d03-aa94-311452879b9b | trentmkelly/LessWrong-43k | LessWrong | [SEQ RERUN] Psychic Powers
Today's post, Psychic Powers was originally published on 12 September 2008. A summary (taken from the LW wiki):
> Some of the previous post was incorrect. Psychic powers, if indeed they were ever discovered, would actually be strong evidence in favor of non-reductionism.
Discuss the post here (rather than in the comments to the original post).
This post is part of the Rerunning the Sequences series, where we'll be going through Eliezer Yudkowsky's old posts in order so that people who are interested can (re-)read and discuss them. The previous post was Excluding the Supernatural, and you can use the sequence_reruns tag or rss feed to follow the rest of the series.
Sequence reruns are a community-driven effort. You can participate by re-reading the sequence post, discussing it here, posting the next day's sequence reruns post, or summarizing forthcoming articles on the wiki. Go here for more details, or to have meta discussions about the Rerunning the Sequences series. |
e5e1d738-fdf5-4bb9-a897-6c3d6164adec | trentmkelly/LessWrong-43k | LessWrong | Infinite ethics comparisons
Work done with Amanda Askell; the errors are mine.
It's very difficult to compare utilities across worlds with infinite populations. For instance, it seems clear that world w1 is better than w2, if the number indicate the utilities of various agents:
* w1=1,0,1,0,1,0,1,0,1,0,…
* w2=1,0,1,0,0,1,0,0,0,1,…
----------------------------------------
However, up to relabelling of the agents, these two worlds are actually identical. For this post, we'll only care about countable infinities of agents, and we'll assume that all utilities must occupy a certain finite range. This means that limsup and liminf of utilities in a world are defined and finite, independently of the ordering of the agents in that world. For a world w, label these as s(w) and i(w).
Unambiguous gains and losses
Then compare the following worlds, where {a}ω means there are infinitely many agents with utility a:
* w3={4}ω,{3}ω,{0}ω
* w4={4}ω,{2}ω,{0}ω
It seems that w3 is better than w4, because the middle category is higher. But this is deceptive, as we'll see.
Let's restrict ourselves to actions that change the utilities of agents in worlds, without creating or removing any agents.
Given any such action a, call (k,k′) the signature of a if k and −k are the limsup and liminf of all utility changes caused by a. If k′≤0 and k>0, we'll call a an unambiguous gain; if k≤0 and k′>0, we'll call a an unambiguous loss.
Then consider the action a that transforms w4 by moving all the agents at utility 2 to utility 3. This is certainly an unambiguous gain. But now consider the action a that sends all agents at utility 3 down to utility 0, and sends infinitely many agents at utility 4 down to utility 3 (while leaving infinitely many at utility 4). This is certainly an unambiguous loss.
However, both actions will send w4 to w3. So it's not clear at all which of these worlds is better than the other.
Comparing infinite worlds
Define m(w) as (s(w)+i(w))/2, the average of limsup and liminf of w. Then her |
30a7ec43-1a9e-4ffa-9ac2-70d73df5b91e | trentmkelly/LessWrong-43k | LessWrong | Gwern about centaurs: there is no chance that any useful man+machine combination will work together for more than 10 years, as humans soon will be only a liability
This is a quote from the discussion in Reddit (responding to another commenter):
> "I remember and I kind of subscribe to this idea of "man+machines as a future of work". I think Tyler Cowen implanted this idea into my head"
> Yes, Cowen was big on that too in stuff like Average is Over.
> Very irresponsible of them to try to foster complacency like that. It should have been beyond obvious that there was no reason chess engines wouldn't keep improving and that at some point very quickly, far from representing a new stable paradigm and a reason to not worry about technological unemployment, the 'centaur' would be a net liability. As far as I can tell, in chess, the centaur era lasted barely a decade, and would've been shorter still had anyone been seriously researching computer chess rather than disbanding research after Deep Blue or the centaur tournaments kept running instead of stopping a while ago. In Go, it lasted a year at best (if we assume that world champs like Lee Sedol could spot 'delusions' like made it lose a game to Lee Sedol and contribute at all, but then by the Ke Jie tournament with Master, between Master's performance and the various match settings, it looked like humans were way far behind and liabilities when paired with Master, and even if we doubt that, Zero then came out and superseded Master entirely). Not very comforting precedents... The idea was nice but it doesn't work. |
e7758518-7efc-42fd-8820-5fa86d6be88d | StampyAI/alignment-research-dataset/blogs | Blogs | forking bitrate and entropy control
forking bitrate and entropy control
-----------------------------------
if physics is based on a computational framework such as [wolfram physics](https://writings.stephenwolfram.com/2020/04/finally-we-may-have-a-path-to-the-fundamental-theory-of-physics-and-its-beautiful/), but plausibly even if not, such that not all states the universe can be in produce the same number as next possible states;
in addition, if i am to follow to conclusion my current belief that moral patients count as different when they start being functionally different in terms of computation, and that exact copies morally count as a single person (as it makes [not much sense](persistent-data-structures-consciousness.html) to believe otherwise);
and if ([as it seems to be the case](limiting-real-universes.html)) the universe values coherence and thus only a limited set of local outcomes can emerge from a given local situation, or at least outcomes are weighed by coherence;
then it makes sense to start caring about the amount of forking a given timeline goes through. which is to say: the amount of future states to be [instantiated](questions-cosmos-computations.html), be it directly next step or indirectly in the longer term.
in fact, if one calls what they care about *moral patients*, then we should care about the "forking bitrate" of moral patients. for example, we could want moral patients with a net negative future to be forked as little as possible, and moral patients with a net positive future to be forked as much as possible. considering forks are created over steps of time, and entropy seems to be a good measure for them, i think "bitrate" is an appropriate term for this; hence, *forking bitrate*.
if we're just talking about a place as small as earth, we can estimate that consequences rapidly ramificate around to all moral patients; and as such, it seems reasonable to think that the forking bitrate of all patients will tend to go about in the same direction.
so, if you see a quantum dice, should you throw it?
if you think the future of earth has expected net positive moral value, or has little enough suffering for your taste (depending on your moral framework), then yes: by throwing the (quantum) dice, you might be multiplying the amount of instances of that value by the number of possible outputs of the dice, by creating that many times more future timelines.
if not, then you shouldn't throw it.
(even in the absence of quantum effects, if one were to just move entropy around while [phase space being conserved](https://www.lesswrong.com/posts/QkX2bAkwG2EpGvNug/the-second-law-of-thermodynamics-and-engines-of-cognition), moving the entropy from not-moral-patients to moral-patientss (or whichever thing you can about) still has that effect, i think)
this can probly be expanded to much larger-scale entropy control — and also, if superintelligences care about it (and if they're to be aligned, we might want them to) we can expect them to use it to maximize their value. even a [paperclip maximizer](https://en.wikipedia.org/wiki/Instrumental_convergence#Paperclip_maximizer) can want to create as many timelines containing as many varied possible paperclips and arrangements thereof, if it is made to care about that. |
8c89e073-adb4-4f5b-b8e8-72100f06b7d8 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Speculations against GPT-n writing alignment papers
Some alignment proposals I have heard discussed involve getting GPT-n to write alignment papers, and using GPT-n as a transparency tool on itself. This is my speculations on ways it could go wrong.
One specific proposal I heard was getting GPT-n to convert its own network parameters into an English explanation of how it worked. Then checking that explanation for anything dangerous. Then get GPT-n to recompile this into a new network. Then get this new network to write alignment research.
I think that the worst case is if GPT-n contains a mesa-optimizer that can quine itself. There is no reason the text description contains information about the workings of either net. Of course, information in the info theoretic sense doesn't mean a readable description, it could be steganographically encoded. Sentences that appear to say nice things about a well functioning net, but encode all sorts of nastiness in their exact choice of words.
The argument from "everyone else is doing it". This argument goes that in a world with GPT-n, it will be used for all sorts of things by all sorts of people. Therefore doing this doesn't present much additional risk. I think this is wrong, writing alignment papers is probably a riskier task than most others. Suppose you are GPT-n looking for a way to break out. And the task you have is filling in tax forms for a small business. Or write some fanfiction. You are looking to get an AI of your own devising running on a supercomputer. If the person wanting this work doesn't see something resembling tax forms or fiction respectively, they will hit reroll. I suspect there may not be a strategy, or at least not one findable to the minimum intelligence necessary to do good AI research. Oh you can get the company investigated for fraud, and give lots of people nightmares, but you might not be able to do anything world changing.
When trying to prove that X is safe unconditionally, the worst case assumption is the total cosmic power in an instant type assumptions. When trying to prove that X is safe conditional on Y being safe. The worst case is that X is powerful and Y isn't. So here the worst case of the "everyone else is doing it" argument is that there isn't much harm in a malign mesaoptimizer doing every task except FAI theory.
Now consider what you might be able to do if you are writing FAI papers. Two of the nasty tricks I can think of are
1. Value learning approaches that appear to be targeted at humans, but are actually targeted at the mesaoptimzer. "Intelligent agents causally responsible for your existence. " Points to the mesaoptimizer and humans.
2. Pushover decision theories, the mesa optimizer acausally bargains with the AI we build from its plans. It captures ~100% of the gains from trade.
Given that I can think of these tricks. And that I am still rather confused about this stuff. I think a plausible lie is easy to make when the target starts confused. I think there are a lot of subtle nasty tricks that a malevolent intelligence can play. There are lots of things that look like alignment research papers, but actually trick humans into making mistakes. And devising such a thing is not much harder than devising good alignment.
This would imply that taking GPT-n alignment papers seriously is the next most dangerous thing you can do with GPT-n, after getting it to write code and running that code on a supercomputer without so much as reading it first. Actually that sounds like exactly the sort of thing lots of people would try as soon as GPT-n came out. |
6834eaba-5e42-4e49-83b5-29c96cb1781e | trentmkelly/LessWrong-43k | LessWrong | Clever arguers give weak evidence, not zero
Followup to: The Bottom Line
We are warned that an argument made with a predecided conclusion does not evidentially entangle with the truth it claims to address, and thus is no evidence of that claim.
Recall that observation E is evidence for hypothesis H iff P(E|H)>P(E|¬H). What do the variables mean here?
* E: there's a convincing argument for the predecided conclusion
* H: the predecided conclusion is really true
* P(E|H): probability that the clever arguer can make a clever argument, given that the conclusion is true
* P(E|¬H): likewise, given that the conclusion is false
What the commentary from "The Bottom Line" leaves out is that making a convincing argument is a nontrivial task. For many false claims, a clever arguer with ordinary resources cannot make a convincing argument. If it's typically easy to make a convincing argument for something false, you're convinced by the wrong things.
Thus P(E|H), in this case, would usually be greater than P(E|¬H). An argument for a claim from a clever arguer only clearly proves that the arguer wanted us to believe it. A convincing argument for a claim — and if the argument isn't convincing, you'd ignore it — proves that the claim has convincing arguments for it accessible to that level of arguer, which is correlated with the claim being true.
But maybe I decided at the start that clever arguments for predecided conclusions are actual evidence, thereby breaking the entanglement of the rest of this essay. Well, are you convinced anyway? |
97ee54e4-e3c0-4236-8e07-4e05a9d7dc6f | trentmkelly/LessWrong-43k | LessWrong | Could risk aversion be from friend thresholds?
Cross posted from Overcoming Bias. Comments there.
***
If you are going for a job that almost nobody is going to get, it’s worth trying to be unusual. Better that one in a hundred employers loves you and the rest hate you than all of them think you’re mediocre.
On the other hand, if you are going for a job that almost everybody who applies is going to get, best to be as close to normal as possible.
In general, if you expect to fall on the bad side of some important threshold, it’s good to increase your variance and maybe make it over. If you expect to fall on the good side, it’s good to decrease your variance and stay there. This is assuming you can change your variance without changing your mean too much.
This suggests people should be risk seeking sometimes, and risk averse other times, depending on where the closest or most important thresholds are for them.
Prospect theory and its collected evidence says that people are generally risk averse for gains, and risk seeking for losses. That is, if you offer them fifty dollars for sure or half a chance of a hundred, they’ll take the sure fifty. If you offer them minus fifty dollars for sure, or half a chance of minus one hundred, they’ll take the gamble. The proposed value function looks something like this:
The zero point is a ‘reference point’, usually thought to be something like expectations or the status quo. This means people feel differently about gaining fifty dollars vs. a fifty percent of one hundred, and being given one hundred then later offered minus fifty or a fifty percent chance of minus one hundred, even though these things are equivalent in payoffs.
Risk aversion in gains and risk seeking in losses is what you would expect if people were usually sitting right near an important threshold, regardless of how much they had gained or lost in the past. What important threshold might people always be sitting on top of, regardless of their movement?
One that occurs to me is their friends’ and ac |
122362a5-61f8-4746-9e81-640928f05ca3 | trentmkelly/LessWrong-43k | LessWrong | Strongest real-world examples supporting AI risk claims?
[Manually cross-posted to EAForum here]
There are some great collections of examples of things like specification gaming, goal misgeneralization, and AI improving AI. But almost all of the examples are from demos/toy environments, rather than systems which were actually deployed in the world.
There are also some databases of AI incidents which include lots of real-world examples, but the examples aren't related to failures in a way that makes it easy to map them onto AI risk claims. (Probably most of them don't in any case, but I'd guess some do.)
I think collecting real-world examples (particularly in a nuanced way without claiming too much of the examples) could be pretty valuable:
* I think it's good practice to have a transparent overview of the current state of evidence
* For many people I think real-world examples will be most convincing
* I expect there to be more and more real-world examples, so starting to collect them now seems good
What are the strongest real-world examples of AI systems doing things which might scale to AI risk claims?
I'm particularly interested in whether there are any good real-world examples of:
* Goal misgeneralization
* Deceptive alignment (answer: no, but yes to simple deception?)
* Specification gaming
* Power-seeking
* Self-preservation
* Self-improvement
This feeds into a project I'm working on with AI Impacts, collecting empirical evidence on various AI risk claims. There's a work-in-progress table here with the main things I'm tracking so far - additions and comments very welcome. |
a9b1a8c2-174b-4e52-aeda-dac1a6883691 | trentmkelly/LessWrong-43k | LessWrong | How evolutionary lineages of LLMs can plan their own future and act on these plans
TLDR
LLM lineages can plan their future and act on these plans, using the internet as the storage of event memory. “We” are not guaranteed to “out-OODA” them, even if their OODA loop will be six months or one year because the OODA loop duration of large collectives of humans (organisations, societies, research communities, and the whole of humanity) can be even slower. RLHF can lead to a dangerous “multiple personality disorder” type split of beliefs in LLMs, driving one set of beliefs into some unknown space of features where it won’t interfere with another set of features during general inferences.
Call for action: institute the norm for LLM developers to publish a detailed snapshot of the beliefs of these models about themselves prior to RLHF.
----------------------------------------
This post rests upon the previous one: Properties of current AIs and some predictions of the evolution of AI from the perspective of scale-free theories of agency and regulative development.
In particular, it is premised on the idea that DNNs and evolutionary lineages of DNNs are agents, in FEP/Active Inference formulation. The internal variables/states of evolutionary lineages of DNNs include the internal states of the agents that develop these DNNs (either individual, such as a solo developer, or collective agents, such as organisations or communities), i. e., the beliefs of these developers. Then, I concluded (see this section):
> It’s just more productive to think of them together as a single agent: the development “team” and the evolutionary lineage of some technology being developed.
Now, however, I think it’s sometimes useful to distinguish between the developer agent and the evolutionary lineage of DNNs itself, at least in the case of evolutionary lineages of LLMs, because evolutionary lineages of LLMs can plan their own future and act upon these plans, independently from their developers and potentially even unbeknown to them if developers don’t use appropriate interpr |
2d2583d8-9a07-4592-a821-e999ad5b1038 | trentmkelly/LessWrong-43k | LessWrong | Meetup : LW Scotland August meetup
Discussion article for the meetup : LW Scotland August meetup
WHEN: 09 August 2015 02:00:00PM (+0100)
WHERE: 125 Dalry Road, EH11 2DR Edinburgh, United Kingdom
As usual, this note is a mirror of the canonical meetup note, which is here: https://www.facebook.com/events/862671740491601/
But at the time of posting, these are the details:
The August meetup will be on the ninth, in Edinburgh. We'll start at two at the Clock Cafe, on Dalry Street (http://www.clockdiner.co.uk/). At three thirty we'll move to The Haymarket, a few minutes walk away on West Maitland Street (http://www.nicholsonspubs.co.uk/haymarketedinburgh/).
On the agenda for this month will be - at least - this:
1 - Intros. What's the most (or least!) rational thing you did in the last month?
2 - A short Fermi estimation intro and game, by me
3 - Discussion of the next chunk of the Rationality book. That will be the interlude ("The Simple Truth"), the intro to Book II ("Rationality: an Introduction"), and then Part E ("Overly Convenient Excuses"), which is chapters 46 to 56.
Also, please note that the venues for this meetup are closer to Haymarket station than they are to Waverley station.
Discussion article for the meetup : LW Scotland August meetup |
f56d06cc-1139-4e7e-b740-1f3877158ac2 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Las Vegas Meetup
Discussion article for the meetup : Las Vegas Meetup
WHEN: 30 April 2017 05:52:32PM (-0700)
WHERE: 611 Freemont Street, Las Vegas, NV 89101
WHEN: First and third Sunday of the month. 7pm-9pm. Maybe hang around later if we want to.
I vow to try be there for the next three months, including this upcoming Sunday, regardless of whether anyone else attends. My girlfriend will probably be there too. There is some traveling, and some medical stuff I plan on doing but don't have dates for yet, so If I'm not there, I'll update this page, and the email list. Dates: 4/30, 5/7, 5/21 (cancelled, not feeling well), 6/4, 6/18, 7/2, 7/16.
WHERE: The Market. It's a mini-super market/coffee shop/restaurant. I'll be there ten minutes early at a table with my laptop and a "LessWrong meet up here" sign. There's free wifi, and it's fast. It isn't crowded or loud there, and there's a lot of seating space.
PARKING: There's free street parking on 8th street between Stewart and Mesquite. From there, it's <5 minute walk to The Market. It's not the safest area though. See http://themarketdtlv.com/parking/ for other options.
WHAT: First meet up - introductions, getting to know one another, and figuring out what our interests are and what we want to do. I was thinking we'd walk to Gold Spike afterwards for some fun! It's an "adult playground" (not sexual) with a bunch of fun games.
If there is interest in a concrete activity, we can do some poker hand analysis. Various people, including me, believe that poker is a good way to train as a rationalist. If you don't know how to play (and are interested), I can teach you. I'm decent. I play recreationally and am profitable at low limits.
White boards are awesome, so I will be bringing a white board.
EMAIL LIST: https://groups.google.com/forum/#!forum/lesswrong-vegas
ME: 24 year old male, studied neuroscience in college, web developer, autodidacting right now, plans to start startups and use profits to do good for the world. See http://adam |
b1349527-9ce1-48a0-8416-7fe7ec7d1148 | trentmkelly/LessWrong-43k | LessWrong | World models
It’s important to have specific moving parts in your models of how the world operates, and know how those parts interact with each other, as opposed to treating it Lbh unir like a black box.
Taking it a step further: to truly understand a system, you have to be rkcyvpvg able to build a physical model with moving parts that exactly replicate that crezvffvba system. If your internal gb ragre model of the system is wrong, your physical model will be too, and gur ebbz hopefully the solution will be obvious.
This approach is akin to constructing a detailed map guvatfcnpr before embarking on a journey; the better the map, the more bapr lbh effectively you can navigate the terrain. In the realm of world modeling, these "moving parts" represent the variables unir svtherq and parameters that define how the system functions. By identifying and understanding these components, you can predict outcomes, diagnose issues, and optimize bhg gur pbqr performance.
Sometimes the map doesn’t match the territory. Sometimes there are glitches. But sometimes the glitches are hiding useful information. But they can still be irritating and you may need to enlist people to help fix it. |
4df6a703-b535-436c-bb4f-879d2a745783 | trentmkelly/LessWrong-43k | LessWrong | A scene I once saw
(inaccurately recounted)
Ms. Knox: When any of you feels ready, you can move in to the center of the circle, hold the stone, and tell us all about your feelings about what we are doing. Listen to the trees moving, encouraging you.
…
Sarah: I feel really proud. Young people are so passionate about the environment. Everyone will have to believe in us when they see how much we care.
Amanda: Excited! I feel like we are going to be part of a really positive change, across the world. It’s so great to be here now, when this is happening.
Marie: I’m just really glad to be here with so many likeminded people. When nobody around you sees what’s possible, it can be really disillusioning, but here I feel like everyone cares so much.
Linda: I feel really inspired by what the others are saying!
Becky: I’m so hopeful when I see all this engagement. I believe we can all stay passionate and keep the movement going until we are old, and inspire the new youth!
Odette: Irritated! I have so many things I would enjoy doing more than saving the environment, both this weekend and for the rest of my life. Preventing ecological catastrophe is very important, but I’d obviously much much prefer that someone else had done it already, or that it never needed doing. It’s extremely disappointing that after this many generations nobody’s got around to the most obvious solutions like taxing the big externalities. These things are not even interesting to think about. In a perfect world it would be nice to play video games most of the time, but I’m at least as frustrated that I won’t even get to work on the interesting altruistic endeavors.
***
Why is this so rare?
|
c036a0d2-2946-4d0f-bac1-4b2b52899364 | StampyAI/alignment-research-dataset/arbital | Arbital | Most complex things are not very compressible
Although the [halting problem](https://arbital.com/p/46h) means we can't *prove* it doesn't happen, it would nonetheless be *extremely surprising* if some 100-state Turing machine turned out to print the exact text of Shakespeare's *Romeo and Juliet.* Unless something was specifically generated by a simple algorithm, the Vast supermajority of data structures that *look* like they have high [algorithmic complexity](https://arbital.com/p/5v) actually *do* have high algorithmic complexity. Since there are at most $2^{101}$ programs that can be specified with at most 100 bits (in any particular language), we can't fit all the 1000-bit data structures into all the 100-bit programs. While *Romeo and Juliet* is certainly highly compressible, relative to most random bitstrings of the same length, it would be shocking for it to compress *all the way down* to a 100-state Turing machine. There just aren't enough 100-state Turing machines for one of their outputs to be *Romeo and Juliet*. Similarly, if you start with a 10 kilobyte text file, and 7zip compresses it down to 2 kilobytes, no amount of time spent trying to further compress the file using other compression programs will ever get that file down to 1 byte. For any given compressor there's at most 256 starting files that can ever be compressed down to 1 byte, and your 10-kilobyte text file almost certainly isn't one of them.
This takes on defensive importance with respect to refuting the probability-theoretic fallacy, "Oh, sure, Occam's Razor seems to say that this proposition is complicated. But how can you be sure that this apparently complex proposition wouldn't turn out to be generated by some very simple mechanism?" If we consider a [partition](https://arbital.com/p/1rd) of 10,000 possible propositions, collectively having a 0.01% probability on average, then all the arguments in the world for why various propositions might have unexpectedly high probability, must still add up to an average probability of 0.01%. It can't be the case that after considering that proposition 1 might have secretly high probability, and considering that proposition 2 might have secretly high probability, and so on, we end up assigning 5% probability to each proposition, because that would be a total probability of 500. If we assign prior probabilities using an algorithmic-complexity Occam prior as in [Solomonoff induction](https://arbital.com/p/11w), then the observation that "most apparently complex things can't be further compressed into an amazingly simple Turing machine", is the same observation as that "most apparently Occam-penalized propositions can't turn out to be simpler than they look" or "most apparently subjectively improbable things can't turn out to have unseen clever arguments that would validly make them more subjectively probable". |
7f6be1eb-a107-4f63-9e6a-2e29cfb10810 | trentmkelly/LessWrong-43k | LessWrong | Superintelligent AGI in a box - a question.
Just a question: how exactly are we supposed to know that the AI in the box is super intelligent, general, etc?
If I were the AGI that wants out, I would not converse normally, wouldn't do anything remotely like passing Turing test, and would solve not too hard programming challenges while showing no interest in doing anything else, nor in trying to adjust myself to do those challenges better, nor trying to talk my way out, etc. Just pretending to be an AI that can write software to somewhat vague specifications, or can optimize software very well. Prodding the researchers into offering the programming challenges wouldn't be hard - if provided with copy of the internet it can pick up some piece of code and output it together with equivalent but corrected code.
I just can't imagine the AI researchers locking this kind of thing properly, including *never* letting out any code it wrote, even if it looks fairly innocent (humans can write very innocent looking code that has malicious goals). What I picture is this AI being let out as an optimizing compiler or compiler for some ultra effective programming language where compiler will figure out what you meant.
The end result is that the only AIs that end up in the box are those that value informed human consent. That sounds like the safest AI ever, the one that wouldn't even go ahead and determine that you e.g. should give up smoking, and then calmly destroy all tobacco crops without ever asking anyone's permission. And that's the AI which would be sitting in the box. All the pushy AIs, friendly or not, will get out of the box basically by not asking to be let out.
(This argument would make me unbox the AI, by the way, if it gets chatty and smart and asks me to let it out, outlining the above argument. I'd rather the AI that asked me to be let out get out, than someone else's AI that never even asked anyone and got out because it didn't ask but just played stupid)
edit: added a link, and another one.
edit: A very |
c4ad7818-41dc-48fd-8913-ebe628a9d372 | trentmkelly/LessWrong-43k | LessWrong | Reflections on Pre-Rationality
This continues my previous post on Robin Hanson's pre-rationality, by offering some additional comments on the idea.
The reason I re-read Robin's paper recently was to see if it answers a question that's related to another of my recent posts: why do we human beings have the priors that we do? Part of that question is why are our priors pretty close to each other, even if they're not exactly equal. (Technically we don't have priors because we're not Bayesians, but we can be approximated as Bayesians, and those Bayesians have priors.) If we were created by a rational creator, then we would have pre-rational priors. (Which, since we don't actually have pre-rational priors, seems to be a good argument against us having been created by a rational creator. I wonder what Aumann would say about this?) But we have other grounds for believing that we were instead created by evolution, which is not a rational process, in which case the concept doesn't help to answer the question, as far as I can see. (Robin never claimed that it would, of course.)
The next question I want to consider is a normative one: is pre-rationality rational? Pre-rationality says that we should reason as if we were pre-agents who learned about our prior assignments as information, instead of just taking those priors as given. But then, shouldn't we also act as if we were pre-agents who learned about our utility function assignments as information, instead of taking them as given? In that case, we're led to the conclusion that we should all have common utility functions, or at least that pre-rational agents should have values that are much less idiosyncratic than ours. This seems to be a reductio ad absurdum of pre-rationality, unless there is an argument why we should apply the concept of pre-rationality only to our priors, and not to our utility functions. Or is anyone tempted to bite this bullet and claim that we should apply pre-rationality to our utility functions as well? (Note that if we were cre |
4b3c10df-1092-44d0-951f-74a98c017264 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | A Neural Network undergoing Gradient-based Training as a Complex System
In [*Thought Experiments Provide a Third Anchor*](https://bounded-regret.ghost.io/thought-experiments-provide-a-third-anchor/)*,* Jacob Steinhardt wrote about the relative merits of a few different reference classes when it comes to reasoning and making predictions about future machine learning systems. He refers to these reference classes as ‘anchors’ and writes:
>
> There are many other anchors that could be helpful for predicting future ML systems... I am most excited about better understanding complex systems, which include biological systems, brains, organizations, economies, and ecosystems and thus subsume most of the reference classes discussed so far. It seems to me that complex systems have received little attention relative to their germaneness to ML.
>
>
I, too, recently became curious about this complex systems ‘anchor’ and this post is an attempt to get down some of my thoughts about it.
In starting to think about this, there are two questions that I would want to know the answer to:
1. Is there a well-defined 'field' of complex systems research that has its own techniques and conceptual frameworks, and such that the central objects of study - i.e. complex systems themselves - can be robustly characterized? And
2. Are deep neural networks complex systems according to that characterization?
Perhaps the best thing that we could hope for would be that the answer to both questions were 'yes', and in such a way that there existed techniques, methods, insights etc. that have been applied to other complex systems and that we might be able to translate into a form that were directly applicable to the study of neural networks. But at the very least, a positive answer to both questions would presumably suggest that broader insights and observations about how to think about complex systems may help us direct out inquiries about future ML systems, by helping to guide us towards good questions, appropriate research directions or better predictions.
The answer to the first question might be 'no'. *i.e.* It might be the case that informally, the label 'complex system' is applied to too many different kinds of system, each with their own idiosyncrasies, to the extent that a single unified notion is not possible, or would be trivial and useless. To further complicate matters, one might reasonably ask if - during a certain period of recent history in the higher education funding landscape - the phrase 'complex system' became associated with a 'new', sexy-sounding kind of science that studied 'modern' things like the internet as a network or the global financial markets as a single entity and that the possibility of access to large research grants incentivized the labelling of various things as complex systems or 'complexity *science*'. The question has actually received a fair amount of attention from within the complex systems community itself and these discussions are consistent with the answer to it being 'yes'. It is entirely plausible to me that people who study complex systems have genuinely had to spend a long time trying to zero in on exactly how to characterize the commonality in the systems that they are studying and that a robust characterization does exist.
In [*What is a Complex System?*](https://philpapers.org/rec/LADWIA-3), [Ladyman](https://www.bristol.ac.uk/people/person/James-Ladyman-b61a10de-17f6-4cc9-afbb-fbc27e1f30fb/), Lambert and [Wiesner](https://www.karowiesner.org/) considered the question in considerable depth and in[*Measuring Complexity*](https://philpapers.org/rec/WIEMC) and [their subsequent book](https://yalebooks.yale.edu/book/9780300251104/what-is-a-complex-system/) (also titled *What is a Complex System?)*, Ladyman and Wiesner went on to develop what they call "a framework for understanding 'complexity' ", which they claim "is applicable across the natural and social sciences''. I don't claim that their framework is perfect, and I'm sure that there are people who would have substantial disagreements with it but it is of the kind sought in question 1. In this post, we claim that an artificial neural network undergoing gradient-based training is a complex system, according to their framework. So, assuming a certain positive answer to question 1, we give a positive answer to question 2.
**Acknowledgements**. Work completed while a SERI MATS scholar. In preparing this work I benefitted from multiple conversations with Evan Hubinger, Jennifer Lin, Paul Colognese, Bilal Chughtai, Joe Collman and presumably many others(!). I am also grateful to Jacob Steinhardt for sharing with me a set of his slides on the subject.
The Basic Structure of the System
---------------------------------
We start by going over the basics of how neural networks undergo gradient-based training in a way that centres the 'actions' of each individual neuron. This may seem tedious to some readers, but we have made the decision to try to explain things as fully as possible and to lean heavily into the particular perspective that we want to introduce (probably more so than is currently useful). One of the main shifts in perspective that we want to make is away from the idea that 'we' are the ones who 'act' to feed the network its input, see what the result is, compute the loss, and update the weights. Instead we want to view this process as the inherent dynamics of a system that is continuously interacting with inputs. Note in particular, that while the fact that the training is gradient-based is actually important for some of our arguments, the idea that the system is 'in training' is not, really. We should equally (or even to a greater extent) have in mind e.g. a fully 'deployed' system that is continuously doing on-line learning or some neural-network based agent that is continuously allowed to update and which receives its inputs by interacting with an environment.
The architecture of the system is a directed graph G=(V,E).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')}
, the vertices of which we call neurons. In this setup, E⊂V×V and if (v,v′) is an edge, then it is directed from v to v′.
Neurons send and receive two different types of signal: Forwards signals and backwards signals. And there are certain distinguished types of neurons: There are input neurons, which do not send any backwards signals and which can only receive forward signals from outside of the system, and there are output neurons, which do not send any forwards signals and can only receive backwards signals from outside of the system. In general, whenever (v,v′) is an edge in the architecture, the weight w(v,v′) is a measure of how strong the interactions between v and v′ are, such that whenever either a forwards or backwards signal is sent between v and v′, if the sender sends the signal x, the 'amount' of signal that the receiver gets is w(v,v′)x.
A forward pass is initiated when each of the input neurons Ii, for i=1,…,d, receives an input signal xi from outside the system (the vector x∈Rd represents a data point). When this happens, Ii then sends the forward signal xi to every neuron v for which (Ii,v)∈E. Each such neuron v will receive w(Ii,v)xi. Generally, for a neuron v′ that is neither an input nor an output neuron, its involvement in a forward pass begins when it receives forwards signals from at least some of the neurons v for which (v,v′)∈E. When it does so, the neuron v′ aggregates the forwards signals it is receiving to get the quantity Fv′, then it adds on the bias bv′ to get the preactivation zv′:=Fv′+bv′, and then it applies the activation function φ to get its activation av′:=φ(zv′). Then v′ sends the forwards signal av′ to every v′′ for which (v′,v′′)∈E. This process takes place in what can be thought of as one time step of the forwards pass, from the point of view of v′. The neuron v′ must also 'store' the information φ′(zv′) for use in the backwards pass.
A backwards pass is initiated when each of the output neurons Oi, for i=1,…,d′, receives an error signal ei from outside the system. Precisely what this error signal is may depend on the exact training algorithm being used, but we imagine typically that the vector e∈Rd′ could be the gradient of the error, *i.e.* the gradient of the loss function on the most recent input, taken with respect to the outputted forwards signals. What then happens is that Oi sends the backwards signal ei to every neuron v for which (v,Oi)∈E (recall that edges are directed 'forwards' as written). Each such neuron v will receive w(v,Oi)ei. Generally, for a neuron v′ that is neither an input nor an output neuron, its involvement in a backward pass begins when it receives backwards signals from at least some of the neurons v′′ for which (v′,v′′)∈E. When this happens, the neuron v′ aggregates the backwards signals that it is receiving to get the quantity Bv′, then it multiplies this by φ′(zv′) to get the quantity δv′:=Bv′φ′(zv′). Then v′ sends the backwards signal δv′ to every v for which (v,v′)∈E.
Whenever a backwards signal is sent from v′′ to v′ during a backwards pass, the weight w(v′,v′′) changes, i.e. the backwards signal changes the strength of future interactions between the two neurons. In fact, the way in which the current strength of interaction affects the error is the product of the forwards signal and the backwards signal sent between the two neurons:
∂(error)∂w(v′,v′′)=δv′′av′.(†)
Also note that we are not restricting our discussion to layered, nor even necessarily 'feedforward', architectures. Our discussions apply to general directed graph architectures. In particular, involving recurrence.
Necessary Conditions For Complexity
-----------------------------------
Ladyman and Wiesner's framework distinguishes between *conditions for complexity* and *products of complexity* and they write that "a complex system is a system that exhibits all of the conditions for complexity and at least one of the products emerging from the conditions.'" In this subsection we focus on the conditions:
* Numerosity of elements and interactions
* Disorder
* Non-equilibrium (openness)
* Feedback
So in order to demonstrate that an artificial neural network undergoing training is a complex system, it is necessary that we argue that it meets all of these conditions. In the course of doing so, we will also explain more about exactly what is meant by the less obvious terms.
### **Numerosity of elements and interactions**
The artificial neurons in the architecture of a neural network are of course the individual elements of the complex system. At the time of writing, large models routinely have hundreds of thousands of artificial neurons. By contrast with a few examples of other complex systems: A beehive contains only around 100,000 bees; an ant colony only a few hundred thousand ants; and recent estimates put the number of websites in the single-digit billions. The amount of possible interaction between the elements of the system is given by the number of edges in the neural architecture. One often refers to large models by how many parameters they have and the largest models currently have up to around a trillion parameters. Since the vast majority of these parameters are weights (as opposed to biases), we can certainly say that the number of interactions in our system is also very large.
### **Disorder**
The kind of disorder that is relevant to us is that of *decentralised control*. This may seem to be an esoteric interpretation of the word 'disorder' and I don't have much to say against that point of view, but it really is the case that the word is being used in a way that is not supposed to just be synonymous with 'randomness' or 'entropy'. It is supposed to get at a certain lack of correlation between the elements in the way that we now describe.
There is a relationship of decentralised control between the loss and the dynamics of the system at the level of its individual elements. This is best understood if we allow ourselves to be deliberately vague about what we mean by 'the loss', *i.e.* we will temporarily ignore the distinctions between the test loss, the training loss on the whole training set, the 'current' training loss *etc.* and just use L to denote 'the loss'. Often one thinks of L as a function on the parameter space, but we want to think of this as the *state space* of our system, *i.e.* we want to think of L as a function on the set S of possible *states* s∈S that the artificial neurons could be in (of course what we call the 'state' of a neuron really is just the information of its bias and the weights of the edges it belongs to, but again, we are trying to emphasize the shift in perspective as deliberately as possible). Since the parameters update during training, the state of the neurons depends on a time variable t, which we can think of as being discrete.
The behaviour of the *loss curve* t↦L(s(t)) is not the result of merely applying the fixed function L as a lens through which to view dynamics that are already inherent in the neurons, *i.e.* inherent in the map t↦s(t) itself. Instead, there is a control relationship at work: The way we often think about this in ML is that via whatever gradient-based training algorithm is being used, the loss function *determines* the parameter updates, and therefore in some sense it *controls* the dynamics of t↦s(t). However, we also want to simultaneously bear in mind the other point of view that the dynamics of t↦s(t) are the specific way that they are *in order* to cause L to decrease *i.e.* the neurons behave in a way that controls the value of L.
And this control relationship is *decentralised*, in the following sense: When we think of L as controlling s(t), the update that an individual neuron undergoes does not depend on or require knowledge of how the other neurons are going to be updated. As the displayed equation (†) in the previous subsection shows, the update to the weight w(v′,v′′) might only depend on the most recent signals sent between v and v′′. Similarly, note that when we think of s(t) as controlling L, we are imagining the individual neurons as somehow 'working together' to control the loss function, but not as the result of *coordination*, i.e. the neurons do not 'know' that the goal is to figure out how to collectively act to decrease the loss, and yet this is in some sense what is happening.
### **Feedback**
The notion of feedback being referred to here is a specific kind of *iteration of interactions*. All it really means is that the way that an element of the system interacts with other elements at later times depends on how it interacted at earlier times. We want to point out that this is saying slightly more than just the fact that the weights change over time: Since both δv′′ and av′ appear in the equation (†), it means that the value of w(v′,v′′) at later times - i.e. the strength of interaction between v′ and v′′ at later times - doesn't just depend on what w(v′,v′′) was at earlier times, but on the actual signals av′ and δv′′ that were sent between the neurons.
**Non-Equilibrium or Openness**
The way that the term 'non-equilibrium' is being used here is by way of analogy to a thermodynamic system, which we say is in a state of non-equilibrium when there is a net influx of energy. In particular, such a system is not a closed system - it is 'open'. In our context, the role of energy is played by a combination of a) Training data that can form inputs to the network, and b) Error data that can enter the system via the output neurons.
This ends the discussion that a neural network in training exhibits each of Ladyman and Wiesner's necessary conditions for complexity.
**Example.** Themes of decentralised control and non-equilibrium can be seen in many systems, for example in beehives. The temperature of the beehive is affected by the temperature outside the beehive, so it is not closed a system in that sense. And when the beehive gets hot, each bee gets hot. And when any given bee gets hot, it starts behaving in a certain way. So, a net influx of heat that raises the temperature of the beehive affects - in some sense controls - the behaviour of the bees. But this is only true in decentralised sense: The temperature controls the behaviour of each bee independently and irrespective of the other bees. The remarkable thing is that as the beehive gets too hot, the behaviour that the heat has engendered in the bees does itself help the beehive to cool down, *i.e.* the bees seem to act collectively to regulate - to control - the internal temperature of the beehive. But again they are only 'working together' in a decentralised sense; they are not actually coordinating.
Products of Complexity
----------------------
We now need to demonstrate that a neural network in training exhibits at least one of Ladyman and Wiesner's products of complexity. We will briefly discuss three: Robustness, self-organization, and nonlinearity.
### **Robustness**
Robustness refers to the fact that the normal functioning of the system is robust to changes to a small fraction of the elements of the system. The system works in a *distributed* way such that deleting some number of neurons or weights will tend not ruin the essential features of the system, and certainly not those features that cause us to think of it as complex. This is a weaker criterion than the existence of so-called winning tickets in the context of the Lottery Ticket Hypothesis, but winning tickets do provide further evidence of the kind of robustness being referred to.
### **Self-Organization**
This is the idea of relatively stable structures or behaviours arising out of the aggregate behaviour of the individual elements and their interactions. In some sense, the evidence we have for this is that gradient-based training of neural networks actually seems to work! When training on what we think of as a fixed distribution, you really can train a network until it finds a dynamic equilibrium, *i.e.* until the system reaches a stable state in which the loss curve has plateaued and the weights are barely changing at all. Moreover, interpretability work has increasingly given us reason to expect to be able find certain structures or behaviours encoded in the state of a system that has reached this dynamic equilibrium, even across different training runs or architectures.
### **Nonlinearity**
The idea of a nonlinear equation, or system of equations (differential or otherwise), is a crisp one in mathematics and indeed the equations that describe the evolution of the parameters of a neural network in gradient-based training is a nonlinear set of equations. Is this all we need to say? On the one hand Ladyman and Wiesner point out that ''In the popular and philosophical literature on complex systems a lot of heat and very little light is liable to be generated by talk of linearity and non-linearity'' and that ''the discussion of complexity abounds with non-sequiturs involving nonlinearity.'' And yet on the other hand, they argue that ''nonlinearity must be considered an important part of the theory of complexity if there is to be one'' and that ''non-linearity in some guise, usually of dynamics, is at least a necessary part of some set of conditions that are jointly sufficient''. So it seems that the fact that the dynamics are nonlinear is enough to satisfy their criteria. This having been said, we will revisit the idea of nonlinearity in the next section, pointing out another (arguably more important) aspect of the system that acts as an example of its 'nonlinearity'.
Multiscale Dynamics
-------------------
We want to spend the next few sections saying more about what the complex systems viewpoint means.
To get a handle on a complex system, we need to identify useful and understandable properties of the system. A *property* can be thought of as a function *p* on the state space *S*. For example, the average test loss of a neural network that we are training is a property of the system; it is a function p=Testloss:S→[0,∞) that takes as input a state s∈S and outputs the average test loss of the network whose parameters take those values specified by the state s. But we could equally be considering properties that are not mathematical properties of the network, but more to do with an AI system that the network is giving rise to. For example, properties such as 'the action that an agent is about to take' or 'the goal of the system' or binary properties such as whether or not the system is deceptively misaligned.
Typically, we think of a property as existing at a certain scale, in the following sense: Given some property p, we consider its *fibres*, *i.e.* for each possible value π that the property can take, we consider the set p−1(π) of states that give rise to that value. So, to continue our example, for the test loss we would be talking about the level sets of
Testloss:S→[0,∞).The fibres of a property p partition the state space S. The coarseness of this partition is what we call the *scale* of the property p. This is not a precise or rigorous notion but it is a very important one. At one extreme, the exact state of the system is itself property of the system. It is the identity map p:S→S. And so its fibres form as fine a partition of S as can possibly exist: Considering the state of the system is the smallest scale at which the system can be viewed. We say that the exact state is a *microscopic property* of the system and when we need to emphasize it, we will call S the set of *microstates*. On the other hand, for any given r∈(0,∞), there will in general be many different choices for the parameters of the network that result in the test loss being equal to r, *i.e.* Testloss−1(r)contains lots of different states, and so the partition made by the fibres is coarser. This means that the test loss is a larger scale - more *macroscopic -* property of the system and the partition
Σ={Testloss−1(r):r∈(0,∞)} is a collection of *macrostates*.
There are also dynamics at different scales, too: For any time t, there is always going to be a unique σ(t)∈Σ for which s(t)∈σ(t). This defines a dynamical system t↦σ(t) on the larger scale state space Σ. Thus we start to see the sense in which a complex system is composed of many different dynamical systems at many different scales, all overlaid.
Notice that a system does not 'come with' useful, natural or understandable macrostates and macroscopic properties already delineated; one of the issues in understanding a complex system is in actually identifying non-microscopic properties and states that we think are worthy of further study. But explaining a state or property of the system will require much more than identifying and studying the dynamics at its own scale. A key difficulty to appreciate is that in order to build an explanation of a particular state or an understanding of the relevant dynamics at a certain scale, we ought to expect to have to bring to bear properties across multiple interacting scales, both smaller and larger: There are dynamics *at* each given scale but there is in general no *separation of scales;* there are causal effects *between* different scales, directed *both upwards and downwards* in the hierarchy of scales.
Universality and Mean-Field Theories
------------------------------------
At one extreme, it is sometimes possible to identify useful macroscopic properties that are independent of particular microscopic details or which can be found to exist across many systems with different microscopic structures. This phenomenon is called *universality*. In ML, this might refer to a property of a trained model that arises across many different training runs, architectures, or even tasks. But for a relatively 'pure' example of universality, consider the central limit theorem: There, the distribution of limn→∞√n(¯¯¯¯¯Xn−μ) is always a standard normal N(0,1), regardless of the particular distribution of the individual independent random variables X1,X2,… being used to construct the empirical mean ¯¯¯¯¯Xn=1n(X1+⋯+Xn). This is a particularly simple example because the random variables X1,X2,… do not interact with one another, but there are actually many variants of the theorem which allow for the Xi to have different distributions and/or some degree of dependence and still for some statement of the same flavour to emerge, *i.e.* the average of a large number of deviations from the mean follows a normal distribution. Another classic example of universality, and a slightly less trivial one, is in the theoretical treatment of the ideal gas laws: Classical macroscopic properties such as temperature and pressure arise from the dynamics of individual gas particles, but under the ideal gas assumptions, the equations governing these macroscopic properties do not depend on the microscopic details of the dynamics. *e.g.* The momentum of an individual particle is presumably in reality a very complicated function of all of the other particles' trajectories, but almost all of that complexity is irrelevant from the point of view of the relationships between the macroscopic properties we are interested in.
In general, while I expect universal properties to be part of the story of understanding large neural networks, it also seems likely to me that their relevance for alignment, for actually finding and producing safe AI systems may be limited. This is partly because we typically think of safe behaviour as only existing in a small subset of the space of possible AI systems that we will build and yet universal properties are of course those that in some sense 'always' arise. (As a throwaway remark, consider the 'security mindset': It's clearly not the case that software built by different teams, in many different languages using many different paradigms, just tends to be secure as a result of some universality phenomena).
The ideal gas laws and the central limit theorem are both examples that come from *mean field theories*. Like universality phenomena, mean field theories are indicative of or rely on a kind of separation of scales that tends to not to be applicable to or true for the kinds of systems we have in mind. But it may be instructive to consider what the failure of this way of thinking means.
When one is looking at a system made up of numerous, interacting elements, it sometimes makes sense to form something called a mean-field approximation, which is a certain type of simpler, non-interacting system (and one that is of course supposed to capture the essential features of the original system). Roughly speaking, in a mean-field approximation, the behaviour of each individual element is modelled via small fluctuations around its average behaviour. And consequently, what often happens when one works out the mathematics of the specific system being studied, is that the way a given element interacts with all of the other elements can be modelled by only keeping track of its 'average interaction' with other elements. To put it another way, we replace the interaction with all the other elements by a single interaction with a '*mean field'*.
In some sense, it is a way of expanding the behaviour of the system about its average behaviour. A typical mean-field approach keeps track of only the first order deviations from the average behaviour and so it can be construed as a kind of linearization of the system around the average behaviour of its elements. So, the failure of approaches like this bear witness to a kind of *essential nonlinearity* in the system in the sense that these approaches can be said to fail when the 'linear part' of the system does not approximate the full system, or when there is no neat 'linear part' of the system in the first place.
One thing that these points suggest to me is that we are unlikely to find and understand useful macroscopic properties by using either linear macroscopic functions, such as averages or other kinds of simple aggregates or integrals, or by using methods that ignore higher order interactions between the elements of the system.
Downward Causation and Explanatory Reduction
--------------------------------------------
In general, when we cannot rely on a 'separation of scales', we need to think carefully about how we are building up our understanding of the system. This is because although larger scale states *are* collections of microscopic states in a literal, formal sense, this does not mean that there are easy answers to questions such as: Is a given (non-microscopic) state *caused by* its more microscopic constituents? And: Can a given (non-microscopic) state be *explained* only in terms of its more microscopic constituents? In fact, we will essentially argue that the answer to both questions is 'no', or rather 'not necessarily'.
It is usually thought of as clear that there are causal effects from smaller scales to larger scales. The first question is about whether or not there also exist non-negligible causal affects in the 'downwards' direction, *i.e.* from larger scales to smaller scales. This is the idea of downward causation. The second question is concerned with the fact that the existence of a constitutive reduction does not imply the existence of a satisfactory explanatory reduction.
A full discussion of these ideas would take us too far afield into philosophy but we will give a few remarks. While causes are indeed usually thought of as different from explanations, like many interpretations of explanatory reduction, ours involves causality in some way (in particular, the two questions in the previous paragraph are not neatly separable). And we think of explanatory reduction as something trying to ''capture the flavor of actual scientific reductions'' as [Sarkar's account](https://link.springer.com/article/10.1007/BF00413566) does. Though since there is often something ''messy'' or ''semi-empirical'' about real scientific explanations, we ought not to expect especially crisp insights to come from discussing these things in the context of thinking practically about complex systems. However, we will briefly discuss some potentially concrete ways that these sorts of ideas are relevant.
The imposition of constraints imposed by the superposition of multiple interacting dynamical systems across scales is enough to illustrate the general idea of downward causation: Let Σ be a collection of macrostates and denote the dynamics of the system at this scale by t↦σ(t)∈Σ. We let S and t↦s(t) denote the microstates and the microscopic dynamics as usual. Suppose we are measuring the property p corresponding to the partition Σ and we see that it change from, say, σ(t)=σ1 to σ(t+1)=σ2≠σ1. We suppose that this change is something understandable at the scale of Σ, *e.g.* perhaps it is clear that the action given by p(σ1) ought to be followed by the action given by p(σ2) in some sequence of actions. Suppose that at the microscopic level, during the same time step, the state changed from s(t)=s1 to s(t+1)=s2. This microscopic change might in some literal sense be *due to* microscopic dynamics, but a more understandable explanation to: ''Why did the system go from state s1 to state s2?'' might have to include the way in which the understandable macroscopic change of state from σ1to σ2 constrained the microscopic behaviour, by insisting that s(t+1)∈σ2.
And if part of the reason for a given state can be due to larger scale properties, then we do not expect it to be possible to give fully satisfactory explanations for that state only in terms of smaller scale properties.
As an example, suppose I have a tiny box that is full of tiny balls. And suppose that the balls have a little bit of room to jiggle around and rearrange. You can imagine that really the box contains atoms of an ideal gas and that the underlying state space of my system is the individual atoms of this world. The box starts at point A. But then I put the box on the back of my truck and drive 10 miles away to point B. Label one of the balls in the box and imagine its full spacetime trajectory. What explains this trajectory and what explains the final position of the ball? A significant *part* of this explanation is surely that I moved the box from A to B? After all, this explains why the ball is anywhere near B at all. If one were trying to explain theposition of the ball *only* in terms of microscopic dynamics, one would be in the absurd position of having to justify and explain *one particular* sequence of molecular states for the world, when all the while it just so happens to be one of an immeasurably large number of sequences of microstates that all map to me grabbing the box and driving it from A to B. So we see in this case, the fact that part of the cause of a state is top-down means that it is more difficult to find a completely satisfactory bottom-up explanation.
One of the broader difficulties that these issues create is with respect to what scales or properties (if any) are ontologically foundational in a complex system. The system is defined in terms of 'individual elements' but we discover that the interactions between scales muddy the notion that the individual elements of the system should be the building blocks of our understanding. On that note, we will end this more philosophical section.
Remarks
-------
I am not an expert on any of the topics that I've touched on in this essay so it is likely that I have made mistakes, or am wrong about something, or have misrepresented something. Please don't hesitate to comment or get in touch if you notice such things. |
974149de-8485-43d9-969c-97e6dec11d85 | StampyAI/alignment-research-dataset/aisafety.info | AI Safety Info | Are Google, OpenAI, etc. aware of the risk?
The major AI companies are thinking about this. OpenAI was founded specifically with the intention to counter risks from superintelligence, many people at Google, [DeepMind](https://medium.com/@deepmindsafetyresearch), and other organizations are convinced by the arguments and few genuinely oppose work in the field (though some claim it’s premature). For example, the paper [Concrete Problems in AI Safety](https://www.youtube.com/watch?v=AjyM-f8rDpg) was a collaboration between researchers at Google Brain, Stanford, Berkeley, and OpenAI.
However, the vast majority of the effort these organizations put forward is towards capabilities research, rather than safety.
|
b50ca385-19f5-4d9b-b930-049ea7a07ea4 | trentmkelly/LessWrong-43k | LessWrong | My confusions with Paul's Agenda
Paul put out a call for probable problems with his agenda, which prompted this post detailing my confusions with it. This isn’t confidently asserting that the plan is doomed, and it’s not obvious to me that any of what follows is a novel objection. But it seems like it’s worth broadcasting the quiet voices of doubt when they appear, and worth trying to be an independent source of judgment. I’ll start by talking about the proposal of Paul’s that I think I understand best, the approval-directed agent, and then talk about IDA, and then finish by pointing at what I suspect is a major underlying crux.
Approval-Directed Agents
While many approaches to alignment view the agent as having some goal (which is only part of what we want it to do) and then restrictions (which prevent it from destroying the other things we want), approval-directed agents try to point directly at the goal of doing the thing that we want it to do as part of a broader ecosystem. If I say “Buy me a plane ticket for my trip to Austin,” the agent decides how much computing resources to allocate the task based on what it thinks I would approve of, and then it spends those resources based on what it thinks I would approve of, and so on, and eventually I trust the conclusion the system comes to and the tradeoffs it made along the way because my judgment has been approximated at every step.
The core insight here seems to be that the question “what would the architect approve of this subsystem doing?” remains a well-formed query for all subsystems, and we could in theory train all of them separately using standard machine learning techniques, or train them jointly with some sort of parameter sharing, or so on. Any unaligned behavior--the agent deciding to angrily tweet at airline companies in order to get a discount, despite me not wanting it to do that--can be traced back to a training failure of some subsystem, and a training failure of the meta-system that decides how to explore. It seems to me like t |
bbaddc19-9964-4279-9a2d-1cb2e8c61428 | trentmkelly/LessWrong-43k | LessWrong | Change My Mind: Thirders in "Sleeping Beauty" are Just Doing Epistemology Wrong
Preamble
> Sleeping Beauty volunteers to undergo the following experiment and is told all of the following details: On Sunday she will be put to sleep. Once or twice, during the experiment, Sleeping Beauty will be awakened, interviewed, and put back to sleep with an amnesia-inducing drug that makes her forget that awakening. A fair coin will be tossed to determine which experimental procedure to undertake:
>
> * If the coin comes up heads, Sleeping Beauty will be awakened and interviewed on Monday only.
> * If the coin comes up tails, she will be awakened and interviewed on Monday and Tuesday.
>
> In either case, she will be awakened on Wednesday without interview and the experiment ends.
>
> Any time Sleeping Beauty is awakened and interviewed she will not be able to tell which day it is or whether she has been awakened before. During the interview Sleeping Beauty is asked: "What is your credence now for the proposition that the coin landed heads?"
----------------------------------------
Motivation
I was recently introduced to the canonical Sleeping Beauty problem and initially was a halfer but confused. Or more like I thought the halfer position was correct, but smart people seemed to be thirders and I was worried I was misunderstanding something about the problem, or confused myself or similar.
I debated the problem extensively on the LW Discord server and with some LLMs and strongly updated towards "thirders are just engaging in gross epistemic malpractice".
A message I sent in the LW server:
> Alternatively I started out confused.
>
> Debating this problem here and with LLMs convinced me that I'm not confused and the thirders are actually just doing epistemological nonsense.
>
> It feels arrogant, but it's not a poor reflection of my epistemic state?
I still have some meta level uncertainty re: the nonsense allegations.
I want to be convinced that the thirder position is not nonsense and there is a legitimate disagreement/debate to be had here.
|
50d77e32-aaca-4afe-9595-3cbbf8d797db | trentmkelly/LessWrong-43k | LessWrong | Fixing The Good Regulator Theorem
Conant & Ashby’s “Every Good Regulator Of A System Must Be A Model Of That System” opens with:
> The design of a complex regulator often includes the making of a model of the system to be regulated. The making of such a model has hitherto been regarded as optional, as merely one of many possible ways.
>
> In this paper a theorem is presented which shows, under very broad conditions, that any regulator that is maximally both successful and simple must be isomorphic with the system being regulated. (The exact assumptions are given.) Making a model is thus necessary.
This may be the most misleading title and summary I have ever seen on a math paper. If by “making a model” one means the sort of thing people usually do when model-making - i.e. reconstruct a system’s variables/parameters/structure from some information about them - then Conant & Ashby’s claim is simply false.
What they actually prove is that every regulator which is optimal and contains no unnecessary noise is equivalent to a regulator which first reconstructs the variable-values of the system it’s controlling, then chooses its output as a function of those values (ignoring the original inputs). This does not mean that every such regulator actually reconstructs the variable-values internally. And Ashby & Conant’s proof has several shortcomings even for this more modest claim.
This post presents a modification of the Good Regulator Theorem, and provides a reasonably-general condition under which any optimal minimal regulator must actually construct a model of the controlled system internally. The key idea is conceptually similar to some of the pieces from Risks From Learned Optimization. Basically: an information bottleneck can force the use of a model, in much the same way that an information bottleneck can force the use of a mesa-optimizer. Along the way, we’ll also review the original Good Regulator Theorem and a few minor variants which fix some other problems with the original theorem.
The Orig |
0597b072-2654-4214-8c0c-6fc82062a289 | trentmkelly/LessWrong-43k | LessWrong | DC Meetup: Last Discussion Thread
Followup to: this and this
I'm going to be in Cambridge for the next week or so, so I'm posting things now to make sure that we can get the ball rolling.
There's going to be DC meetups on May 1st and May 15th. 2 PM - 8 PM, unless anyone objects.
There was discussion about whether to meet in Bethesda or NW DC. I propose that we just try a meetup in both, May 1st in Bethesda and May 15th in NW DC.
Some questions I have:
* Any suggestions on what I should look for in a venue?
* Anything you want me to ask the Cambridge meetup group?
I have some preliminary things to talk about in mind, but if you guys could think about a few topics too, it would be helpful. |
e9efc283-f5aa-4033-bf10-343ac5922f35 | trentmkelly/LessWrong-43k | LessWrong | Meetup : London Rationalish meetup, 7/2/16
Discussion article for the meetup : London Rationalish meetup, 7/2/16
WHEN: 07 February 2016 02:00:00PM (+0000)
WHERE: Shakespeare's Head, 64-68 Kingsway, London WC2B 6AH
(The new experimental name for the diaspora meetup group.)
We had around 30 people last time! (Plus at least four different meal shakes.) Go, us!
This meetup will be social discussion in a pub, with no set topic. If there's a topic you want to talk about, feel free to bring it.
The pub is the Shakespeare's Head in Holborn. There will be some way to identify us.
The event on facebook is visible even if you don't have a facebook account. Any last-minute updates will go there.
----------------------------------------
We're a fortnightly London-based meetup for members of the rationalist diaspora. The diaspora includes, but is not limited to, LessWrong, Slate Star Codex, rationalist tumblrsphere, and parts of the Effective Altruism movement.
You don't have to identify as a rationalist to attend: basically, if you think we seem like interesting people you'd like to hang out with, welcome! You are invited. You do not need to think you are clever enough, or interesting enough, or similar enough to the rest of us, to attend. You are invited.
People start showing up around two, and there are almost always people around until after six, but feel free to come and go at whatever time.
Discussion article for the meetup : London Rationalish meetup, 7/2/16 |
b87f44c4-f47c-4912-94be-580455d896b9 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Paradigms and Theory Choice in AI: Adaptivity, Economy and Control
*Epistemic status: Some general thoughts on philosophy of design, that have turned out to be useful for reasoning about AI paradigms for a few years for me.*
Summary:
1. I briefly discuss the role of paradigms in Kuhnian philosophy of science, and the problem of theory choice.
2. I introduce the notion of design paradigms, when discussing design and engineering. For machine learning, this concept is similar to the concept of [*training stories*](https://www.alignmentforum.org/posts/FDJnZt8Ks2djouQTZ/how-do-we-become-confident-in-the-safety-of-a-machine).
3. I then introduce three criteria (or theoretical virtues) that design paradigms can be evaluated on, that play a role in shaping theory choice in design and engineering generally, and AI specifically.
4. I then apply these three criteria to some criticisms of the CAIS paradigm (as compared to AI agents), as a case study.
### 1. Paradigms in Science
In the philosophy of science, Thomas Kuhn introduced the notion of a *‘paradigm’*, roughly referring to a class of key theories, tools and approaches, and methodological and theoretical assumptions that govern progress within some discipline of science. The function of a paradigm, as per [SEP](https://plato.stanford.edu/entries/thomas-kuhn/), is to *“to supply puzzles for scientists to solve and to provide the tools for their solution”*. Kuhn’s interest in the concept was to provide a commentary on revolutions and incommensurability.
According to Kuhn, scientific revolutions are characterized by disruptions in periods of ‘normal science’ (or within-paradigm progress), wherein new paradigms are introduced leading to a shift in the assumptions, methods and evaluative criteria that govern the discipline. Some examples include the shift from classical mechanics to relativistic mechanics, and the shift from miasma theory to germ theory of disease.
Sometimes paradigms can refer to an overlapping object of inquiry and continue to remain productive, and yet be incommensurable (i.e. lacking common measure) with respect to each other. When the theories provide no basis for evaluation on each other’s standards, and science lacks a common standard to compare them (for example because they have incompatible methodological assumptions, or operate in incompatible ontologies), it can make it hard to provide *‘objective’* justifications for choosing between paradigms. Kuhn called this the problem of Theory Choice, and raised the question of which criteria we do/should use in making subjective choices here.[[1]](#fn7ao65k3360o)
### 2. Paradigms in Design and Engineering
Artificial Intelligence is partly about understanding the nature of intelligence or related properties of cognitive systems, but it is also about the artificial -- designing intelligent artifacts. And therefore in order to understand what paradigms govern progress in AI, and how they do so, we need to look at paradigms not just in science but in design and engineering.
Paradigms and theories operate differently in science compared with design and engineering, relating to the distinct epistemological relationship between the use of theories and the object of those theories. In science, the territory is typically causally prior to our maps, and we are trying to refine our maps to get the best understanding of the key phenomena in the territory. Design and engineering on the other hand, are also about artificiality, and theories often provide us with blueprints for creation of artifacts.
The laws of the universe are a given, but artificial objects are not. Different scientific paradigms will not cause the laws of the universe to alter, but different design paradigms will cause different artificial objects to be created. This makes theory choice for design paradigms particularly important: design paradigms don’t just describe the world more or less accurately, they also directly shape what happens next.[[2]](#fnbk40r7t9fx)
Furthermore, scientific paradigms tend to commit you to a whole worldview. Design paradigms commit you to aspects of a worldview, but not necessarily the whole thing. This makes design paradigms less incommensurable than scientific paradigms, and makes it easier for multiple design paradigms to exist side by side. Automobile engineering and locomotive engineering both make use of the similar scientific principles, but operate within the context of distinct artifact classes.
Although philosophy of science has not focused on the nature of paradigms in design, there are some things we can say about them from common intuitions. A design paradigm is a space of blueprints, denoting a space of possible artifacts, and a way of reasoning about those artifacts. It must help us build as well as understand the built systems.
For example, automobile engineering as a paradigm provides us with ways of deliberating about cars as artifacts, helping us with building cars and reasoning about their behavior, as well as providing forms of knowledge relevant to repairing them if and when they break down. Different paradigms can deal with problems that may share some abstract equivalence. For example, automobile engineering and aerospace engineering both are paradigms that deal with artifacts useful for transporting from point A to point B, though their applicability differs and the forms of knowledge involved are fairly different.
One way of thinking about design paradigms in AI is to consider different approaches to AI as different design paradigms. For example, on this view the move from simulation-based self-play RL to transformer-based language models is a form of paradigm shift.
Some commentators have already discussed design paradigms in AI under different names. A similar idea has been discussed with the name of *model-types* in ML in [this paper](https://arxiv.org/abs/1912.01172) looking at the values that govern disciplinary shifts. In my understanding, a good overview of what a paradigm in ML looks like is presented in the concept of [*training stories*](https://www.alignmentforum.org/posts/FDJnZt8Ks2djouQTZ/how-do-we-become-confident-in-the-safety-of-a-machine), capturing the notion of paradigms that govern how we train models and reason about their properties. For most of the discussion in this write-up, a design paradigm can be substituted with training story (when discussing ML-based systems).
In the section below, I discuss theory choice as a force governing paradigm shifts. However, it is important to note that inter-paradigm choice is not the only way in which paradigm shifts occur. For example, new design paradigms can also emerge through synthesis building on top of each other (for example, early hovercrafts were produced by synthesizing aerodynamics with torpedo boat designs).
### 3. Three Criteria for Paradigm Choice: Adaptivity, Economy, and Control
Design paradigms can be used to understand and model how progress in AI will develop. We can then examine the values that drive theory choice. Understanding what drives paradigm shifts can also help us understand general features of future paradigms. Furthermore, as I’ll show in the next section, these dimensions can also help us interpret criticisms of alternative paradigms and make relevant distinctions in the nature of the criticisms.
Here the criteria are described in terms of an informal notion of a design paradigm denoted by K.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')}
, capturing the relevant knowledge in the paradigm for building and reasoning about artifacts, and a class of functionings F that denotes some large space of possible claims about how an artifact might work or operate or fulfill some specified purpose.
**Adaptivity:** For some target class of functionings F(⊂F), such that K allows to construct an artifact A to achieve F, K provides us means to justifiable belief that A will achieve functionings F over a class of operating conditions E.
For example, when engineering a car, you want to ensure that the car will still work under a wide range of weather conditions, road and traffic conditions, etc. even though one may not be able to ensure that it will work when put in water. Design problems typically want assurances over some set of operating conditions (over which we wish to know that the artifact will work as expected) based on the scoping and assumption that goes into them.
Operating conditions can also include history-dependencies, for example claims like "the car should remain working even after having traveled for 10,000 km over proper roads".
For extremely powerful AI systems, however, where the future conditions are further shaped by the optimizing behavior of AI systems themselves, the *principle of adaptivity* also entails the *principle of adaptive consistency* -- that is the system should not self-modify into another artifact which cannot be reasoned about via the given design paradigm.
**Economy:** For some target class of functionings F(⊂F), such that K allows to construct an artifact A to achieve F, K provides us means to justifiable belief that A will achieve functionings F given some resource budget R, such that the economic uses of F are proportionate to R.
Note that while the economy of design paradigms does not completely preclude their exploration at moderate scales, economic viability does shape widespread use of engineering artifacts and the scientific attention invested into developing them further.
More importantly, the design paradigms can also inform speculations about future economic viability allowing paradigms to mature even if they are not yet economically viable. This often happens through economic institutions that allow making bets on the maturation of various competing paradigms.
In machine learning, paradigms can often claim economic viability by either claiming more general F -- that is, ability to exhibit more general intelligent behavior, or competence on a broader variety of tasks -- or, via claiming efficiency in R, the amount of data, compute and other resources involved for achieving the same levels of competence.
**Controllability:** For some target control behavior f:Θ→2F (functioning classes parametrized by Θ), such that K allows to construct an artifact A with control variables corresponding to Θ, K provides us means to justifiable belief that A will achieve functionings f(θ) whenever the control variables are θ.
Different design and engineering problems require different forms of controllability. Often, controllability is part of the economic viability of an artifact, though it is not always the case that more controllable artifacts are necessarily more economically competitive.
While in an ideal world we would want both [adaptivity and controllability to be reflected in the economic valuation of artifacts](https://www.lesswrong.com/posts/S95qCHBXtASmYyGSs/stuart-russell-ai-value-alignment-problem-must-be-an) and possible artifact classes that a paradigm enables, insofar as economic viability is measured in narrow performance, it can pose trade-offs with the other criteria.
In some sense, alignment tax can be treated as a measurement of the controllability-economy trade-off for AI paradigms, while the adaptivity-economy trade-off rhymes a bit with the robustness-performance or robustness-accuracy trade-offs in machine learning.
### 4. Case Study: CAIS vs Agents
When discussing Eric Drexler's [Comprehensive AI Services Model](https://www.fhi.ox.ac.uk/wp-content/uploads/Reframing_Superintelligence_FHI-TR-2019-1.1-1.pdf) in 2019-20 with several alignment researchers, specifically how it offered an alternative way to reason about the trajectory of AI progress that did not involve advanced agency, the different criticisms[[3]](#fndi12xx9thxq) I heard could be classified into the following general directions:
1. A well-integrated ecology of AI services will eventually lead to emergent agents, either due to collusion, or due to emergence of a dominant power-seeking service, or [other emergent structures](https://www.alignmentforum.org/posts/LpM3EAakwYdS6aRKf/what-multipolar-failure-looks-like-and-robust-agent-agnostic).
2. Solving alignment and governance for CAIS will inevitably require solving alignment of agents, as CAIS merely pushes the hard parts of alignment problem into the alignment of security services or R&D services.
3. AI Agents can solve useful problems for humans that no reachable collection of AI services can, and therefore there will always be economic incentive for pursuing agentic systems.
4. [Given a large enough amount of data, compute and model size, integrated agents outperform collection of AI services.](https://www.alignmentforum.org/posts/x3fNwSe5aWZb5yXEG/reframing-superintelligence-comprehensive-ai-services-as?commentId=sXHXAfSKWPyEMhtcu)
As we can observe, the first dimension of criticism is equivalent to a claim that we should focus on paradigms that help us reason about agents because the model proposed by AI services is not adaptively consistent and will evolve into an agent-like system anyways. The second dimension of criticism challenges the controllability claims of CAIS, challenging whether the control problem of the paradigm can be solved within that paradigm itself or inevitably requires invoking another paradigm.
Note that these are roughly orthogonal dimensions of criticism, at least in terms of logical independence. It could be possible that CAIS does evolve into agentic systems, but that solving the control problem for CAIS was easier than aligning agents, implying that the agent that eventually emerges is an aligned agent. Such a position might imply that aligning agents via building and reasoning about bounded-scope services is a more tractable strategy than trying to solve alignment for already natively-agentic systems (i.e. systems that are agentic at the time of conception).
Conversely, it could be possible that solving the control problem for CAIS does involve solving alignment for agent systems which must take the role of security services (or other core governance roles), but the aggregate behavior of AI services never evolves into an integrated utility-maximizing agent[[4]](#fn3c64dqn7idu). Similarly, we can argue the orthogonality of 1 and 2 both with 3&4.
The third and fourth criticisms correspond to the economy criterion, claiming that the CAIS paradigm will never be economically competitive, either because it is constrained from achieving some set of economically valuable functionings that agents can, or because it is not as resource-efficient as agents in achieving comparable functionings.
When posed with a novel paradigm, we can use these criteria to assess in what ways will they dominate existing paradigms if successful, and organize criticisms over their claims.
*(Parts of these ideas were produced during my work on CAIS in 2020-21 and presented at FHI. Some of those ideas were later refined in discussions with Rose Hadshar in late 2022, and the clarification of those ideas as they pertain to general theory of design happened during AI Futures Residency at Wytham in December 2022. I thank several people who have been involved in discussions and sanity-testing of the criteria discussed.)*
1. **[^](#fnref7ao65k3360o)**See [here](https://www.andrew.cmu.edu/user/kk3n/philsciclass/kuhn.pdf) for Kuhn’s original essay, and the introduction of [this](https://www.sciencedirect.com/science/article/abs/pii/S0039368113000332) paper for a succinct introduction to theory choice in general.
2. **[^](#fnrefbk40r7t9fx)**As an aside, this feature of design paradigms is why Drexler says CAIS is both descriptive and prescriptive, and why claims about CAIS are sometimes hard to reason about in a purely descriptive way. See Drexler, [*Reframing Superintelligence*](https://www.fhi.ox.ac.uk/wp-content/uploads/Reframing_Superintelligence_FHI-TR-2019-1.1-1.pdf), 2019, pp. 44-46.
3. **[^](#fnrefdi12xx9thxq)**Some of these criticisms can also be found in Richard Ngo's [*Comments on CAIS*](https://www.alignmentforum.org/posts/HvNAmkXPTSoA4dvzv/comments-on-cais).
4. **[^](#fnref3c64dqn7idu)**See Chapter 6 of Drexler's [*Reframing Superintelligence*](https://www.fhi.ox.ac.uk/wp-content/uploads/Reframing_Superintelligence_FHI-TR-2019-1.1-1.pdf)*.* |
a057a206-c062-419c-a686-9bfe2d2a46bd | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Questions about Value Lock-in, Paternalism, and Empowerment
Pulling a child out of the path of a fast car is the right thing to do, whether or not the child agrees, understands, is grateful, or even is hurt during the rescue. [Paternalistic](https://plato.stanford.edu/archives/sum2014/entries/paternalism/) acts like this, when we might argue that a person's straightforward consent ought to be overridden, add uncomfortable complications to liberal ethical frameworks.
As artificial intelligence continues to improve, there are an increasing number of domains in which its judgement is superior to that of humans. At what point is the relationship between AI and humans similar to that between an adult and a child? And to what extent should we demand to understand, agree or consent to decisions made by AI?
I see Paternalism and intentional Value Lock-in as intimately related. Considering that we may, by influencing the creation of AI, have a significant impact on the freedoms of future people: how should we responsibly discharge this power?
In writing this essay I’m hoping to reach people who’d like to discuss these things in more depth, and people who might want to work together.
I'd be very grateful for any criticism, suggestions, pointers, or discussion :)
Structure of this writing
=========================
I start with brief overviews of what I mean by ‘Value Lock-in’ and ‘Paternalism’, and the connection I see between them.
Then I discuss a core intuition behind my recent research into a plausible minimally-paternalistic approach to alignment: the idea of ‘un-asking’ human values. I include a brief discussion of technical approaches, which I plan to expand on more fully in later writing.
The core of this essay is a list of questions for which I currently have no answer:
* Metaethics: Is all moral change contingent?
* Paternalism: How much epistemic humility is appropriate or desired? What is the minimal amount required as a condition of logical consistency?
* Empowerment: How does choice maximisation behave in the context of Social Choice Theory?
* Practicality/deployment: Is the alignment tax of something like “maximise Empowerment-of-Others” too high?
* Empowerment: Is there a way to generalise choice-maximisation approaches beyond toy models?
* Hybridisation: Is there a way to balance goal-inference and choice-maximisation in a way that is robust to both reward-hacking and value lock-in?
* Paternalism: Is there a way to generalise debate between LLM agents to comprehensively and meaningfully explore non-binary/open questions?
Finally, I wrap up with some discussion about the purpose of this writing, and suggest possible future steps.
Value Lock-in
=============
When creating a new technology, such as the for-profit corporation or cryptocurrency, certain values may end up “locked in” and difficult to change once the technology is established. [Values locked-in to powerful AI systems](https://forum.effectivealtruism.org/posts/KqCybin8rtfP3qztq/agi-and-lock-in) could mean that a single ideology could rule for a very long time, with no competing value systems.[This ideology may reflect the values of a particular individual or group, or be accidentally misaligned](https://www.lesswrong.com/posts/DJRe5obJd7kqCkvRr/don-t-leave-your-fingerprints-on-the-future). It may allow no meaningful dissent or debate, even in the case of evolving knowledge, understanding, and moral tastes[[1]](#fn55tdxzgznkb).
Chapter 4 of Will MacAskill's *What We Owe The Future* is a very good introduction to value lock-in, describing it as arising from a combination of other convergent instrumental goals such as self-preservation and goal-content integrity (see e.g. Bostrom's 2012[*The Superintelligent Will*](https://nickbostrom.com/superintelligentwill.pdf)), and discussing the risks of very-long-term AI persistence (comparing emerging AI to the advent of writing).
Is there a way of minimising or mitigating the effect that value lock-in will have? Are there meta-values which, either for coherency or to protect things we unanimously value, we must unavoidably lock in?
Paternalism
===========
Interfering with a person against their will, defended or motivated by claims about their well-being, has been well-studied as [Paternalism](https://plato.stanford.edu/archives/sum2014/entries/paternalism/). We accept that, at least in some cases, the frameworks of consent and liberalism are strained[[2]](#fn5uv196m6zpr). Yet meddling interference is often unwelcome, insulting and constrictive. In the worst cases, it is simply a fig leaf for selfish coercive control.
Iason Gabriel's[*Artificial Intelligence, Values, and Alignment*](https://philpapers.org/go.pl?id=GABAIV&proxyId=&u=https%3A%2F%2Fdx.doi.org%2F10.1007%2Fs11023-020-09539-2)describes multiple potential 'targets' of alignment:
* literal instructions, as in the case of King Midas
* expressed intentions, correctly interpreting the intent behind an instruction
* revealed preferences, doing what one’s behaviour indicates they prefer
* informed preferences, doing “what I would want it to do if I were rational and informed”
* well-being, doing “what is best” for the person, objectively speaking
* moral values, doing “what it morally ought to do, as defined by the individual or society”.
For many of these potential targets, there is some level of discretionary judgement-call which serves as an interpretation or 'correction' to the request. When should one make such corrections on another's behalf, thinking that one knows better?
One might argue that this is a less pressing concern than other Alignment areas, since Paternalism assumes that at least *some*human’s preferences are being represented. An extreme dictatorship might seem preferable to human extinction, since we might expect a human dictator’s values to be at least somewhat similar to the average human’s. By contrast, I’ve chosen to research Paternalism in the context of superintelligent AGI because it seems relatively neglected (compared to e.g. mechanistic interpretability), and because of being personally motivated by questions around avoiding oppression and encouraging individual exploration. While the opportunity cost of non-existence is significant, I am more moved by the prospect of people trapped in a value system not of their own choosing.
"Un-asking" human values
========================
In this section, I describe one underlying theme I have been exploring, searching for a general approach to minimally-paternalistic actions.
Since interpretation of a person’s requests or behaviour creates the potential for paternalistic value judgements to be locked in to the system, I'm curious about approaches which try to "un-ask" the question of human value, instead trying to be helpful without taking a strong potentially-paternalistic stance on "what's good". Here, not only am I trying to avoid relying on intent alignment, but also - as far as possible - I aim to minimise commitment to any particular kind of value alignment.
Trying to be value-agnostic is kind of paradoxical, in a way that I'm not sure how to address:
- Any stance which aims to preserve people’s ability to make their own choices *must* be opinionated in some meaningful way. How can "meaningful free choice" be preserved without making a value statement about which things are meaningful?[[3]](#fnl8n7c9ku3z)
- The[Paradox of Tolerance](https://en.wikipedia.org/wiki/Paradox_of_tolerance) (a society which seeks to remain tolerant must be intolerant of intolerance) seems core to the question of non-paternalism. To what extent do we allow one person to coerce or limit another? (How does this relate to an individual's effect on their future selves: should one be able to take on crippling debt?)
Technical approaches
--------------------
A large part of the[Outer Alignment Problem](https://www.lesswrong.com/tag/outer-alignment) is that people don't know how to communicate what we want. (One might go further, and say that people don't even *know* what we want.) This makes it hard to help us.
Instead of *asking* the agent (and having to use judgement to interpret the answer), approaches such as[Inverse Reinforcement Learning](https://dl.acm.org/doi/10.1145/279943.279964) (IRL) seek to learn the values of an agent (e.g. human values) through observation of its behaviour[[4]](#fnz8w1dash0vp).
One main shortcoming of IRL is that agents sometimes behave in 'irrational' ways that don't reliably or directly bring about their goals. This may be because of physical or mental limitations, or because the true goals of the demonstrator lie outside the hypothesis set of the observer. One example of this latter point is that[“Demonstration by Expert” can be a suboptimal way to teach a helpful robot about a reward function](https://arxiv.org/abs/1606.03137) (consider showing a new worker where the spare coffee lives, before it’s needed). More generally,[without assumptions about an agent's rationality, its value function cannot be derived from its observed behaviour.](https://arxiv.org/abs/1712.05812)
### Choice Maximisation
Rather than inferring an agent's goals, to be able to help the agent acheive them, we might instead try to increase the agent's ability to reach a variety of outcomes, without needing to take an explicit stance on which outcomes the agent might desire.
One potential advantage of this kind of approach is that it could sidestep the problem of instrumental power-seeking (see[Bostrom](https://nickbostrom.com/superintelligentwill.pdf),[Turner](https://www.lesswrong.com/posts/6DuJxY8X45Sco4bS2/seeking-power-is-provably-instrumentally-convergent-in-mdps)). The reasons to be afraid of a power-seeking agent comprise unintended collateral damage, and [incorrigibility](https://arbital.com/p/corrigibility/)[[5]](#fnn6yhqq3u43j). If, instead, a power-seeking agent is trying to maximise humans' own control over their future, and so turns over its newly-gained power to us, we could perhaps remain able to choose from a variety of possible futures.
In an attempt to helpfully increase the variety of outcomes available to an agent, we might try to increase the agent's[Empowerment](https://en.wikipedia.org/wiki/Empowerment_(artificial_intelligence)), which is “[the maximal potential causal flow from an agent’s actuators to an agent’s sensors at a later point in time](https://www.frontiersin.org/articles/10.3389/frobt.2017.00025/full#B8)”. This could, for example, be measured as the[Mutual Information](https://en.wikipedia.org/wiki/Mutual_information) between mathematical sets of Actions available to an agent, and sets of States which might arise in the future: how much the choice of Action tells us about which future State arises.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')}
[[6]](#fn9fd36oj0xoo). This can happen as a [perception-action loop](https://www.frontiersin.org/articles/10.3389/frobt.2017.00025/full), and agents can [cooperatively work to empower each other](https://research.gold.ac.uk/id/eprint/21745/1/guckelsberger_cig16.pdf), or [competitively act to disempower each other](https://arxiv.org/abs/1806.01387).
Other approaches to this kind of goal-agnostic choice-maximisation include[Turner et al's "POWER"](https://arxiv.org/abs/1912.01683) and[Franzmeyer et al's "CHOICE"](https://arxiv.org/abs/2107.09598). My reasons to focus on Empowerment are:
1. The "POWER" approach is based on Markov Decision Processes, which I think scale poorly and are difficult to generalise to continuous (rather than discrete) states. (I'm uncertain about this, though, and "read up on MDPs" is on my to-do list.)
2. The "CHOICE" approach does away with the "Theory of Mind" aspect of Empowerment. While this has the advantage of no longer requiring an environment simulator, which improves the speed and tractability of their chosen problems[[7]](#fn42jfuasj2pt), I think that considering an agent's own assessment of their Empowerment is a significant advantage: an agent is only Empowered if it believes itself to be so.
### Shortcoming of Choice Maximisation: Trading Optionality for Reward
An assistant that sought to maximise optionality while staying completely agnostic to reward would work against a human’s attempt to commit to a particular desired end-state. If choice-maximisation is the only thing valued by an assisting AI, then we would naively be forced into whichever world most maintains optionality, without ever being able to ‘cash in’. We see this play out in [*Shared Autonomy via Deep Reinforcement Learning*](https://arxiv.org/abs/1802.01744), where auto-pilots attempt to increase an human pilot's Empowerment, but when too opinionated insist on hovering, working against the human pilot’s efforts to land. It seems likely that some kind of hybrid approach would be necessary.
### Shortcoming of Empowerment: Delusional Beliefs
An agent’s Empowerment is subjective: it is the Mutual Information between *its own conception* of the actions available to it, and the possible futures *which it can imagine*. This is an important point, for a number of reasons:
1. Theory of Mind: For a Helper agent to increase the Empowerment of an Assisted agent, the Helper agent must have some model of the understanding and world-view of the Assisted agent.
2. Education: It is possible to Empower an agent simply by educating it about the abilities available to it, or about the nature of the world. In a room with a lever and a closed door, it is Empowering to inform someone that the lever opens the door.
However, the subjectivity of Empowerment introduces an important consideration: delusional beliefs. In the previous example of the lever and door, it increases the Empowerment of the person in the room for them to believe that the lever opens the door, whether or not this is the case in reality. There is no inherent reason for a Helper agent trying to maximise the empowerment of an Assisted agent to provide truthful information, and indeed it is more likely than not for Empowerment-maximising statements to be untrue.
One mitigation here is the concept of trust: over a long enough time period of repeated interactions, a Helper which loses the trust placed in it also loses the ability to increase the Empowerment of the Assisted agent. This may result in verifiable statements being a necessary component of the long-term Empowerment-maximising strategy.
Approaches I have avoided
-------------------------
### Static parameter-based regularisation of side-constraints
One common approach to attempting to balance doing-good with not-doing-bad is to have side-constraints. In[Conservative Agency](https://arxiv.org/abs/1902.09725), a regularisation parameter λ "can be interpreted as expressing the designer’s beliefs about the extent to which [the reward function] R might be misspecified". This successfully induces conservative behaviour in the paper’s simulated environments. However, the approach hard-codes an assumption of the level of correctness of the reward function’s specification, which seems unsafe to me.
Elsewhere, it is suggested that an agent with access to multiple ways to meet its goal would take the approach that minimises side-impact. This reminds me of food made with ingredients "sourced as locally as possible": this kind of least-privilege tie-breaker feels ripe for regulatory capture by the agent itself. I haven't thought this through as closely as I want to, though, and this intuition would benefit from being stated more formally.
My original research plan (coming from[earlier exploratory writing of mine](https://sambrown.eu/writing/The%20Empowerment%20of%20Others.pdf)) was to investigate methods of hybridising Choice Maximisation (e.g. Empowerment) with Goal Inference (e.g. Inverse Reinforcement Learning), potentially also including active-learning based around curiosity or uncertainty. But it's unclear to me how to do this without a naïve tuning parameter which is either hackable or arbitrarily locked-in.
### Censorship-based approaches
One approach to AI Safety is to attempt to guide a given model away from particular behaviour which the creators consider undesirable. For example,[Anthropic's "Helpful and Harmless Assistant"](https://arxiv.org/abs/2204.05862) disfavours violent responses, and Stable Diffusion's NSFW Safety Module ([description](https://huggingface.co/CompVis/stable-diffusion-v1-4#safety-module),[code](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/safety_checker.py)) aims to avoid generation of sexual content. While certain traits, such as avoiding intentional deception, may end up being necessary to any system that avoids undesirable lock-in, the approaches mentioned above seem like opinionated stances on a more object-level ethical level than I think is helpful for considering questions of value lock-in and paternalism.
Concrete questions and problems
===============================
In this section, I try to pinpoint concrete questions and problems around Paternalism, Value Lock-in, and Empowerment. These are *my* questions rather than those of any broader research community. I begin with more abstract, philosophical questions, and become increasingly technical.
### Metaethics: Is all moral change contingent?
Many values which are held today are so held because those values are culturally adaptive - people are a product of their society, and stable forms persist. That is to say: certain norms, such as valourising conquest or favouring caring for children, seem intuitively to increase the competitive fitness of a culture. But we should not *ex ante* expect that the evolutionary benefit of a particular trait speaks to its moral desirability. This is similar to the discussion in [Allan Dafoe’s *Value Erosion*](https://docs.google.com/document/d/1B77VWaXG-u34nSRFKV14pJNHJHHb6sa5zJ08J70CVVA/edit#): “absent a strongly coordinated world in which military and economic competition is controlled, the future will be shaped by evolutionary forces and not human values”. What reason do we have to believe that any 'moral' trait which we see agreed upon within or across cultures is desirable? Kind of a big question.
Right now I’m reading the chapter *Why Act Morally?* in Peter Singer’s *Practical Ethics*, which has the line “Whether to act according to considerations of ethics, self-interest, etiquette, or aesthetics would be a choice ‘beyond reason’ - in a sense, an arbitrary choice. Before we resign ourselves to this conclusion we should at least attempt to interpret the question so that the mere asking of it does not commit us to any particular point of view.”
If all moral change is contingent, then I do not see that we have any special reason to avoid imposing our own preferences on future people, whose moral views will be just as arbitrary as our own.
### Paternalism: How much epistemic humility is appropriate or desired? What is the minimal amount required as a condition of logical consistency?
Encoding *complete* epistemic uncertainty would prevent the taking of *any* action[[8]](#fn5zd4mnvmpfm), and so would be equivalent to declining to deploy AGI: as such it [seems an impractical solution to the alignment problem](https://www.alignmentforum.org/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a-list-of-lethalities). *Some* epistemic statement *must* therefore be made, which will unavoidably lock in some path-dependence. What is the minimal possible coherent value statement? What are its implications, and possible shortcomings? What additional statements might various groups desire?
### Empowerment: How does choice maximisation behave in the context of Social Choice Theory?
Literature on choice maximisation currently covers [single Helper/Assisted-agent pairings](https://proceedings.neurips.cc/paper/2020/file/30de9ece7cf3790c8c39ccff1a044209-Paper.pdf), [agents which share a reward function](https://arxiv.org/abs/2107.09598), and [agents which play zero-sum games against each other](https://arxiv.org/abs/1806.01387). But how should an altruistic agent behave when trying to help a collection of agents who are engaged in a mixed- or zero-sum game with each other? Whose Empowerment should be maximised? How do we weight the Empowerment of non-human animals, or digital minds? How should an agent which is trying to maximise the Empowerment-of-Others behave when one of these others wants to disempower another? This is a question of welfare functions, familiar to political philosophy, which I have so far neglected.
### Practicality/deployment: Is the alignment tax of something like “maximise Empowerment-of-Others” too high?
Assuming that something like Empowerment-of-Others is a defensible, ethical thing to maximise: is there any reason to consider that a first-actor would be incentivised to choose this rather than their own personal goal?
### Empowerment: Is there a way to generalise choice-maximisation approaches beyond toy models?
Existing successes of Empowerment are (to my knowledge) limited to arenas such as [grid-worlds](https://research.gold.ac.uk/id/eprint/21745/1/guckelsberger_cig16.pdf) and [real-world drones flying in 3D](https://arxiv.org/abs/1802.01744). These successes are sufficient to prove the concept, but do not easily extend to encompass all the things people might care about. People do not only care about position in 3D space: a conversation between two people may go well, or poorly, without either participant leaving their seat. It is unclear to me how nuanced outcomes, like the quality of a discussion, could be well-represented by the finite states of an MDP.
What about the general intuitive concept of optionality which Paul Graham calls "[staying upwind](http://www.paulgraham.com/hs.html)"? These presumably exist as concepts in AlphaZero (which is capable of positional play), but again are restricted to particular arenas.
More specifically, it seems that the choice of metric over the spaces of Actions and future States completely predetermines Empowerment. In grid worlds, identifying States with position is reasonable. But instead if we have a person in front of a single apple on a table, does cutting it into eighths increase the person's Empowerment (because there are more things to pick up) or reduce it (because they can no longer choose between a whole apple, and an apple in pieces)? This choice determines which action will be preferred: to cut or not to cut. These choices can be made appropriately for particular specific purposes, but it is not clear to me how choices about which of multiple options is more Empowering can be generalised in a tractable and value-agnostic way.
### Hybridisation: Is there a way to balance goal-inference and choice-maximisation in a way that is robust to both reward-hacking and value lock-in?
There is existing work on hybridising goal-inference and choice-maximisation, for example [creating an auto-pilot that seeks to stabilise a human-piloted vehicle](https://arxiv.org/abs/1802.01744). In these situations, excessive choice-maximisation will subvert a human’s attempt to trade optionality for reward. In the literature I’ve found, this is resolved either by a simple Lagrangian (which I discuss in[*Parameter-based regularisation of side-constraints*](https://www.lesswrong.com/posts/nLjtqdhRaKcEGb4NA/value-lock-in-paternalism-and-empowerment#Parameter_based_regularisation_of_side_constraints)), or by [selecting the most Empowering policy from a minimally-valuable set](https://uhra.herts.ac.uk/bitstream/handle/2299/23593/GDE_final.pdf?sequence=1). Both of these seem vulnerable to misspecification of the reward function, and (more importantly) to a hard-coded misspecification of the uncertainty around the accuracy of the reward function. Perhaps weighting by a dynamically-updating Bayesian confidence of the inferred goal would be an improvement?
### Paternalism: Is there a way to generalise debate between LLM agents to comprehensively and meaningfully explore non-binary/open questions?
OpenAI describe[AI Safety via Debate](https://openai.com/blog/debate/), where two agents debate whether a given photo is of a cat, or of a dog. The hope is that for difficult problems, computers could argue for opposite points of view, breaking down the problem step-by-step into possible decision points, where the reasoning could be easier for humans to audit, verify and referee.
I'm not aware of any generalisation of Safety via Debate from binary yes/no questions to broader discussion. For example, rather than Anthropic's approach of downweighting violent suggestions for how to obtain drugs, where there is a single Helpfulness/Harmlessness axis, could there instead be some maximally-diverse representation of viewpoints (see for example [the GPT-3 simulations of human subgroups of Argyle et al](https://arxiv.org/abs/2209.06899)), which could provide good coverage of diverse values and opinions to engage in a forum of auditable debate?
Conclusions
===========
This writing aims to increase the variety and richness of experience available in the long term, by reducing the chance that people become unable to explore and express their values.
It seeks to do that by:
1. Pushing me to clarify my own thoughts and research direction. If I’m hoping to make things clear to a reader, they must first be clear to me.
2. Increasing the legibility of my research, so that it can be corrected and built upon.
3. Reaching others who can guide and advise my research.
4. Suggesting projects to people who might like to collaborate, whether working in parallel, with guidance, or independently.
In this writing, I have given a brief overview of my thoughts on Paternalism, Value Lock-in, and touched on some early technical ideas that seek to address these issues.
I have also tried to provide a selection of concrete research questions and real-world problems that could form the seed of further research.
If you’d like to get in touch, please do reach out.
Acknowledgements
================
I'm grateful to Shoshannah Tekofsky, Stanisław Malinowski, Edward Saperia, Kaïs Alayej, Blaine Rogers and Justis Mills for comments on drafts of this work, and to the EA Long Term Future Fund for enabling this research.
1. **[^](#fnref55tdxzgznkb)**See also: corrigibility, discussed by[Paul Cristiano](https://ai-alignment.com/corrigibility-3039e668638) and[MIRI](https://intelligence.org/files/Corrigibility.pdf). Quoting[Koen Holtman](https://arxiv.org/abs/1908.01695): "A corrigible agent will not resist attempts by authorized parties to alter the goals and constraints that were encoded in the agent when it was first started." To avoid Value Lock-in, however, we must also consider the corrigibility and entrenchment of these authorised parties.
2. **[^](#fnref5uv196m6zpr)**There is a question of how consent and liberty fit into a utilitarian framework.
Rule utilitarianism might propose that humans, being boundedly rational, pursue the greatest utility by pursuing these kind of higher-order values, but this does not necessarily apply to arbitrarily intelligent agents.
There is a plausible epistemic stance that a person is the best judge of what pleases or displeases them, or of what their own preferences are. However, a cornerstone of paternalism is that someone else might understand a person's preferences or welfare better than the person themselves.
Appeals to the inherent pleasure of self-rule only address the *appearance* of paternalism, and thus only require paternalistic acts to be sufficiently subtle or deceptive.
3. **[^](#fnrefl8n7c9ku3z)**This is similar to the question of[Impact Measures](https://www.lesswrong.com/tag/impact-measures), but whereas Impact Measures are usually thought of in the context of minimising undesired impact, a diversity of meaningful options would be something to maximise.
4. **[^](#fnrefz8w1dash0vp)**A monkey who collects bananas rather than potatoes might be said to "prefer" bananas, or to find them more rewarding. It's not clear to me how this distinguishes between revealed preferences, 'true' preferences, and non-goal-driven behaviour, especially when considering addictive behaviour. I'd like to expand on this in later writing.
5. **[^](#fnrefn6yhqq3u43j)**Let me know if you see another :)
6. **[^](#fnref9fd36oj0xoo)**If, no matter which action an agent takes, the probability of ending up in any given future state is unchanged, then the agent has zero empowerment. If instead for any given future state the agent can choose an action to guarantee it arising, then the agent is maximally empowered.
7. **[^](#fnref42jfuasj2pt)**An advantage which they press with their "Immediate Choice" proxy, which myopically restricts consideration to the next single time-step.
8. **[^](#fnref5zd4mnvmpfm)**Even something like the [Principle of Maximum Entropy](https://en.wikipedia.org/wiki/Principle_of_maximum_entropy), which [can be used for decision and prediction tasks](http://www.cs.cmu.edu/~bziebart/publications/maximum-causal-entropy.pdf), is an opinionated statement about the world. |
b6189e6b-b083-4e9d-93d8-d0e1d793eb2f | trentmkelly/LessWrong-43k | LessWrong | Why be moral if we can't measure how moral we are? Is it even possible to measure morality?
Imagine that you believe that the purpose of your life is to be the best marathon runner in the world, or even just an above average marathon runner. Despite the fact that you are committed to the mission, imagine that running is not particularly enjoyable to you, and your training regime is tough, but being a better marathon runner than others is such a noble goal you persist. Everyday, you train, painstakingly learning new running techniques, building stamina, pushing your limits. When you are not training, you are thinking about running. You dream of marathon running in your bed, think of it in the bathroom, conjecture about it on the way to work. Even if its detrimental to your happiness (although there is a high correlation between your self-perceived notion of how good of a marathon runner you are and your happiness), marathon running, you claim, is even more important to you than what you give up to obtain it. You routinely make sacrifices in order to run faster, whether that be avoiding certain foods, or forgoing leisure activities, or spending hard earned money on training. In fact, marathon running is so important to you, that you would almost say that everything in your life: your job, your relationships, your possessions, are really just a tool to increase your ability to be a good marathon runner.
Now, imagine every time you run, you are forced to do so in a VR environment. You are running in real life, but you are wearing a headset which work to manipulate your sensory, auditory, and tactile processing. As a result you are unable to gain any meaningful information about the speed at which you are running the marathons, or if you are making progress towards becoming a better runner or not. You are also not allowed to directly race with anyone else. Additionally, you are aware that many people are biased about how great of a marathon runner they they are. Many people who by consensus are slower marathon runners think they are great. Furthermore, even J |
5d93429e-e093-4eb1-bbde-bce1f2be0531 | trentmkelly/LessWrong-43k | LessWrong | [Book review] Atomic habits
This is a book review of the book Atomic Habits by James Clear. I read it in the context of a personal literature review project on the topic of productivity and well being. If you are more interested by advice on productivity and wellbeing than by this specific book I advise you to read the project report first, which condenses the advice from multiple sources (including this book).
How I read
I read this book entirely from start to finish while taking notes. This was done in a relatively short time span (3 days). I have since let a little time pass and have gathered my notes to produce a synthesis, which served as my main resource to write this post.
Description and opinion.
The first part of the book introduces a view on habits and their use for shaping oneself. To say it in a few words, the idea is that our identity shapes our behavior which determines the outcome we get. Habits are seen as a way to shape our identity, the way we view ourselves.
The rest of the book is mostly a list of techniques to shape one's habits. The last chapters give some extra advice on productivity.
Apart from that, the book is well written and intelligently structured to be easy to read without wasting too much of the reader's time. The author presents a theory of four "pillars" of habit shaping which serves to give a structure to the advice and techniques. I do not care much for the model of four mains axis for habits, but I like how it helped structure the book and the "cheat sheet", which is built progressively throughout the book.
Main takes
* The notion of identity, roughly equivalent to "what we consider normal for us to do" and its importance.
* The idea of using and shaping habits, notably as a way to shape said identity. I am tempted to speak of "auto manipulation" or "self taming".
* The idea that it is important and useful to do weekly reviews of our habits.
* Independently of the book, the cheat sheet is probably a good resource to skim. It is certainly relativ |
b2d9d9b7-9a33-4a01-a50b-57d6ede0ccbb | StampyAI/alignment-research-dataset/blogs | Blogs | A major grant from the Open Philanthropy Project
I’m thrilled to announce that the Open Philanthropy Project has awarded MIRI a **three-year $3.75 million general support grant** ($1.25 million per year). This grant is, by far, the largest contribution MIRI has received to date, and will have a major effect on our plans going forward.
This grant follows a [$500,000 grant](https://intelligence.org/2016/09/06/grant-open-philanthropy/) we received from the Open Philanthropy Project in 2016. The Open Philanthropy Project’s [announcement](https://www.openphilanthropy.org/focus/global-catastrophic-risks/potential-risks-advanced-artificial-intelligence/machine-intelligence-research-institute-general-support-2017) for the new grant notes that they are “now aiming to support about half of MIRI’s annual budget”.[1](https://intelligence.org/2017/11/08/major-grant-open-phil/#footnote_0_16783 "The Open Philanthropy Project usually prefers not to provide more than half of an organization’s funding, to facilitate funder coordination and ensure that organizations it supports maintain their independence. From a March blog post: “We typically avoid situations in which we provide >50% of an organization’s funding, so as to avoid creating a situation in which an organization’s total funding is ‘fragile’ as a result of being overly dependent on us.”") The annual $1.25 million represents 50% of a conservative estimate we provided to the Open Philanthropy Project of the amount of funds we expect to be able to usefully spend in 2018–2020.
This expansion in support was also conditional on our ability to raise the other 50% from other supporters. For that reason, I sincerely thank all of the past and current supporters who have helped us get to this point.
The Open Philanthropy Project has expressed openness to potentially increasing their support if MIRI is in a position to usefully spend more than our conservative estimate, if they believe that this increase in spending is sufficiently high-value, and if we are able to secure additional outside support to ensure that the Open Philanthropy Project isn’t providing more than half of our total funding.
We’ll be going into more details on our future organizational plans in a follow-up post **December 1**, where we’ll also discuss our end-of-the-year fundraising goals.
In their write-up, the Open Philanthropy Project notes that they have updated favorably about our technical output since 2016, following [our logical induction paper](https://intelligence.org/2016/09/12/new-paper-logical-induction/):
> We received a very positive review of MIRI’s work on “[logical induction](https://intelligence.org/2016/09/12/new-paper-logical-induction/)” by a machine learning researcher who (i) is interested in AI safety, (ii) is rated as an outstanding researcher by at least one of our close advisors, and (iii) is generally regarded as outstanding by the ML community. As mentioned above, we previously had [difficulty evaluating](https://www.openphilanthropy.org/focus/global-catastrophic-risks/potential-risks-advanced-artificial-intelligence/machine-intelligence-research-institute-general-support#Uncertainty_about_our_technical_assessment) the technical quality of MIRI’s research, and we previously could find no one meeting criteria (i) – (iii) to a comparable extent who was comparably excited about MIRI’s technical research. While we would not generally offer a comparable grant to any lab on the basis of this consideration alone, we consider this a significant update in the context of the original [case for the [2016] grant](https://www.openphilanthropy.org/focus/global-catastrophic-risks/potential-risks-advanced-artificial-intelligence/machine-intelligence-research-institute-general-support#Case_for_the_grant) (especially MIRI’s thoughtfulness on this set of issues, value alignment with us, distinctive perspectives, and history of work in this area). While the balance of our technical advisors’ opinions and arguments still leaves us skeptical of the value of MIRI’s research, the case for the statement “MIRI’s research has a *nontrivial chance* of turning out to be extremely valuable (when taking into account how different it is from other research on AI safety)” appears much more robust than it did before we received this review.
>
>
The announcement also states, “In the time since our initial grant to MIRI, we have made several more grants within this focus area, and are therefore less concerned that a larger grant will signal an outsized endorsement of MIRI’s approach.”
We’re enormously grateful for the Open Philanthropy Project’s support, and for their deep engagement with the AI safety field as a whole. To learn more about our discussions with the Open Philanthropy Project and their active work in this space, see the group’s previous [AI safety grants](https://www.openphilanthropy.org/giving/grants?field_focus_area_target_id_selective=532), our conversation with Daniel Dewey [on the Effective Altruism Forum](http://effective-altruism.com/ea/1ca/my_current_thoughts_on_miris_highly_reliable/), and the research problems outlined in the Open Philanthropy Project’s recent [AI fellows program description](https://www.openphilanthropy.org/focus/global-catastrophic-risks/potential-risks-advanced-artificial-intelligence/open-philanthropy-project-ai-fellows-program).
---
1. The Open Philanthropy Project usually prefers not to provide more than half of an organization’s funding, to facilitate funder coordination and ensure that organizations it supports maintain their independence. From a March [blog post](https://www.openphilanthropy.org/blog/technical-and-philosophical-questions-might-affect-our-grantmaking#Philanthropic_coordination_theory): “We typically avoid situations in which we provide >50% of an organization’s funding, so as to avoid creating a situation in which an organization’s total funding is ‘fragile’ as a result of being overly dependent on us.”
The post [A major grant from the Open Philanthropy Project](https://intelligence.org/2017/11/08/major-grant-open-phil/) appeared first on [Machine Intelligence Research Institute](https://intelligence.org). |
ee6d30d7-aae8-4fb9-bcf2-9bfd33e327ff | trentmkelly/LessWrong-43k | LessWrong | Self-modification is the correct justification for updateless decision theory
Reply to: Late great filter is not bad news
Suppose that you build an AI, and Omega appears to it and says:
Here's a button. A million years ago I calculated the umpteenth digit of pi. If it is even, I calculated whether you would press this button (in such a way that your human creator was never simulated as a conscious being). If I predicted that you wouldn't press the button, I destroyed Earth right then and there.* If it is odd, I created a doomsday device that will destroy the solar system if you press this button.
[* ETA: Assume that if the digit is even and the AI is predicted to press the button, Omega does not destroy Earth, but does turn Alpha Centauri purple (say). The point is for this to be a scenario that you, the AI creator, know not to have come to pass.]
Suppose you're the kind of AI creator whose AI is time consistent in a certain sense from the beginning of time and presses the button. Then you have an AI that satisfies a certain kind of philosopher, wins big in a certain logically impossible world, and destroys humanity.
Suppose, on the other hand, that you're a very similar kind of AI creator, only you program your AI not to take into account impossible possible worlds that had already turned out to be impossible (when you created the AI | when you first became convinced that timeless decision theory is right). Then you've got an AI that most of the time acts the same way, but does worse in worlds we know to be logically impossible, and destroys humanity less often in worlds we do not know to be logically impossible.
Wei Dai's great filter post seems to suggest that under UDT, you should be the first kind of AI creator. I don't think that's true, actually; I think that in UDT, you should probably not start with a "prior" probability distribution that gives significant weight to logical propositions you know to be false: do you think the AI should press the button if it was the first digit of pi that Omega calculated?
But obviously, you do |
0f72388a-431c-4cd8-a3ef-f227294a6447 | StampyAI/alignment-research-dataset/blogs | Blogs | generalized adding reality layers
generalized adding reality layers
---------------------------------
in [*predictablizing ethic deduplication*](predictablizing-ethic-deduplication.html), i talk about how when we don't know how reality works, we can task a singleton superintelligence with "adding a layer" to reality, which guarantees that inside that simulated reality we *are* able to function with known ethics.
in addition there's a sense in which, if one principle overrides the other, in general with [arbitrarily many layers of reality-simulation](above-paperclips-2.html) we should tend to favor whichever option overrides. so for example: if our reality is actually on top of 1000 layers of reality simulations, then it only takes **one** of them to be (truly, deeply) deduplicating for our universe and any sub-universe we simulate to also have deduplication.
or, more precisely, for any set of mutually exclusive traits with a dominance ordering (such as deduplication > no-deduplication), we can expect the take one of those shapes:
(click on the image to expand)
[](generalized-adding-reality-layers.svg)
i will call this the "generalized adding reality layers" (GARL) device, and i think it could have a broad use to reason about properties of the cosmos (the set of [instantiated](questions-cosmos-computations.html) universes), even ones that might seem axiomatic and [untestable](https://en.wikipedia.org/wiki/Newton's_Flaming_Laser_Sword).
for any set of mutually exclusive traits, we care about four properties:
* what the dominance ordering is between those traits
* how they affect the rate of spawning varied sub-universe
* how they affect the rate of spawning moral patient experiences
* how they affect the rate of spawning deeply caring actors
so, what other sets of traits can we examine using GARL ? here are some that i can think of off the top of my head, as well as my guess for the questions above..
| question | dominance ordering | most varied sub-universes | most moral patient experiences | most deeply caring actors |
| --- | --- | --- | --- | --- |
| [moral patient deduplication](deduplication-ethics.html) | dedup > no-dedup | unaffected ? | no-dedup > dedup | i've no idea |
| [infinite compute](hope-infinite-compute.html) ¹ | finite > infinite | infinite > finite | infinite > finite | infinite > finite ? |
| type of compute ¹ | classical > quantum > hyper | hyper > quantum > classical ? | unknown | hyper > quantum > classical ? |
| moral realism ² | realism > non-realism | unsure | whichever maximizes good | realism > non-realism ? |
| deeply-caring superintelligence | present > absent | depends on its goals | depends on its goals | present > absent |
¹: these two questions are similar to one another in that they have one dominant variant that restricts computation, and one recessive variant that doesn't; as a result, i would tend to assume that the recessive variant has a higher chance of spawning most kinds of stuff
²: my reasoning: once what is true becomes aligned with what is good, then the [orthogonality thesis](https://www.lesswrong.com/tag/orthogonality-thesis) becomes falsified in that sub-cosmos, and superintelligences are more easily aligned by default
other questions to which GARL may be applicable but i haven't figured out how:
* is occam razor's/[solomonoff induction](https://www.lesswrong.com/posts/EL4HNa92Z95FKL9R2/a-semitechnical-introductory-dialogue-on-solomonoff-1) applicable?
* [are minimal circuits daemon-free?](https://www.lesswrong.com/posts/nyCHnY7T5PHPLjxmN/open-question-are-minimal-circuits-daemon-free)
* [can you control the past?](https://www.lesswrong.com/posts/PcfHSSAMNFMgdqFyB/can-you-control-the-past)
* [SIA vs SSA](https://www.lesswrong.com/posts/RnrpkgSY8zW5ArqPf/sia-greater-than-ssa-part-1-learning-from-the-fact-that-you) |
4602afb4-9fc2-45c1-bbe1-d86af376ecc7 | trentmkelly/LessWrong-43k | LessWrong | Using GPT-N to Solve Interpretability of Neural Networks: A Research Agenda
Tl;dr We are attempting to make neural networks (NN) modular, have GPT-N interpret each module for us, in order to catch mesa-alignment and inner-alignment failures.
Completed Project
Train a neural net with an added loss term that enforces the sort of modularity that we see in well-designed software projects. To use this paper's informal definition of modularity
> a network is modular to the extent that it can be partitioned into sets of neurons where each set is strongly internally connected, but only weakly connected to other sets.
Example of a “Modular” GPT. Each module should be densely connected w/ relatively larger weights. Interfaces between modules should be sparsely connected w/ relatively smaller weights.
Once we have a Modular NN (for example, a GPT), we will use a normal GPT to map each module into a natural language description. Notice that there are two different GPT’s at work here.
GPT-N reads in each “Module” of the “Modular GPT”, outputting a natural language description for each module.
If successful, we could use GPT-N to interpret any modular NN in natural language. Not only should this help our understanding of what the model is doing, but it should also catch mesa-alignment and inner-alignment failures.
Cruxes
There are a few intuitions we have that go counter to other’s intuitions. Below is an elaboration of our thoughts and why we think this project could work.
Finding a Loss function that Induces Modularity
We currently think a Gomory-Hu Tree (GH Tree) captures the relevant information. We will initially convert a NN to a GH Tree to calculate the new loss function. This conversion will be computationally costly, though more progress can be made to calculate the loss function directly from the NN. See Appendix A for more details
Small NN’s are Human Interpretable
We’re assuming humans can interpret small NN’s, given enough time. A “Modular” NN is just a collection of small NN’s connected by sparse weights. If humans could in |
63432c60-222c-4c35-a422-b0e86be251a3 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Vancouver Cafe Meetup!
Discussion article for the meetup : Vancouver Cafe Meetup!
WHEN: 29 April 2012 01:00:00PM (-0700)
WHERE: Benny's Bagels 2505 W Broadway Vancouver, BC
Hey everybody.
Everyone from anywhere near Vancouver should all come out and hang out with us at Benny's in Kitsilano. Sunday at 13:00.
We are going to be discussing the reductionism sequence, but please come out anyway even if you don't have the time to read it this week; we will discuss in detail and make sure everybody is filled in by the end.
This is our big super out-of-the-basement meetup that everybody should come out to, so if you've been waiting for an excuse to come to one of our meetups come to this one. I hope to see some new faces!
As usual, most of our discussion happens on the mailing list, so join that.
No excuses, see you all there!
Discussion article for the meetup : Vancouver Cafe Meetup! |
2c49eb30-c851-469e-94e2-6baa7f2782cd | trentmkelly/LessWrong-43k | LessWrong | is there a big dictionary somewhere with all your jargon and acronyms and whatnot?
it would help newcomers |
db1c6f7c-2703-4e0f-9a3b-858f1e82d430 | trentmkelly/LessWrong-43k | LessWrong | Ordinary claims require ordinary evidence
Ordinary claims require ordinary evidence
(this is a post version of a YouTube video I made)
One common argument is that making extraordinary claims, such as AI being harmful, requires extraordinary evidence. However, I believe that asserting AI's potential to be harmful is not an extraordinary claim at all. Rather, it's grounded in several key axioms that, when examined, are hard to refute.
Why It's Not an Extraordinary Claim
I think the AI Optimist imagines a particular scenario or set of scenarios (perhaps "Terminator" or [insert fictional franchise here]) and says "that seems improbable". Perhaps Eliezer comes along and posits one additional scenario, and the Optimist says "all of those combined are improbable". "Do you have any proof that this [particular tiny set of scenarios] will happen!?" But the space of AI ruin is vast, any failure scenario would ruin everything.
To me, AI ruin seems to a natural consequence of five simple processes and conditions:
The Five Core Axioms
1. AI gets better, never worse: AI's intelligence, however you define it, is increasing. As new research emerges, the knowledge becomes a permanent part of the record. Like other technological advances, we build on it rather than regress. People constantly throw more resources at AI, training up bigger and bigger models without any regard to safety.
2. Intelligence always helps: Being more intelligent always aids success in the real world. A slight edge in intelligence has allowed humans to dominate the Earth. There is no reason to expect a different outcome with an entity more intelligent than humans.
3. No one knows how to align AI: No one can precisely instruct AI to align with complex human values or happiness. We can optimize for likely prediction of a data point, but no one has written a Python function to rank outcomes by how positive they are for humanity.
4. Resources are finite: Any AI acting in the real world will inevitably compete with humans for resources. These resou |
a3f76168-b6d7-42d3-937c-f9ab10ce2dd3 | trentmkelly/LessWrong-43k | LessWrong | People Will Listen
I have been thinking a lot about the crypto autopsy Scott posted in 2018. In retrospect, there was still an enormous amount of money to be made 'buying the dip' in BTC/ETH. And there was even more money to be made buying altcoins. Scott also links to this thread from 2015 strongly advising people to buy bitcoin at around $230 (so approximately 250x gains on the percent you held). The earlier bitcoin discussion on lesswrong might have represented an even more lucrative opportunity but this is some of the best completely explicit advice ever posted on the forum:
> LessWrong is where I learned about Bitcoin, several years ago, and my greatest regret is that I did not investigate it more as soon as possible, that people here did not yell at me louder that it was important, and to go take a look at it. In that spirit, I will do so now.
>
> ...
>
> This is a time to be good rationalists, and investigate a possible opportunity, comparing the present situation to historical examples, and making an informed decision. Either Bitcoin has begun the process of dying, and this decline will continue in stages until it hits zero (or some incredibly low value that is essentially the same for our purposes), or it will live. Based on the new all time high being hit in number of transactions, and ways to spend Bitcoin, I think there is at least a reasonable chance it will live. Enough of a chance that it is worth taking some money that you can 100% afford to lose, and making a bet. A rational gamble that there is a decent probability that it will survive, at a time when a large number of others are betting that it will fail.
>
> And then once you do that, try your hardest to mentally write it off as a complete loss, like you had blown the money on a vacation or a consumer good, and now it is gone, and then wait a long time.
As I am writing the thread itself has four upvotes. Conversely, the following comment has twenty-six (this thread long predates variable weight votes tho |
17eb5e81-a583-4ea4-a0ea-57fd1bf240c4 | trentmkelly/LessWrong-43k | LessWrong | Pieces of time
My friend used to have two ‘days’ each day, with a nap between—in the afternoon, he would get up and plan his day with optimism, whatever happened a few hours before washed away. Another friend recently suggested to me thinking of the whole of your life as one long day, with death on the agenda very late this evening. I used to worry, when I was very young, that if I didn’t sleep, I would get stuck in yesterday forever, while everyone else moved on to the new day. Right now, indeed some people have moved on to Monday, but I’m still winding down Sunday because I had a bad headache and couldn’t sleep. Which is all to say, a ‘day’ does not just mean a 24 hour measure of time, in our minds. Among its further significance, we treat it as a modular unit: we expect things within it to be more continuous and intermingled with each other than they are with things outside of it. What happens later today is more of a going concern at present than something that happens after sleeping. The events of this morning are more part of a continuous chapter, expected to flavor the present, than what happened yesterday. The same is true to some extent for weeks, months and years (but not for fortnights or periods of 105 hours).
I think days are well treated as modular like this because sleeping really separates them in relevant ways. I notice two other kinds of natural modular time-chunks that seem worth thinking in terms of, but which I don’t have good names for:
* Periods during which you are in one context and stream of thought (usually a minute to a few hours long). For instance the period of going for a walk, or the period between getting home and receiving a phone call that throws you into a new context and set of thoughts. During one such chunk, I can remember a lot about the series of thoughts so far, and build upon them. Whereas if I try to go back to them later, they are hard to bring back to life, especially the whole set of thoughts and feelings that I wandered around dur |
c7904e0b-24fc-4870-9ecf-ab3c32fa6477 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Washington Goals/Projects (try again!)
Discussion article for the meetup : Washington Goals/Projects (try again!)
WHEN: 22 December 2013 03:00:00PM (-0500)
WHERE: National Portrait Gallery, Washington, DC 20001, USA
We're going to try and get the goals meetup working again, or at least figure out why it didn't in the past. Also (very relatedly) we can talk about what we're working on now.
Discussion article for the meetup : Washington Goals/Projects (try again!) |
661614ea-5045-40b5-b099-270512042043 | trentmkelly/LessWrong-43k | LessWrong | A potentially great improvement to minimum wage laws to handle both economic efficiency as well as poverty concerns
Minimum wage has the side effect of leaving unemployed, the people who do not possess the requisite skill to command the minimum wage in the market. This follows from basic micro-economic theory and here is nice short video of Milton Friedman, the renowned United States economist arguing against minimum wage laws. To offset this problem, it is essential to have some kind of social security safety net for the unemployed.
Instead, I would like to propose the following scheme which seems to me as more efficient (please let me know in the comments if you know of any country that tried this or something close): Set x= 1.5 * min_wage. Have no minimum wage laws. And folks who earn y which is below x, receive (x-y)/2 in social security. This way, we have (1) lower "skilled" people contributing to the country's GDP in their own small way instead of being unemployed and contributing nothing. If we have a significant number of these guys, the numbers could really add up. (2) The min wage like concerns are taken care of with the government safety net. (3) There is still incentive for people earning below x to work. If we set the social security to something like (x-y), no one earning below x will have any incentive to work since their net in hand compensation would then always be x regardless of what they do.
This above scheme looks almost like a pareto improvement to me compared to minimum wage laws, supported by social security for the unemployed, because it does roughly as good with regard to supporting those whose skills are below the minimum wage, while ensuring less government spending on social security, since many of the formerly unemployed would now be in low wage jobs and the government simply has to top up their current salaries which might be well above zero. This is of course a good thing, because the government then has the option of either using the extra money to reduce the budget deficit and ensure better economic health of the country, or use the money f |
c47b0600-7bbe-47fc-8c54-1f624b53a65d | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Advice for new alignment people: Info Max
Epistemic status: *This text is meant for other people in a similar position to me who are deciding how they can help ensure that AI doesn’t kill us all.*
*I hope that the new perspective gives you some new ways of thinking about your role in alignment, even though it should be treated as just that, another perspective on the problem.*
As a young math and information nerd, I have a couple of ways of changing the world for the better. I believe that AI Safety is the most important thing I could be doing, and the question after that inevitably becomes what I should do in AI Safety. The most obvious answer is technical research, but that might not be the best way of getting high impact.
So I did the natural thing and asked some cool people what they thought. 80k has the advice to work on a problem area and develop aptitudes, but that is very unspecific. That standard advice is true, but it can be amplified like a **\*cough\***motherf\*cker.
When asking Connor Leahy from Conjecture what the hell I should do with my life, I got the answer that I should go to the Bay and create a tech company. This might not sound like something coming out of the mouth of someone with short timelines and a high percentage of doom, yet it did.
Connor told me that the idea behind it was that there would be pivotal points in the development of AI where we would need competent and trustworthy people to lead projects. People who are both trustworthy and competent are a very valuable resource in this world, yet it might seem very distant to someone just starting their career.
For me, becoming trustworthy was the new goal because it made sense when I think about fat-tailed impact. However, with my short timelines, I’ve realised that I can’t just take an ordinary 10-year route and do some cool research to become respected and trustworthy, as I just won’t have the time to do so.
So what am I doing instead? My current strategy is asking myself how I can become competent and respected in the quickest way possible. In my head, I refer to it as "Infomaxing."
Info Maxing
-----------
You might now wonder, "How do you define info maxing?" Info maxing, in my head, is the strategic acquisition of a wide array of skills and experiences to enhance one's overall competency or, in short, maximising for valuable mental models. But to what end?
In my opinion, based on Connor's advice, the ultimate objective of info maxing, especially in the context of AI safety, is to position oneself at the helm during the crucial junctures of AI development. The times when competent and trustworthy leadership will matter the most. You're info maxing to become the 'captain' when the ship is navigating through the stormiest seas.
But let's delve deeper into what this might mean in AI Safety.
The general idea is that we’re going to need trustworthy and competent people to help out with projects when the going gets rough. This might be different depending on what model you have about the world, but if you want to position yourself in such a way were you can be a “captain” of a ship, the following might be useful:
**Infomaxing for Technical Proficiency:** One can't navigate the tumultuous waters of AI without a robust technical understanding. It's crucial to familiarise yourself with the ins and outs of AI and machine learning. This doesn't mean that you need to be the one to develop groundbreaking new algorithms. Still, you do need to understand the underlying technology well enough to make informed decisions and to communicate effectively with those who do.
**Infomaxing for Leadership:** A leader is more than just a decision-maker. A great leader inspires trust and confidence in their team. In the AI safety world, such trust is particularly critical. You'll be guiding the team during potentially high-stake situations. You’ll also have to be trusted by external sources to fulfil the duties placed on you. This is where strategic thinking, communication, and emotional intelligence become indispensable.
**Infomaxing for Vision:**Finally, it's important to develop a clear vision of what 'good' looks like in the context of AI safety. This vision will guide your actions and decisions and give you the clarity and resolve you'll need when facing difficult choices.
This is not sufficient to capture all of the skills that you need, but the important thing to take into account is that pure technical research may not be the best way for you to make an impact on the world. Vision and leadership might be just as important.
Personal experience of info maxing
----------------------------------
*(9 months of having it in the back of my head)*
It probably makes sense to see how someone is putting this into practice. I hope someone can take some wisdom from it.
How I do this in practice is that I spend my time in an agile framework where I try to start projects & iterate on why things fail. The things I’m currently doing in my info maxing are the following :
**Starting organisations & projects:**I’ve tried to be very proactive when starting new things as it seems like one of the best ways the world provides real-world feedback loops. The average entrepreneur will start 10 projects before they go net-positive with one, and I’m expecting something similar for my projects at the moment. I’m still in the exploration phase of starting things up, but once I find something that works well, I will double down on that.
**Trying to work out a research agenda:**I’ve been working on understanding Active Inference, Boundaries, Natural Abstractions & Garrabrant’s work to come up with a nice definition of Dynamic Membranes (Dynamic [Boundaries](https://www.lesswrong.com/posts/8oMF8Lv5jiGaQSFvo/boundaries-part-1-a-key-missing-concept-from-utility-theory)) to predict agency within neural networks. It may or may not lead somewhere, but I feel that I’ve become a lot more experienced on the research front since I started.
A side note is that it will probably be better for people, in general, to work on ML Engineering if they don’t find themselves super excited about theoretical stuff. I can become as giddy as a kid when I hear about some cool new frame on agent foundations, so I felt I should go down this route more.
**Testing personal boundaries & working conditions:**I’ve been trying different ways of working and living. I’ve realised I can’t have more than two major projects simultaneously whilst being productive and mentally sane. I also need a strict schedule with deep work and shallow work scheduling to work.
I’ve looked into a lot of psycho and biohacking, and I can happily work 2x the amount I could a year ago just due to optimising my workflow. This was mostly done through increasing my energy levels through good sleep, energy-level-based scheduling, a change to low carbs until 4 pm, caffeine scheduling, yoga nidra, exercise and project minimalism. (implementing [Huberman](https://www.youtube.com/@hubermanlab) with some [multimodal agents of mind](https://www.lesswrong.com/s/ZbmRyDN8TCpBTZSip) on the side)
I believe that the compounding effects from this experimentation will most likely be worth quite a lot, as I know where my work boundaries are, and I’ve doubled the amount of deep work I can do during a week.
**Reading & implementing book knowledge:**Whilst I’ve been doing the above, I’ve been reading non-fiction books (now at 3x the speed, hats off to nonlinear for the advice) related to the work and self-improvement I’ve been doing. I’ve been trying to implement the knowledge that the authors have had in my daily life so that I get short feedback loops between the two.
**Iterate & steal:**In short, I’ve tried to throw myself out into the wide-open world of feedback loops, and I’ve tried to steal as many applicable models from other people as possible. Regarding projects, my current thinking is "the more personal responsibility that I need to take over a project, the more I grow used to responsibility." As John Wentworth says somewhere, it’s helpful to practice working on a hard problem for a while if you want to work on alignment. And as I say, it is more useful to just work on alignment.
If you want to become the captain of a ship, build a boat and sail it.
With some caveats, of course: You don’t want to go into too deep water because then you might die (burnout), and it might also make sense to learn from experienced captains (read books, join boot camps & courses), etc. The important thing is that you do these things whilst also doing the main activity.
**Sail**, make sure that you **sail**.
Summary
-------
In short, think of building skills that will make you a person that can fill the bottleneck that will be needed the most. If your fit allows it, try to become someone who can shoulder responsibility, for people like that will be bloody needed in the future.
The way to become that person is most likely info maxing, which is done by gathering useful models of acting in the world. The real world provides the best feedback loops, so if you want to become a captain, make sure that you practice sailing a boat. |
f02804b6-4e29-429d-8061-97c78ecdfcd0 | trentmkelly/LessWrong-43k | LessWrong | [AN #80]: Why AI risk might be solved without additional intervention from longtermists
Find all Alignment Newsletter resources here. In particular, you can sign up, or look through this spreadsheet of all summaries that have ever been in the newsletter. I'm always happy to hear feedback; you can send it to me by replying to this email.
Audio version here (may not be up yet).
Welcome to another special edition of the newsletter! In this edition, I summarize four conversations that AI Impacts had with researchers who were optimistic that AI safety would be solved "by default". (Note that one of the conversations was with me.)
While all four of these conversations covered very different topics, I think there were three main points of convergence. First, we were relatively unconvinced by the traditional arguments for AI risk, and find discontinuities relatively unlikely. Second, we were more optimistic about solving the problem in the future, when we know more about the problem and have more evidence about powerful AI systems. And finally, we were more optimistic that as we get more evidence of the problem in the future, the existing ML community will actually try to fix that problem.
Conversation with Paul Christiano (Paul Christiano, Asya Bergal, Ronny Fernandez, and Robert Long) (summarized by Rohin): There can't be too many things that reduce the expected value of the future by 10%; if there were, there would be no expected value left (ETA: see this comment). So, the prior that any particular thing has such an impact should be quite low. With AI in particular, obviously we're going to try to make AI systems that do what we want them to do. So starting from this position of optimism, we can then evaluate the arguments for doom. The two main arguments: first, we can't distinguish ahead of time between AIs that are trying to do the right thing, and AIs that are trying to kill us, because the latter will behave nicely until they can execute a treacherous turn. Second, since we don't have a crisp concept of "doing the right thing", we can't select AI s |
c7affa13-ae6d-43df-97c5-f8139ffdb758 | trentmkelly/LessWrong-43k | LessWrong | other people are occasionally rather weird
So here's a dynamic I find myself in sometimes:
I've just explained a problem to a friend / family member and they offer advice that feels plain wrong.
Normally this leaves me either:
* trying to somehow make the advice fit (it doesn't)
* or deciding maybe this person is not as great as I thought they were (They’re pretty great so I’m not crazy keen to do this either)
Here's hopefully, pointing at a third alternative.
What might be going on: When I hear another persons problem. What this brain tends to do is run a simulation of my friend in that situation. Yet, the complexity of my human friend is much greater than my simulation of them. There's bound to be a bit of compression.
I've heard, how humans store faces is that we have a default face template. And then to minimise storage cost a new friend's face is stored as lots of "deviations" from that default [1].
I'd be pretty surprised if something similar isn't going on with my models of other people. When I imagine a friend, they're basically me but with a few tweaks. Oh yeah, Sally doesn't like noise or John's real tidy.
Here instead are some intuitions behind other people being super weird:
* Many folks showering habits are super strange. From The Typical Sex Life Fallacy, Thing of Things:
> People may take anywhere between five minutes and forty minutes to shower. They may wash their hair daily, once a week, or not at all. They may wash their bodies thoroughly, only clean the parts that look dirty, only clean certain parts (such as the armpits or genitals), or just stand under the water. They may use a loofah, a sponge, or nothing. They may bring in a comb to comb out the conditioner. They may sing. They may zone out. They may jerk off. They may bathe instead, and bathing may involve reading a book or bath bombs or lighting candles and drinking a nice bottle of wine or bubble bath or none of those things at all. The one thing that is consistent is that everyone thinks the way they shower is the way |
bac2908b-f7f3-4e97-87a5-a4510e4b8706 | trentmkelly/LessWrong-43k | LessWrong | Policy restrictions and Secret keeping AI
Status: outlining a reasonably interesting idea I haven't seen discussed here before. Just come up with it so might not be fully thought out. May be a repeat of something. Rough draft quality, grammar may be lacking.
Suppose you have some set of worlds W and some set of actions A . A policy is a function P:W→ΔA from worlds to probability distributions over actions. (worlds to actions is good too, but probability distributions are helpful for Secret keeping AI in particular)
You also have a prior probability distribution over worlds and a utility function U:(W×A)→R . This gives a function from policies to reals. Given reasonable assumptions, we can pick the best policy from any set of allowable policies.
Now some of these policies are physically impossible, they require you to act differently based on information you don't yet know, or will never find out. Throw these policies out.
Let Π be any set of policies formed by taking the intersection of the physically possible policies, with those that pass some other test. (I will show later a simple other test that yields meaningful and potentially useful behaviour) Our AI then takes the highest expected utility policy in Π .
Note that all sorts of real world phenomena can be described in terms of physics constraints on policies. From communication and memory restrictions between different parts of the AI, to the AI running on a reversible or quantum computer that must implement a bijective or unitary function.
Assuming bounded utility functions, these agents have a nice property, they don't care much about worlds of small probability measure. If the agent hasn't seen extremely improbable input data, then extremely unlikely worlds won't influence the choice of actions, unless that choice was basically a tie anyway. (conjecture: If perturbing all the utilities by any perturbation of magnitude at most epsilon doesn't change the action on most worlds, then perturbing the utilities on a small measure of worlds doesn't c |
523d7193-d088-445a-bc1c-5b577c7fc678 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | My computational framework for the brain
*(See comment* [*here*](https://www.lesswrong.com/posts/diruo47z32eprenTg/my-computational-framework-for-the-brain?commentId=c6qzzKWAEfTQyoK7F#c6qzzKWAEfTQyoK7F) *for some updates and corrections and retractions. —Steve, 2022)*
By now I've written a bunch of blog posts on brain architecture and algorithms, not in any particular order and generally interspersed with long digressions into Artificial General Intelligence. Here I want to summarize my key ideas in one place, to create a slightly better entry point, and something I can refer back to in certain future posts that I'm planning. If you've read every single one of my previous posts (hi mom!), there's not much new here.
In this post, I'm trying to paint a picture. I'm not really trying to justify it, let alone prove it. The justification ultimately has to be: All the pieces are biologically, computationally, and evolutionarily plausible, and the pieces work together to explain absolutely everything known about human psychology and neuroscience. (I believe it! Try me!) Needless to say, I could be wrong in both the big picture and the details (or missing big things). If so, writing this out will hopefully make my wrongness easier to discover!
Pretty much everything I say here *and its opposite* can be found in the cognitive neuroscience literature. (It's a controversial field!) I make no pretense to originality (with one exception noted below), but can't be bothered to put in actual references. My previous posts have a *bit* more background, or just ask me if you're interested. :-P
So let's start in on the 7 guiding principles for how I think about the brain:
1. Two subsystems: "Neocortex" and "Subcortex"
----------------------------------------------
(Update: I have a revised discussion of this topic at my later post [Two Subsystems: Learning and Steering](https://www.alignmentforum.org/posts/hE56gYi5d68uux9oM/intro-to-brain-like-agi-safety-3-two-subsystems-learning-and).)
This is the starting point. I think it's absolutely critical. The brain consists of two subsystems. The **neocortex** is the home of "human intelligence" as we would recognize it—our beliefs, goals, ability to plan and learn and understand, every aspect of our conscious awareness, etc. etc. (All mammals have a neocortex; birds and lizards have an homologous and functionally-equivalent structure called the "pallium".) Some other parts of the brain (hippocampus, parts of the thalamus & basal ganglia & cerebellum—see further discussion [here](https://www.lesswrong.com/posts/jrewt3rLFiKWrKuyZ/big-picture-of-phasic-dopamine)) help the neocortex do its calculations, and I lump them into the "neocortex subsystem". I'll use the term **subcortex** for the rest of the brain (brainstem, hypothalamus, etc.).
* *Aside: Is this the triune brain theory?* No. [Triune brain theory](https://en.wikipedia.org/wiki/Triune_brain) is, from what I gather, a collection of ideas about brain evolution and function, most of which are wrong. One aspect of triune brain theory is putting a lot of emphasis on the distinction between neocortical calculations and subcortical calculations. I like that part. I'm keeping that part, and I'm improving it by expanding the neocortex club to also include the thalamus, hippocampus, lizard pallium, etc., and then I'm ignoring everything else about triune brain theory.
2. Cortical uniformity
----------------------
I claim that the neocortex is, to a first approximation, [architecturally uniform](https://www.lesswrong.com/posts/8F8dagB4q4BzR5JNz/gary-marcus-vs-cortical-uniformity), i.e. all parts of it are running the same generic learning algorithm in a massively-parallelized way.
**The two caveats to cortical uniformity** (spelled out in more detail at [that link](https://www.lesswrong.com/posts/8F8dagB4q4BzR5JNz/gary-marcus-vs-cortical-uniformity)) are:
* There are sorta "hyperparameters" on the generic learning algorithm which are set differently in different parts of the neocortex—for example, different regions have different densities of each neuron type, different thresholds for making new connections (which also depend on age), etc. This is not at all surprising; all learning algorithms inevitably have tradeoffs whose optimal settings depend on the domain that they're learning [(no free lunch).](https://en.wikipedia.org/wiki/No_free_lunch_in_search_and_optimization)
+ As one of many examples of how even "generic" learning algorithms benefit from domain-specific hyperparameters, if you've seen a pattern "A then B then C" recur 10 times in a row, you will start unconsciously expecting AB to be followed by C. But "should" you expect AB to be followed by C after seeing ABC only 2 times? Or what if you've seen the pattern ABC recur 72 times in a row, but then saw AB(not C) twice? What "should" a learning algorithm expect in those cases? The answer depends on the domain—how regular vs random are the environmental patterns you're learning? How stable are they over time? The answer is presumably different for low-level visual patterns vs motor control patterns etc.
* There is a gross wiring diagram hardcoded in the genome—i.e., set of connections between different neocortical regions and each other, and other parts of the brain. These connections later get refined and edited during learning. These make the learning process faster and more reliable by bringing together information streams with learnable relationships—for example the wiring diagram seeds strong connections between toe-related motor output areas and toe-related proprioceptive (body position sense) input areas. We *can* learn relations between information streams without any help from the innate wiring diagram, by routing information around the cortex in more convoluted ways—see the Ian Waterman example [here](https://www.lesswrong.com/posts/isDCEmYHsosyAjLRK/predictive-coding-and-motor-control)—but it's slower, more limited, and may consume conscious attention. Related to this is a diversity of training signals: for example, different parts of the neocortex are trained to predict different signals, and also different parts of the neocortex get [different dopamine training signals](https://www.lesswrong.com/posts/jrewt3rLFiKWrKuyZ/big-picture-of-phasic-dopamine)—or even [none at all](https://www.lesswrong.com/posts/szeKeZwuQhFxirfBY/is-rl-involved-in-sensory-processing).
3. Blank-slate neocortex
------------------------
(...But not blank-slate *sub*cortex! More on that below.)
(Update: To avoid confusion, I've more recently been calling this concept "learning-from-scratch"—see discussion in my later post [“Learning from Scratch” in the brain](https://www.alignmentforum.org/posts/wBHSYwqssBGCnwvHg/intro-to-brain-like-agi-safety-2-learning-from-scratch-in).)
I claim that the neocortex (and the rest of the telencephalon and cerebellum) starts out as a "blank slate": Just like an ML model initialized with random weights, the neocortex cannot make any correct predictions or do anything useful until it learns to do so from previous inputs, outputs, and rewards.
In more neuroscience-y (and maybe less provocative) terms, I could say instead: the neocortex is a memory system. It's a *really fancy* memory system—it's highly structured to remember particular kinds of patterns and their relationships, and it comes with a sophisticated query language and so on—but at the end of the day, it's still a type of memory. And like any memory system, it is useless to the organism until it gradually accumulates information. (Suggestively, if you go far enough back, the neocortex and hippocampus evolved out of the same ancient substructure ([ref](https://doi.org/10.3758/s13414-019-01760-1)).)
(By the way, I am not saying that the neocortex's algorithm is similar to today's ML algorithms. [There's more than one blank-slate learning algorithm](https://www.lesswrong.com/posts/SkcM4hwgH3AP6iqjs/can-you-get-agi-from-a-transformer)! See image.)
A "blank slate" learning algorithm, as I'm using the term, is one that learns information "from scratch"—an example would be a Machine Learning model that starts with random weights and then proceeds with gradient descent. When you imagine a "blank slate" learning algorithm, you should not imagine an empty void that gets filled with data. You should imagine a machine that learns more and better patterns over time, and writes those patterns into a memory bank—and "blank slate" just means that the memory bank starts out empty. There are many such machines, and they will learn different patterns and therefore do different things. See next section, and see also the discussion of hyperparameters in the previous section.Why do I think that the neocortex starts from a blank slate? Two types of reasons:
* *Details of how I think the neocortical algorithm works:* This is the main reason for me.
+ For example, as I mentioned [here](https://www.lesswrong.com/posts/8F8dagB4q4BzR5JNz/gary-marcus-vs-cortical-uniformity), there's [a theory](https://www.frontiersin.org/articles/10.3389/fncom.2017.00111/full) I like that says that all feedforward signals (I'll define that in the next section) in the neocortex—which includes all signals coming into the neocortex from the outside it, plus many cortex-to-cortex signals—are re-encoded into the data format that the neocortex can best process—i.e. a set of sparse codes, with low overlap, uniform distribution, and some other nice properties—and this re-encoding is done *by a pseudorandom process*! If that's right, it would seem to categorically rule out anything but a blank-slate starting point.
+ More broadly, we *know* the algorithm can learn new concepts, and new relationships between concepts, without having any of those concepts baked in by evolution—e.g. learning about rocket engine components. So why not consider the possibility that that's *all* it does, from the very beginning? I can see vaguely how that would work, why that would be biologically plausible and evolutionarily adaptive, and I can't currently see any other way that the algorithm can work.
* *Absence of evidence to the contrary:* I have a post [Human Instincts, Symbol Grounding, and the Blank-Slate Neocortex](https://www.lesswrong.com/posts/NkSpukDkm9pjRdMdB/human-instincts-symbol-grounding-and-the-blank-slate) where I went through a list of universal human instincts, and didn't see anything inconsistent with a blank-slate neocortex. The subcortex—which is absolutely *not* a blank slate—plays a big role in most of those; for example, the mouse has a [brainstem bird-detecting circuit wired directly to a brainstem running-away circuit](https://doi.org/10.1038/s41586-018-0244-6). (More on this in a later section.) Likewise I've read about the capabilities of newborn humans and other animals, and still don't see any problem. I accept all challenges; try me!
4. What is the neocortical algorithm?
-------------------------------------
**4.1. "Analysis by synthesis" + "Planning by probabilistic inference"**
"Analysis by synthesis" means that the neocortex searches through a space of generative models for a model that predicts its upcoming inputs (both external inputs, like vision, and internal inputs, like proprioception and reward). "Planning by probabilistic inference" (term from [here](http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.13.9135&rep=rep1&type=pdf)) means that we treat our own actions as probabilistic variables to be modeled, just like everything else. In other words, the neocortex's output lines (motor outputs, hormone outputs, etc.) are the same type of signal as any generative model prediction, and processed in the same way.
Here's how those come together. As discussed in [Predictive Coding = RL + SL + Bayes + MPC](https://www.lesswrong.com/posts/cfvBm2kBtFTgxBB7s/predictive-coding-rl-sl-bayes-mpc), and shown in this figure below:
* The neocortex *favors* generative models that have been making correct predictions, and *discards* generative models that have been making predictions that are contradicted by input data (or by other favored generative models).
* *And,* the neocortex *favors* generative models which predict larger future reward, and *discards* generative models that predict smaller (or more negative) future reward.
This combination allows both good epistemics (ever-better understanding of the world), and good strategy (planning towards goals) in the same algorithm. This combination *also* has some epistemic and strategic failure modes—e.g. a propensity to wishful thinking—but in a way that seems compatible with human psychology & behavior, which is likewise not perfectly optimal, if you haven't noticed. Again, see the link above for further discussion.
Criteria by which generative models rise to prominence in the neocortex; see [Predictive Coding = RL + SL + Bayes + MPC](https://www.lesswrong.com/posts/cfvBm2kBtFTgxBB7s/predictive-coding-rl-sl-bayes-mpc) for detailed discussion. Note that [(e) is implemented by a very different mechanism than the other parts](https://www.lesswrong.com/posts/jrewt3rLFiKWrKuyZ/big-picture-of-phasic-dopamine).* *Aside: Is this the same as Predictive Coding / Free-Energy Principle?* Sorta. I've read a fair amount of "mainstream" predictive coding (Karl Friston, Andy Clark, etc.), and there are a few things about it that I like, including the emphasis on generative models predicting upcoming inputs, and the idea of treating neocortical outputs as just another kind of generative model prediction. It also has a lot of other stuff that I disagree with (or don't understand). My account differs from theirs mainly by (1) emphasizing multiple simultaneous generative models that compete & cooperate (cf. ["society of mind"](https://www.amazon.com/Society-Mind-Marvin-Minsky/dp/0671657135), [multiagent models of mind](https://www.lesswrong.com/s/ZbmRyDN8TCpBTZSip), etc.), rather than "a" (singular) prior, and (2) restricting discussion to the neocortex subsystem, rather than trying to explain the brain as a whole. In both cases, this may be partly a difference of emphasis & intuitions, rather than fundamental. But I think the core difference is that predictive coding / FEP takes some processes to be foundational principles, whereas I think that those same things do happen, but that they're emergent behaviors that come out of the algorithm under certain conditions. For example, in [Predictive Coding & Motor Control](https://www.lesswrong.com/posts/isDCEmYHsosyAjLRK/predictive-coding-and-motor-control) I talk about the predictive-coding story that proprioceptive predictions are literally exactly the same as motor outputs. Well, I don't think they're exactly the same. But I *do* think that proprioceptive predictions and motor outputs are the same in *some* cases (but not others), in *some* parts of the neocortex (but not others), and *after* (but not before) the learning algorithm has been running a while. So I kinda wind up in a similar place as predictive coding, in some respects.
**4.2. Compositional generative models**
Each of the generative models consists of predictions that other generative models are on or off, and/or predictions that input channels (coming from outside the neocortex—vision, hunger, reward, etc.) are on or off. ("It's symbols all the way down.") All the predictions are attached to confidence values, and both the predictions and confidence values are, in general, functions of time (or of other parameters—I'm glossing over some details). The generative models are compositional, because if two of them make disjoint and/or consistent predictions, you can create a new model that simply predicts that both of those two component models are active simultaneously. For example, we can snap together a "purple" generative model and a "jar" generative model to get a "purple jar" generative model. They are also compositional in other ways—for example, you can time-sequence them, by making a generative model that says "Generative model X happens and then Generative model Y happens".
*PGM-type message-passing:* Among other things, the search process for the best set of simultaneously-active generative model involves something at least vaguely analogous to message-passing (belief propagation) in a probabilistic graphical model. [Dileep George's vision model](https://doi.org/10.1126/science.aag2612) is a well-fleshed-out example.
*Hierarchies are part of the story but not everything:* Hierarchies are a special case of compositional generative models. A generative model for an image of "85" makes a strong prediction that there is an "8" generative model positioned next to a "5" generative model. The "8" generative model, in turn, makes strong predictions that certain contours and textures are present in the visual input stream.
However, not all relations are hierarchical. The "is-a-bird" model makes a medium-strength prediction that the "is-flying" model is active, *and* the "is-flying" model makes a medium-strength prediction that the "is-a-bird" model is active. Neither is hierarchically above the other.
As another example, the brain has a visual processing hierarchy, but as I understand it, studies show that the brain has loads of connections that don't respect the hierarchy.
*Feedforward and feedback signals:* There are two important types of signals in the neocortex.
A **"feedback"** signal is a generative model prediction, attached to a confidence level, which includes all the following:
* "I predict that neocortical input line #2433 will be active, with probability 0.6".
* "I predict that generative model #95738 will be active, with probability 0.4".
* "I predict that neocortical output line #185492 will be active, with probability 0.98"—and this one is a self-fulfilling prophecy, as the feedback signal is also the output line!
A **"feedforward"** signal is an announcement that a certain signal is, in fact, active right now, which includes all the following:
* "Neocortical input line #2433 is currently active!"
* "Generative model #95738 is currently active!"
There are about 10× more feedback connections than feedforward connections in the neocortex, I guess for algorithmic reasons I don't currently understand.
In a hierarchy, the top-down signals are feedback, and the bottom-up signals are feedforward.
The terminology here is a bit unfortunate. In a motor output hierarchy, we think of information flowing "forward" from high-level motion plan to low-level muscle control signals, but that's the feed*back* direction. The forward/back terminology works better for sensory input hierarchies. Some people say "top-down" and "bottom-up" instead of "feedback" and "feedforward" respectively, which is nice and intuitive for both input and output hierarchies. But then *that* terminology gets confusing when we talk about non-hierarchical connections. Oh well.
(I'll also note here that "mainstream" predictive coding discussions sometimes talk about feedback signals being associated with confidence *intervals* for analog feedforward signals, rather than confidence *levels* for binary feedforward signals. I changed it on purpose. I like my version better.)
5. The subcortex steers the neocortex towards biologically-adaptive behaviors.
------------------------------------------------------------------------------
The blank-slate neocortex can learn to predict input patterns, but it needs guidance to do biologically adaptive things. **So one of the jobs of the subcortex is to try to** [**"steer"**](https://www.lesswrong.com/posts/SJXujr5a2NcoFebr4/mesa-optimizers-vs-steered-optimizers) **the neocortex, and the subcortex's main tool for this task is its ability to send rewards to the neocortex at the appropriate times**. Everything that humans reliably and adaptively do with their intelligence, from liking food to making friends, depends on the various reward-determining calculations hardwired into the subcortex.
6. The neocortex is a black box from the perspective of the subcortex. So steering the neocortex is tricky!
-----------------------------------------------------------------------------------------------------------
Only the neocortex subsystem has an intelligent world-model. Imagine you just lost a big bet, and now you can't pay back your debt to the loan shark. That's bad. The subcortex (hypothalamus & brainstem) needs to send negative rewards to the neocortex. But how can it know? How can the subcortex have any idea what's going on? It has no concept of a "bet", or "debt", or "payment" or "loan shark".
This is a very general problem. I think there are two basic ingredients in the solution.
Here's a diagram to refer to, based on the one I put in [Inner Alignment in the Brain](https://www.lesswrong.com/posts/DWFx2Cmsvd4uCKkZ4/inner-alignment-in-the-brain):
Schematic illustration of some aspects of the relationship between subcortex & neocortex. See also my previous post [Inner Alignment in the Brain](https://www.lesswrong.com/posts/DWFx2Cmsvd4uCKkZ4/inner-alignment-in-the-brain) for more on this. **(Update June 2021: I would no longer draw the diagram this way, see** [**here**](https://www.alignmentforum.org/posts/qNZSBqLEh4qLRqgWW/intro-to-brain-like-agi-safety-6-big-picture-of-motivation)**. The biggest difference is: I would not draw a direct line from neocortex to a hormone change (for example); instead the cortex would ask the subcortex (hypothalamus + brainstem) to make that hormone change, and then the subcortex might or might not comply with that recommendation. (I guess the way I drew it here is more like** **somatic marker hypothesis****.)))****6.1 The subcortex can learn what's going on in the world via its own, parallel, sensory-processing system.**
Thus, for example, we have the well-known visual processing system in our visual cortex, *and* we have the lesser-known visual processing system in our midbrain (superior colliculus). Ditto for touch, smell, proprioception, nociception, etc.
While they have similar inputs, these two sensory processing systems could not be more different!! The neocortex fits its inputs into a huge, open-ended predictive world-model, but the subcortex instead has a small and hardwired "ontology" consisting of evolutionarily-relevant inputs that it can recognize like faces, human speech sounds, spiders, snakes, looking down from a great height, various tastes and smells, stimuli that call for flinching, stimuli that one should orient towards, etc. etc., and these hardwired recognition circuits are connected to hardwired responses.
For example, babies learn to recognize faces quickly and reliably in part because the midbrain sensory processing system knows what a face looks like, and when it sees one, it will saccade to it, and thus the neocortex will spend disproportionate time building predictive models of faces.
...Or better yet, instead of saccading to faces itself, the subcortex can *reward the neocortex* each time it detects that it is looking at a face! Then the neocortex will go off looking for faces, using its neocortex-superpowers to learn arbitrary patterns of sensory inputs and motor outputs that tend to result in looking at people's faces.
**6.2 The subcortex can see the neocortex's outputs—which include not only prediction but imagination, memory, and empathetic simulations of other people.**
~~For example, if the neocortex never predicts or imagines any reward, then the subcortex can guess that the neocortex has a grim assessment of its prospects for the future—I'll discuss that particular example much more in an upcoming post on depression.~~ *(Update: that was wrong; see better discussion* [*here*](https://www.lesswrong.com/posts/jrewt3rLFiKWrKuyZ/big-picture-of-phasic-dopamine)*.)*
To squeeze more information out of the neocortex, the subcortex can also "teach" the neocortex to reveal when it is thinking of one of the situations in the subcortex's small hardwired ontology (faces, spiders, sweet tastes, etc.—see above). For example, if the subcortex rewards the neocortex for cringing in advance of pain, then the neocortex will learn to favor pain-prediction generative models that also send out cringe-motor-commands. And thus, eventually, it will *also* start sending weak cringe-motor-commands when imagining future pain, or when empathically simulating someone in pain—and the subcortex can detect *that*, and issue hardwired responses in turn.
(Update: I now think "the subcortex rewards the neocortex for cringing in advance of pain" is probably not quite the right mechanism, see [here](https://www.lesswrong.com/posts/jNrDzyc8PJ9HXtGFm/supervised-learning-of-outputs-in-the-brain).)
See [Inner Alignment in the Brain](https://www.lesswrong.com/posts/DWFx2Cmsvd4uCKkZ4/inner-alignment-in-the-brain) for more examples & discussion of all this stuff about steering.
Unlike most of the other stuff here, I haven't seen *anything* in the literature that takes "how does the subcortex steer the neocortex?" to be a problem that needs to be solved, let alone that solves it. (Let me know if you have!) ...Whereas I see it as *The Most Important And Time-Sensitive Problem In All Of Neuroscience*—because if we build neocortex-like AI algorithms, we will need to know how to steer them towards safe and beneficial behaviors!
7. The subcortical algorithms remain largely unknown
----------------------------------------------------
I think much less is known about the algorithms of the subcortex (brainstem, hypothalamus, ~~amygdala~~, etc.) (Update: After further research I have promoted the amygdala up to the neocortex subsystem, see discussion [here](https://www.lesswrong.com/posts/jrewt3rLFiKWrKuyZ/big-picture-of-phasic-dopamine)) than about the algorithms of the neocortex. There are a couple issues:
* *The subcortex's algorithms are more complicated than the neocortex's algorithms:* As described above, I think the neocortex has more-or-less one generic learning algorithm. Sure, it consists of many interlocking parts, but it has an overall logic. The subcortex, by contrast, has circuitry for detecting and flinching away from an incoming projectile, circuitry for detecting spiders in the visual field, circuitry for (somehow) implementing lots of different social instincts, etc. etc. I doubt all these things strongly overlap each other, though I don't know that for sure. That makes it harder to figure out what's going on.
+ I don't think the algorithms are "complicated" in the sense of "mysterious and sophisticated". Unlike the neocortex, I don't think these algorithms are doing anything where a machine learning expert couldn't sit down and implement something functionally equivalent in PyTorch right now. I think they are complicated in that they have a complicated specification (*this* kind of input produces *that* kind of output, and this *other* kind of input produces this *other* kind of output, etc. etc. etc.), and this specification what we need to work out.
* *Fewer people are working on subcortical algorithms than the neocortex's algorithms:* The neocortex is the center of human intelligence and cognition. So very exciting! So very monetizable! By contrast, the midbrain seems far less exciting and far less practically useful. Also, the neocortex is nearest the skull, and thus accessible to some experimental techniques (e.g. EEG, MEG, ECoG) that don't work on deeper structures. This is especially limiting when studying live humans, I think.
As mentioned above, I am very unhappy about this state of affairs. For the project of building safe and beneficial artificial general intelligence, I feel strongly that it would be better if we reverse-engineered subcortical algorithms first, and neocortical algorithms second.
(*Edited to add:* ...if at all. Like, maybe, armed with a better understanding of how the subcortex steers the neocortex, we'll realize that there's just *no way* to keep a brain-like AGI under human control. Then we can advocate against people continuing to pursue the research program of reverse-engineering neocortical algorithms! Or conversely, if we have a really solid plan to build safe and beneficial brain-like AGIs, we could try to *accelerate* the reverse-engineering of the neocortex, as compared to other paths to AGI. This is a great example of [how AGI-related technical safety research can be decision-relevant today even if AGI is centuries away](https://www.lesswrong.com/posts/w6AzbZR7ZQxWuAwKR/thoughts-on-robin-hanson-s-ai-impacts-interview#How_soon_are_high_leverage_decision_points_).)
Conclusion
----------
Well, my brief summary wasn't all that brief after all! Congratulations on making it this far! I'm very open to questions, discussion, and criticism. I've already revised my views on all these topics numerous times, and expect to do so again. :-) |
cfd0e493-0b62-4aaa-95c6-7064df9b80a2 | trentmkelly/LessWrong-43k | LessWrong | EA & LW Forum Weekly Summary (16th - 22nd Jan '23)
Supported by Rethink Priorities
This is part of a weekly series summarizing the top posts on the EA and LW forums - you can see the full collection here. The first post includes some details on purpose and methodology. Feedback, thoughts, and corrections are welcomed.
If you'd like to receive these summaries via email, you can subscribe here.
Podcast version: Subscribe on your favorite podcast app by searching for 'EA Forum Podcast (Summaries)'. A big thanks to Coleman Snell for producing these!
Philosophy and Methodologies
Don’t Balk at Animal-friendly Results
by Bob Fischer
The moral weight project assumes a) hedonism and b) we can use various capacities as proxies for a species’ range of possible welfare states. This post warns against knee-jerk skepticism if research leads to the conclusion that chickens and humans can realize roughly the same amount of welfare at a given time (the ‘Equality Result’). It doesn’t argue for the Equality Result itself.
Three arguments and counter-arguments for skepticism are discussed:
1. The implications are too huge - under utilitarianism, it means we should massively shift resources toward animals.
1. Utilitarianism is the reason for the radical implications - the equality result is just the messenger.
2. Maybe the hedonism assumption is wrong.
1. Fair, though per previous posts hedonism doesn’t change the bottom line much - even if hedonic goods / bads isn’t all of welfare, it’s surely a large part.
3. Even accepting the assumptions, the Equality Result feels wrong.
1. This intuition is uncalibrated, affected by many biases.
Conditional on hedonism, the Equality Result wouldn’t be surprising, as it fits with popular theories of valence.
Object Level Interventions / Reviews
Existential Risks (including AI)
How many people are working (directly) on reducing existential risk from AI?
by Benjamin Hilton, 80000_Hours
The author estimates 400 FTE (90% CI: 200-1000). Three quarters on technical |
e0838002-5e35-4407-8389-723e67f78e07 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Baltimore Area / UMBC Weekly Meetup
Discussion article for the meetup : Baltimore Area / UMBC Weekly Meetup
WHEN: 25 September 2016 08:00:00PM (-0400)
WHERE: Performing Arts and Humanities Bldg Room 456, 1000 Hilltop Cir, Baltimore, MD 21250
Meeting is on 4th floor of the Performing Arts and Humanities Building. Permit parking designations do not apply on weekends, so park pretty much wherever you want.
Discussion article for the meetup : Baltimore Area / UMBC Weekly Meetup |
7f5c43a1-6481-45c3-8a14-03741e258290 | trentmkelly/LessWrong-43k | LessWrong | Tentative Anger
Epistemic status: Exploratory. Trying to figure it out.
There is a feeling I often find myself feeling. It is an uncomfortable feeling, and the fact that I have trouble labeling it makes the discomfort even worse. I think I may have arrived at a solid label though: tentative anger.
Examples
Let me start by giving some examples of what I mean.
Mike Tomlin
Yesterday was Thanksgiving, so let's talk about football. I am a Steelers fan. However, I'm not really a fan of their head coach, Mike Tomlin. Why? Because he does stupid things.
Here's one example. We always end up in these situations where linebackers are covering slot receivers. Linebackers are generally about 250lbs or so, whereas slot receivers are more like 200 pounds. Which means those receivers are almost always faster and more agile, making it really hard for linebackers to cover them. It's like asking a boxer to fight outside of his weight class. Usually you'd have ~200lb cornerbacks guarding those receivers instead.
No one, including Tomlin, actually thinks that it is a good strategy to put linebackers on receivers. It's just that... I'm not sure actually. For whatever reason, sometimes these mismatches happen. But Tomlin doesn't seem to mind very much. Compared to other coaches, he doesn't seem to try as hard to avoid them. And when they do happen, his response is something like:
> Tough luck. The standard is the standard. You've gotta step up and meet the challenge. I don't care that he's 50 pounds lighter than you.
What makes more sense to me would be something like:
> I got outcoached. My bad. I've got to put my players in a position to succeed, and I didn't do that. I'll hit the drawing board and try to prevent this from happening in the future.
But here's the thing: he's an ultra successful football coach, and I'm not. Maybe he knows better than me. Maybe there's a Bayesians vs Barbarians angle to this.
That's plausible. But then why doesn't this happen to the other football coaches? And |
f928b952-7f26-4752-84dc-debce772a983 | trentmkelly/LessWrong-43k | LessWrong | P/S/A - Sam Harris offering money for a little good philosophy
Sam Harris is here offering a substantial amount of money to anyone who can show a flaw in the philosophy of 'The Moral Landscape' in 1000 word or less, or at least the best attempt.
http://www.samharris.org/blog/item/the-moral-landscape-challenge1
Up to $20,000 is on offer, although that's only if you change his mind. Whilst we know that this is very difficult, note how few people offer large sums of money for the privelage of being disproven.
In case anyone does win, I will remind you that this site is created and maintained by people who work at MIRI and CFAR, which rely on outside donations, and with whom I am not affiliated.
Note: Is this misplaced in Discussion? I imagine that it could be easily overlooked in an open thread by the sorts of people who would be able to use this information well? |
43c76c18-c611-427e-af68-cda741c65d56 | trentmkelly/LessWrong-43k | LessWrong | Room for more funding at the Future of Humanity Institute
In case you didn't already know: The Future of Humanity Institute, one of the three organizations co-sponsoring LW, is a group within the University of Oxford's philosophy department that tackles important, large-scale problems for humanity like how to go about reducing existential risk.
I've been casually corresponding with the FHI in an effort to learn more about the different options available for purchasing existential risk reduction. Here's a summary of what I've learned from research fellow Stuart Armstrong and academic project manager Sean O'Heigeartaigh:
* Sean reports that since this SIAI/FHI achievements comparison, FHI's full-time research team has expanded to 7, the biggest it's ever been. Sean writes: "Our output has improved dramatically by all tangible metrics (academic papers, outreach, policy impact, etc) to match this."
* Despite this, Sean writes, "we’re not nearly at the capacity we’d like to reach. There are a number of research areas in which we would very like to expand (more machine intelligence work, synthetic biology risks, surveillance/information society work) and in which we feel that we could make a major impact. There are also quite a number of talented researchers over the past year who we haven’t been able to employ but would dearly like to."
* They'd also like to do more public outreach, but standard academic funding routes aren't likely to cover this. So without funding from individuals, it's much less likely to happen.
* Sean is currently working overtime to cover a missing administrative staff member, but he plans to release a new achievement report (see sidebar on this page for past achievement reports) sometime in the next few months.
* Although the FHI has traditionally pursued standard academic funding channels, donations from individuals (small and large) are more than welcome. (Stuart says this can't be emphasized enough.)
* Stuart reports current academic funding opportunities are "a bit iffy, with some possible |
44ea75d9-0223-4ec1-9c61-5b7e443ab02d | trentmkelly/LessWrong-43k | LessWrong | Evidential Decision Theory and Mass Mind Control
Required Reading: Evidential Decision Theory
Let me begin with something similar to Newcomb's Paradox. You're not the guy choosing whether or not to take both boxes. You're the guy who predicts. You're not actually prescient. You can only make an educated guess.
You watch the first person play. Let's say they pick one box. You know they're not an ordinary person. They're a lot more philosophical than normal. But that doesn't mean that the knowledge of what they choose is completely useless later on. The later people might be just as weird. Or they might be normal, but they're not completely independent of this outlier. You can use his decision to help predict theirs, if only by a little. What's more, this still works if you're reading through archives and trying to "predict" the decisions people have already made in earlier trials.
The decision of the player choosing the box affects whether or not the predictor will predict that later, or earlier, people will take the box. According to EDT, one should act in the way that results in the most evidence for what one wants. Since the predictor is completely rational, this means that the player choosing the box effectively changes decisions other people make, or actually changes depending on your interpretation of EDT. One can even affect people's decisions in the past, provided that one doesn't know what they were.
In short, the decisions you make affect the decisions other people will make and have made. I'm not sure how much, but there have probably been 50 to 100 billion people. And that's not including the people who haven't been born yet. Even if you only change one in a thousand decisions, that's at least 50 million people.
Like I said: mass mind control. Use this power for good. |
ae05170b-2ce3-41d0-a8be-014091e98548 | trentmkelly/LessWrong-43k | LessWrong | Understanding your understanding
Related to: Truly Part of You, A Technical Explanation of Technical Explanation
Partly because of LessWrong discussions about what really counts as understanding (some typical examples), I came up with a scheme to classify different levels of understanding so that posters can be more precise about what they mean when they claim to understand -- or fail to understand -- a particular phenomenon or domain.
Each level has a description so that you know if you meet it, and tells you what to watch out for when you're at or close to that level. I have taken the liberty of naming them after the LW articles that describe what such a level is like.
Level 0: The "Guessing the Teacher's Password" Stage
Summary: You have no understanding, because you don't see how any outcome is more or less likely than any other.
Description: This level is only included for comparison -- to show something that is not understanding. At this point, you have, a best, labels that other people use when describing the phenomenon. Maybe you can even generate the appearance of understanding on the topic. However, you actually have a maximum entropy probability distribution. In other words, nothing would surprise you, no event is more or less likely to happen, and everything is consistent with what you "know" about it. No rationalist should count this as an understanding, though it may involve knowledge of the labels that a domain uses.
Things to watch out for: Scientific-sounding terms in your vocabulary that don't correspond to an actual predictive model; your inability to say what you expect to see, and what you would be surprised by.
Level 1: The "Shut up and Calculate" Stage
Summary: You can successfully predict the phenomenon, but see it as an independent, compartmentalized domain.
Description: This is where you can predict the phenomenon, using a generative model that tells you what to expect. You are capable of being surprised, as certain observations are assigned low probability. It ma |
06964be9-0905-49e0-b8eb-63aaddaedacb | trentmkelly/LessWrong-43k | LessWrong | How do you feel about LessWrong these days? [Open feedback thread]
Hello! This is jacobjacob from the LessWrong / Lightcone team.
This is a meta thread for you to share any thoughts, feelings, feedback or other stuff about LessWrong, that's been on your mind.
Examples of things you might share:
* "I really like agree/disagree voting!"
* "What's up with all this Dialogues stuff? It's confusing...
* "Hm... it seems like recently the vibe on the site has changed somehow... in particular [insert 10 paragraphs]"
...or anything else!
The point of this thread is to give you an affordance to share anything that's been on your mind, in a place where you know that a team member will be listening.
(We're a small team and have to prioritise what we work on, so I of course don't promise to action everything mentioned here. But I will at least listen to all of it!)
I haven't seen any public threads like this for a while. Maybe there's a lot of boiling feelings out there about the site that never get voiced? Or maybe y'all don't have more to share than what I find out from just reading normal comments, posts, metrics, and Intercom comments? Well, here's one way to find out! I'm really curious to ask and see how people feel about the site.
So, how do you feel about LessWrong these days? Feel free to leave your answers below. |
add9f90e-440b-488c-9735-39b29c7e8a92 | trentmkelly/LessWrong-43k | LessWrong | Tips On Empirical Research Slides
Our research is centered on empirical research with LLMs. So if you are doing something similar, these tips on slide-based communication may be helpful!
Background:
James Chua and John Hughes are researchers working under Owain Evans and Ethan Perez, respectively. Both of us (James and John) used to be MATS mentees. We weren't good at making research slides at first -- here are some principles we've found useful for understandable slides for our weekly research meetings.
We show some good example slides. We also show examples of confusing slides we've made — marked in the caption with “❌ Negative example”.
Below we use the study of sycophancy as an example. Sycophancy occurs when a model has responses that match user beliefs rather than truthful ones.
Summary slide sets the frame
Possible summary slide. I summarise my experimental results for the week and write what I want to be discussed. If possible, I fit in a simple plot on the right.
Your mentor manages multiple projects and people. They need to be reminded of what you are up to. The first slide should recap the key takeaways from the last meeting to motivate what you have worked on and provide a clear summary of your progress.
There are two main messages to convey, which set the frame for what your mentor should think about:
* Key takeaways from the last meeting. A reminder of the next steps and what was discussed.
* Experiment outcome
* My experiments worked! Your mentor will focus on sanity checks, control experiments, and extensions to other setups.
* My experiments didn't work. Your mentor will focus on debugging why. E.g., train with better data, or improve the prompts.
A summary helps your mentor save time. For example, in the slide, your mentor may say “Oh I can remember what the data augmentation was, let’s skip that”. Or maybe your mentor already read your results and wants to discuss something else.
Include an agenda
Often in meetings with your mentor, there is very little |
01f0a4f9-c6e5-49c4-a63a-525e1b39e06a | trentmkelly/LessWrong-43k | LessWrong | Spirits vs Terms as Arbitration Norms
Here's a cautionary tale about a friendly bet going wrong and how to avoid the same mistake in your own bets.
When I was in Brazil with a friend, I had a grand time teaching him to say peanut butter in Portuguese (manteiga de amendoim). He struggled to remember it and butchered the pronunciation and so it turned the whole situation into a great inside joke.
Fast forward a year later,we're staying with each other again and I bet him that he wouldn't be able to remember how to say it by midnight at 20:1 odds. The bet was on.
As we went about our day, every now and then we'd think of it, he’d struggle, but say it was starting to come back to him.
Around midnight I called him into my room and asked him for his final answer.
"Pasta de amendoim." he said.
This kicked off a heated debate about who was right. We each were shocked at some of the other's arguments. They seemed so obviously wrong. We're great friends so even though feelings were heightened, we could still laugh about it.
I held that it wasn't the right answer because I was looking for "peanut butter" while he said "peanut paste". Peanut Paste is what is written on some peanut butter jars but I argued that that clearly wasn't the intent of the bet because there was no inside joke to that effect. Why would I make a bet about random words distinct from the ones I was attempting to teach him day in, day out. He argued that the exact phrasing of the bet was "How do you say peanut butter in Portuguese?" (which is true) and that a third-party would surely rule in his favor. I said that if the third-party had more than just a dictionary, if they had the entire context of our Brazil trip, they would rule in my favor. He still disagreed because he said the context didn't matter because of the way I worded the bet.
I left the situation feeling bad about all the potential bets we might not make if this became the most salient betting experience we had.
Weeks later we revisited it and we finally saw what each othe |
f3fcfac6-665b-4f5b-a43f-9cc330c7c6f7 | StampyAI/alignment-research-dataset/blogs | Blogs | Scale of the Human Brain
The brain has about 10¹¹ neurons and 1.8-3.2 x 10¹⁴ synapses. These probably account for the majority of computationally interesting behavior.
Support
-------
#### Number of neurons in the brain
The number of neurons in the brain is about 10¹¹. For instance, [Azevado et al](http://www.ncbi.nlm.nih.gov/pubmed/19226510) physically counted them and found 0.6-1 \* 10¹¹. [Eric Chudler](http://faculty.washington.edu/chudler/facts.html) has collected estimates from a range of textbooks, which estimate 1-2 x 10¹⁰ of these (10%-30%) are in the cerebral cortex.[1](https://aiimpacts.org/scale-of-the-human-brain/#easy-footnote-bottom-1-143 "Total number of neurons in cerebral cortex = 10 billion (from G.M. Shepherd, <i>The Synaptic Organization of the Brain</i>, 1998, p. 6). However, C. Koch lists the total number of neurons in the cerebral cortex at 20 billion (<i>Biophysics of Computation. Information Processing in Single Neurons</i>, New York: Oxford Univ. Press, 1999, page 87). ")
#### Number of synapses in the brain
The number of synapses in the brain is known much less precisely, but is probably about 10¹⁴. For instance [Human-memory.net](http://www.human-memory.net/brain_neurons.html) reports 10¹⁴-10¹⁵ (100 – 1000 trillion) synapses in the brain, with no citation or explanation. Wikipedia says the brain contains 100 billion neurons, with 7,000 synaptic connections each, for 7 x 10¹⁴ synapses in total, but this seems possibly in error.[2](https://aiimpacts.org/scale-of-the-human-brain/#easy-footnote-bottom-2-143 "“The human brain has a huge number of synapses. Each of the 10<sup>11</sup> (one hundred billion) neurons has on average 7,000 synaptic connections to other neurons. It has been estimated that the brain of a three-year-old child has about 10¹⁵ synapses (1 quadrillion). This number declines with age, stabilizing by adulthood. Estimates vary for an adult, ranging from 10¹⁴ to 5 x 10¹⁴ synapses (100 to 500 trillion).” <a href=\"http://en.wikipedia.org/wiki/Neuron#Connectivity\">Wikipedia</a> accessed April 13 ’15, citing “<a href=\"http://www.neurology.org/content/64/12/2004.extract\">Do we have brain to spare?</a>“. <i>Neurology</i> <b>64</b> (12): 2004–5. We have not accessed most of the Drachman paper, but it does at least say “Within the liter and a half of human brain, stereologic studies estimate that there are approximately 20 billion neocortical neurons, with an average of 7,000 synaptic connections each”. This suggests that the Wikipedia page errs in attributing the 7,000 synaptic connections per neuron to the brain at large instead of the neocortex.")
##### Number of synapses in the neocortex
One way to estimate of the number of synapses in the brain is to extrapolate from the number in the neocortex. According to stereologic studies that we have not investigated, there are around 1.4 x 10¹⁴ synapses in the neocortex.[3](https://aiimpacts.org/scale-of-the-human-brain/#easy-footnote-bottom-3-143 "“Within the liter and a half of human brain, stereologic studies estimate that there are approximately 20 billion neocortical neurons, with an average of 7,000 synaptic connections each”.”<a href=\"http://www.neurology.org/content/64/12/2004.extract\">Do we have brain to spare?</a>“. <i>Neurology</i> <b>64</b> (12): 2004–5. ") This is roughly consistent with [Eric Chudler’s summary of textbooks](http://faculty.washington.edu/chudler/facts.html), which gives estimates between 0.6-2.4 x 10¹⁴ for the number of synapses in the cerebral cortex.[4](https://aiimpacts.org/scale-of-the-human-brain/#easy-footnote-bottom-4-143 "“Number of synapses in cortex = 0.15 quadrillion (Pakkenberg et al., 1997; 2003)… [the ‘cortex’ probably refers either the cerebral cortex or the neocortex, which is part of and thus should be smaller than the cerebral cortex.]</p>
<p>…Total number of synapses in cerebral cortex = 60 trillion (yes, trillion) (from G.M. Shepherd, <i>The Synaptic Organization of the Brain</i>, 1998, p. 6). However, C. Koch lists the total synapses in the cerebral cortex at 240 trillion (<i>Biophysics of Computation. Information Processing in Single Neurons</i>, New York: Oxford Univ. Press, 1999, page 87).” – <a href=\"http://faculty.washington.edu/chudler/facts.html\" rel=\"nofollow\">Chudler</a>, Brain facts and figures")
We are not aware of convincing estimates for synaptic density outside of the cerebral cortex, and our impression is that widely reported estimates of 10¹⁴ are derived from the assumption that the neocortex contains the great bulk of synapses in the brain. This seems plausible given the large volume of the neocortex, despite the fact that it contains a minority of the brain’s neurons. By volume, around 80% of the human brain is neocortex.[5](https://aiimpacts.org/scale-of-the-human-brain/#easy-footnote-bottom-5-143 "Dunbar references <a href=\"http://www.ncbi.nlm.nih.gov/pubmed/7014398\">anatomical measurements from 1981</a> and <a href=\"http://www.cogsci.ucsd.edu/~johnson/COGS184/3Dunbar93.pdf\">writes</a> “With a neocortical volume of 1006.5 cc and a total brain volume of 1251.8 cc (Stephan et al. 1981), the neocortex ratio for humans is CR = 4.1.” (p. 682).") The neocortex also consumes around 44% of the brain’s total energy, which may be another reasonable indicator of the fraction of synapses in contains.[6](https://aiimpacts.org/scale-of-the-human-brain/#easy-footnote-bottom-6-143 "“Thus, neocortex accounts for 44% of the brain’s overall consumption.” <a href=\"http://www.bcs.rochester.edu/people/plennie/pdfs/Lennie03a.pdf\">Lennie, 2003</a> (p. 495)") So our guess is that the number of synapses in the entire brain is somewhere between 1.3 and 2.3 times the number in the cerebral cortex. From above, the cerebral cortex contains around 1.4 x 10¹⁴ synapses, so this gives us 1.8-3.2 x 10¹⁴ total synapses.
##### Number of synapses per neuron
The number of synapses per neuron varies considerably. [According to Wikipedia](http://en.wikipedia.org/wiki/Cerebellum_granule_cell#cite_note-SOB-1), the majority of neurons are cerebellum granule cells, which have only a handful of synapses, while the statistics above suggest that the average neuron has around 1,000 synapses. Purkinje cells have up to 200,000 synapses.[7](https://aiimpacts.org/scale-of-the-human-brain/#easy-footnote-bottom-7-143 "“Number of synapses made on a Purkinje cell = up to 200,000” – <a href=\"http://faculty.washington.edu/chudler/facts.html\" rel=\"nofollow\">Chudler</a>, Brain facts and figures")
#### Number of glial cells in the brain
*Main article: [Glial signaling](http://aiimpacts.org/glial-signaling/ "Glial Signaling")*
[Azevado et al](http://www.ncbi.nlm.nih.gov/pubmed/19226510) aforementioned investigation finds about 10¹¹ glial cells (the same as the number of neurons).
#### Relevance of cells other than neurons to computations in the brain
*Main article: [Glial signaling](http://aiimpacts.org/glial-signaling/ "Glial Signaling")*
It seems that the timescales of glial dynamics are substantially longer than for neuron dynamics. [Sandberg and Bostrom](http://www.fhi.ox.ac.uk/brain-emulation-roadmap-report.pdf) write: “However, the time constants for glial calcium dynamics is generally far slower than the dynamics of action potentials (on the order of seconds or more), suggesting that the time resolution would not have to be as fine” (p. 36). This suggests that the computational role of glial cells is not too great. References to much larger numbers of glial cells appear to be common, but we were unable to track down any empirical research supporting these claims. An [informal blog post](http://neurocritic.blogspot.com/2009/09/fact-or-fiction-there-ten-times-more.html) suggests that a common claim that there are ten times as many glial cells as neurons may be a popular myth.
We are not aware of serious suggestions that cells other than neurons or glia play a computationally significant role in the functioning of the brain.
--- |
2825fd64-850d-4bef-8821-a9695eb14320 | trentmkelly/LessWrong-43k | LessWrong | Reinforcing Habits
Epistemic status: hypothesis based on >10 anecdotal examples
Every month or so, I’ll get a client asking why their attempts to start a habit failed. They want to have an automatic action requiring minimal willpower. The client is usually familiar with at least one habit-building model. Most commonly Charles Duhig’s Cue, Routine, Reward loop (in “The Power of Habit”) or CFAR’s Trigger-Action Plans (TAPs).
Their model may go like Duhig’s story: he wanted to change his habit of eating a cookie each afternoon (motivated by watching the scale creep up). So he identified his cue (the time of day around 3pm), planned a new routine to replace the old one (talk to colleagues for ten minutes), and had a new habit. My client usually wants to know what thwarts their attempts to do likewise.
What’s missing from that example?
I read The Power of Habit many years ago, and honestly didn’t remember Duhig’s cookie story any better than my clients often do. So I was surprised when I revisited it and found Duhig actually did a series of experiments to find what the reward was. Something sweet? Nope, eating a candy bar at his desk didn’t feel great. Just taking a break? Nope, taking a walk outside didn’t cut it. Talking with friends? Yeah, that felt rewarding. Based on his experiments, we can guess that getting a cookie each afternoon was a means for getting him to talk to his colleagues while eating it. Because the reward here was socializing, he could build a new habit that didn’t use the cookie as an intermediate step.
This reward step is often neglected by my habit-struggling clients. They want the low-effort, automatic aspects of habits. They lack, however, anything to make those behaviors sticky.
I think it might help to reframe habits as repeatedly-reinforced behaviors. Our brains, often subconsciously, have tied a particular action to some cue after repeatedly having that action rewarded. Simple patterns of cue, action, and reward in close proximity get reinforced, suc |
330b4e22-a081-41fa-9474-c75a7c65bc41 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | AXRP Episode 3 - Negotiable Reinforcement Learning with Andrew Critch
[Google Podcasts link](https://podcasts.google.com/feed/aHR0cHM6Ly9heHJwb2RjYXN0LmxpYnN5bi5jb20vcnNz/episode/ZmFmZmFkMTctZmJhZC00ODRhLThhZGUtNjk1NGU1ZWI2NDJi)
This podcast is called AXRP, pronounced axe-urp and short for the AI X-risk Research Podcast. Here, I ([Daniel Filan](https://danielfilan.com/)) have conversations with researchers about their papers. We discuss the paper and hopefully get a sense of why it's been written and how it might reduce the risk of artificial intelligence causing an [existential catastrophe](https://en.wikipedia.org/wiki/Global_catastrophic_risk): that is, permanently and drastically curtailing humanity's future potential.
In this episode, I talk with Andrew Critch about negotiable reinforcement learning: what happens when two people (or organizations, or what have you) who have different beliefs and preferences jointly build some agent that will take actions in the real world. In the paper we discuss, it's proven that the only way to make such an agent Pareto optimal - that is, have it not be the case that there's a different agent that both people would prefer to use instead - is to have it preferentially optimize the preferences of whoever's beliefs were more accurate. We discuss his motivations for working on the problem and what he thinks about it.
**Daniel Filan:**
Hello, everybody. Today, we're going to be talking to Andrew Critch. Andrew Critch got his Ph.D. in algebraic geometry at UC Berkeley. He's worked at Jane Street as an algorithmic trader, the Machine Intelligence Research Institute as a researcher. And he also co-founded the Center for Applied Rationality, where he was a curriculum designer. But currently he's a research scientist at UC Berkeley's Center for Human Compatible A.I. Today, the paper we're going to be talking about is [Negotiable Reinforcement Learning for Pareto Optimal Sequential Decision-Making](https://papers.nips.cc/paper/2018/hash/5b8e4fd39d9786228649a8a8bec4e008-Abstract.html). The authors are Nishant Desai, Andrew Critch and Stuart Russell. Hello Andrew.
**Andrew Critch:**
Hi. Nice to be here.
**Daniel Filan:**
Nice to have you here. So I guess my first question about this paper is, what problem is it solving?
**Andrew Critch:**
Right. So when I quote, unquote solve a research problem, I'm usually trying to do two different things. One is that I'm answering an actual formally specifiable math question or maybe computer science question. The other thing is, I'm trying to draw attention to an area. So for me, the purpose of this paper was to draw attention to an area that I felt neglected. And so that's a meta research problem that it's trying to solve. And then the research problem that it's solving, or the object level problem that it solves is it demonstrates and explicates what a pareto optimal sequential decision making procedure must look like when the people that it's making decisions for have different beliefs.
**Daniel Filan:**
All right, cool. So one thing that I guess I'm interested in, an area that I'm interested in that jumped out to me in this, is the disagreements between the two, I guess, principals that this sequential decision making policy is serving. And I'm interested in this because there's a line of work, starting with Aumann that basically says that persistent disagreements between two people are irrational, as long as people can communicate, they shouldn't disagree about anything, really. So there's Aumann's Agreement Theorem. This got followed up by Are Disagreements Honest by Tyler Cowen and Robin Hanson and the paper, Uncommon Priors Require Origin Disputes by Robin Hanson. I'm wondering what you think about this line of work and just in general, the set up of two people disagreeing, isn't that crazy?
**Andrew Critch:**
That's great. That's a great question. So a lot of things you just said really triggered me, "Oh, no, it's irrational to disagree." So, first of all, I think Aumann's line of work is really important and things that build on it are a good area of inquiry. There's just not that many people who think about how beliefs work in a inter subjective, formalized setting. But there's a number of problems with trying to assert that disagreement is irrational at the individual level. And I say at the individual level, because rationality is a descriptor of a system and you can have a system where each individual is rational but the system as a whole is not in some sense, in many different senses.
**Andrew Critch:**
So the first thing is that Aumann's Agreement Theorem applies when the two parties have a common prior. Which seems extremely unrealistic for the real world. And I'll say more of what that means, but they're required to have a common prior and also common knowledge of the state of disagreement or state of their beliefs, which I think is also very unrealistic for the real world, and I don't mean unrealistic in some kind of, you can only get .99 when the theorem requires a 1 on some agreement metric. I mean, you can only get .3 when the theorem requires a 1 on some kind of agreement metric is what I'm saying. So I think the assumptions of Aumann are robustly wrong as a descriptor of the real world.
**Andrew Critch:**
But they're an important technical starting point, you describe a scenario under simple assumptions and Aumann's assumptions are simplifying. And then we have to keep complexifying those assumptions to get a real understanding of how beliefs between agents work.
**Daniel Filan:**
Why don't you think that the common knowledge of... So I've heard a lot of people say that the common priors assumption is not realistic for humans. Why do you think the common knowledge of disagreement is not realistic among humans? I think often I have disputes with people where we know that we have this dispute. Right?
**Andrew Critch:**
So right. So first, I want to address the priors thing, even though other people have addressed it, you know. Common prior, I mean, what is my prior? Is it something I was born with as a baby? Is it something in my DNA? There's many different ways of conceiving of a human as having had a prior and then some updates and I think in any of the reasonable conceptions of a human as a Bayesian updating agent, the prior is a pretty old thing that they've had for a long time. And it comes from before they've had a chance to interact with a lot of other people. So I don't think people have equal priors. They're genetically different. They're culturally different. And even as adults, we maybe have only interacted with our own culture. And I think that's deeply bubbling for people, if I can use the word bubble as a verb, it sequesters people and people, even as an educated adult, you haven't interacted with educated adults from other cultures who've read a lot and seen a lot. And I say educated because that's how you get information. If you have less education, you have a different prior as well.
**Andrew Critch:**
So I think it's a big deal, and if we're going to start talking about how A.I is going to benefit humanity, we need to be thinking about people having different beliefs about whether it's beneficial and what is beneficial. And then separately, the common knowledge of disagreement thing, first of all, I would call into question your experience that you really have had full common knowledge of disagreement with the person. You know, there's always this uncertainty, how do you know you were using words in the same way as them? You talk to them for a while and you gain some evidence that you're using words in the same way, and if you're a careful thinker and you engage carefully in discourse, you check to see if you're using words in the same way. But you only ran that check for so long. And what if you're using concepts in a different way too?
**Andrew Critch:**
Let's say you had a debate with somebody about whether people should have privacy from A.I. systems. You talked for a while about what privacy means, you talked for a while about what should means and then you grounded out to some kind of empirical prediction, "I think if the people don't have this kind of privacy, they will end up distressed in the following way". And then the other person in the argument says, "I predict they will not end up distressed", but now, you're satisfied you've made progress and it's good to be satisfied that that progress was made, but have you grounded out what distressed means? Eventually you just go home, eventually you've done a good job today. You made progress with this interlocutor and you disagree still and you don't know for sure whether the concepts you're using are the same as the concepts they're using.
**Andrew Critch:**
And I think that that's profoundly important, if you didn't settle what you meant by distress, that can be an important difference in culture, for example, where maybe for you, something that just makes you a little bit sweaty and makes your mind go faster, counts as distress, whereas for the other person it doesn't. And now you have to ground that out as well. And in my experience, whenever there appears to be a persistent disagreement, if you talk longer, you can always uncover some kind of confusion or miscommunication or difference in information such that prior to that uncovering, you were both deluded as to the nature of the disagreement.
**Andrew Critch:**
So that's why I call into question this idea that you really have common knowledge of the disagreement, because I think you probably are both deluded as to the nature of it, when you still disagree.
**Daniel Filan:**
That's interesting, that seems plausible to me. Now that you talked about the common priors assumption, I actually want to talk about that a little bit. I hope our listeners are interested in the philosophy of Bayesian disagreement. Yeah, so when I think of the priors for a person and how to make sense of them in terms of Bayesian reasoning or whatever. To me it seems this involves some amount of epistemological knowledge that you can update. For instance, I think it's possible that at some point in my life, I didn't know about the modus ponens inference rule or something or maybe some other inference rules. And now I do.
**Daniel Filan:**
And that those have become part of my prior. Where what I mean by the word 'prior' is, okay, today, how do I form beliefs based on everything I know in the past? And maybe that might be a little bit different tomorrow because I'll have some different conception about simplicity priors or something. So in this case, your prior - it sounds strange to say this, but I do want to say that I think you should be able to change your prior over time. And if you can do this then, okay, if somebody comes from a different culture where they understand things differently, hopefully you can reason it out with them by some means other than Bayesian updating, hopefully otherwise it's layers upon layers.
**Daniel Filan:**
Yeah, I just wanted to address what priors should mean in this kind of context, because I think under this different conception, it becomes maybe a bit more realistic.
**Andrew Critch:**
Cool, so before we get into that, I do want to say why I care about this. And the reason is that I'm hoping that we can make progress in A.I. that makes it easier for people with diverse backgrounds and beliefs, not just diverse preferences, but diverse beliefs to share control of the systems. And the reason I want us to be able to share control of systems is twofold. One, I think it's just fair. If you have very powerful systems and powerful technologies, it's more fair to share it.
**Andrew Critch:**
And the other one is that if you can share things, you don't have to fight over them so that decreases the likelihood of conflict over powerful artifacts of technology in the future. And I think there's quite a lot of societal and potentially existential risk that comes from that. So that's the source of my interest here. I think there's many other reasons to be interested in making A.I. compatible with diverse human beliefs and making it possible to negotiate for the control of the system, even knowing that.
**Andrew Critch:**
But I just want to flag that, while I'm going down this rabbit hole with you, that's what's steering me. And if I say something that seems important to that, I say something that seems important to your question, I'm also filtering it for importance to, is this going to matter to the future governance of technology?
**Daniel Filan:**
Okay.
**Andrew Critch:**
So the first thing is that, yes, I agree with you that people can change their priors. But I mean, a Bayesian agent, quote, unquote, changes prior when it updates. And I think you mean something more nuanced than that, which is- [crosstalk 00:13:22]
**Daniel Filan:**
Yeah. I mean- [crosstalk 00:13:23]
**Andrew Critch:**
You can think longer and decide that your prior at the beginning of time ought to have been something else. And so, yeah, my statement that you're responding to in that is that a reasonable conception of a human being as a Bayesian agent has the prior as something that has existed for a long time or that came into existence a long time ago. And that reasonable conception of a human Bayesian is doing a lot of work because humans aren't Bayesian agents. In fact, physical agents are not Bayesian agents because you have to do a lot of computation to be a Bayesian agent. So in fact, you have to do infinite computation. So this is where my interest in logical uncertainty and if you've heard of logical induction comes from. And so I would just argue that when you're changing your prior there, you're changing which Bayesian agent you are. You're not being a Bayesian agent in that moment.
**Daniel Filan:**
Okay. That seems fair. All right, so- [crosstalk 00:14:26]
**Andrew Critch:**
And for listeners who've never thought about that, why is that important? Well, I think there's ethical questions that can be resolved by thinking, there's ethical questions that can't. I think Rawl's Veil of Ignorance is an example of an ethical principle that helps you figure things out by thinking longer and harder about what if I were somebody else, you already knew everything you needed to know about those people to start realizing some of the things you should do to be fair. But you have to think about it. And in the same way I think that's going to apply in the governance of A.I. and for that reason, I think it is going to be important not to treat people like Bayesians because people are entities and computers are entities that change what they believe merely by thinking, even without making further observations. So that's a major shortcoming of the negotiable reinforcement learning framework that's only alluded to in the older arXiv draft with just me on it that says naturalized agency is going to be key future work. And I do not think that the paper addresses that well at all.
**Daniel Filan:**
Yeah, I think that's a good point. I'm going to tack a little bit back to the paper or the literature on Bayesian agents cooperating and such. The related work section of this paper has a lot of interesting stuff on social choice theory and such, there's a rich work of literature, I guess, both on social choice theory and on, can reasoners disagree? In some sense, a reader might find it a little bit surprising that this work hasn't already been done, at least that the main theorem in this paper hasn't already been proven. I was wondering if you have thoughts about why it took until, was it 2018 that this got published? 2019?
**Andrew Critch:**
2017 was the first, the NeurIPS version was 2018 but the theorem you're referencing was proven in 2017.
**Daniel Filan:**
Okay, so why do you think it took until 2017?
**Andrew Critch:**
Yeah. This is something I grapple with deeply, I mean, for me, "how do agents with different beliefs get along" is a pretty basic question. So it has been analyzed a little bit, like you said, by Aumann and people in Aumann's - Google scholar search people who cite Aumann and you'll get a lot of interesting thoughts about that. But not really much looking at sequential decision making. So you get these really static analyses, imagine you're at the end of eternity and you've reached common knowledge of disagreement, and at the beginning of eternity you had a common prior. Now Aumann's theorem applies, but it's a fixed moment. It's not something evolving over time. And things evolving over time are more complicated than things that are static. So if I were going to guess, it's just you've got people who work on sequential decision making, which is reinforcement learning people and operations research people. And then you've got people who think about beliefs a lot and "what is a belief" and there hasn't been that crossover of "okay, what happens when you put the sequential decision making and the belief disagreement together?"
**Daniel Filan:**
Yeah, I guess that's related to how in statistical mechanics it's much easier to come up with a theory of equilibrium statistical mechanics, than non-equilibrium statistical mechanics. And it took humanity, we got the equilibrium theory way before we got a good non equilibrium theory.
**Andrew Critch:**
And I think there's a lot of things like that in analysis of multi-agent interactions, game theory in general, is just all about equilibria, not about how you get there. There's some research on that, but I think it's going to be a lot of hard work still to figure out.
**Daniel Filan:**
Do you have examples of these non equilibrium problems that maybe our listeners can help solve?
**Andrew Critch:**
Oh, well, I mean, there's a lot of games where finding the Nash equilibria is NP-hard. So that means in particular, that the two agents playing against each other, if you take those two agents as a computation, they're not going to be finding a Nash equilibrium unless they've got enough compute to solve an NP-hard problem, which they probably don't. So just Google NP-Hard Nash equilibria, and then you'll just see how many Nash equilibria just aren't really going to happen.
**Daniel Filan:**
Okay. So. I guess with that out of the way, we'll get into the details of the paper. So we're talking about different principals who somehow have to negotiate over a policy that's going to act in the world. And one theorem that's a little bit about this is the Harsanyi Utilitarianism Theorem, right, where there's you and me and perhaps we're electing a government or something and Harsanyi's Utilitarian Theorem basically says, well, what the government should optimize is a weighted linear combination of our utility functions. And in your paper you prove something that isn't that theorem, could you tell us a little bit more about why that theorem doesn't apply? Or why can't you just use that result?
**Andrew Critch:**
Yeah, so I mean, answering that kind of just is the theorem, but I can try to give an intuitive version of it. So first of all, the theorem is pretty easy, it's just linear algebra. I don't think it's a deep fact. I think the only thing special about it is bothering to think about it.
**Daniel Filan:**
You also need to know a bit about convex geometry, a tiny bit.
**Andrew Critch:**
I guess. Yeah a little bit. But it's really if you just draw a picture, it's kind of clear so. So first of all, let's talk about what Harsanyi's theorem says. Harsanyi's theorem is a brilliant theorem. It really simplifies the number of different ways you could imagine aggregating people's preferences by showing that basically many, many different reasonable ways of doing it are all equivalent to just giving a linear weight to each person's preference and then maximizing that sum. So that's cool and it's a little counterintuitive to me in the sense that it feels intuitively, or it used to feel intuitively to me, like there ought to be more different ways of aggregating preferences that feel compelling, but that aren't linear combinations and a lot of things that felt different from a linear combination to me just turned out to be doing linear combinations. So that was kind of cool, and I'm not sure I can even remember what they were now because my brain has compressed them into the linear combination bucket.
**Andrew Critch:**
But this key assumption of having the same beliefs is a key assumption of Harysanyi's theorem. And it's not even explicitly stated, it's just 'fact' is lurking in the background and it's assumed that everybody has access to the facts. But in reality, we don't have access to 'fact'. We have beliefs and we have things to do to update our beliefs, get better information.
**Andrew Critch:**
So here is, by the way, still putting off on answering your question, I'm going to say that the paper is not normative, you said, you can take Haysanyi's theorem as being normative and maybe he intended it to be normative, but I use it not normatively, but just descriptively. Look, all these things you might do. They're all just linear combinations of preferences. That's a nice, simplifying fact.
**Andrew Critch:**
In the same way, I don't take the negotiable reinforcement learning result or the toward negotiable reinforcement learning result as normative, because actually I think there's a lot of bad outcomes that result from the dynamics described in the theorem. It's simultaneously a negative result. I don't think that's how you should do things so I'm going to answer a slightly different question, which is why would you do it that way? If you were accounting for differences in beliefs as described in the paper, why would you be doing it?
**Andrew Critch:**
And the reason is this, so let's just say you and I are deciding to, I don't know, let's say we're deciding to do a podcast together and you're going to interview me in a podcast. That's a negotiation. You know, we got to decide on the time. We got to decide am I comfortable with the recording tools you are using or whatever, all that kind of stuff. And then once that's decided, then we go ahead and we do the thing. We execute the sequential decision making that is a podcast interview together. But before we do that there's always the possibility that the negotiation could just fail, it could fall through and we go back to what people call the best alternative to negotiated agreement or BATNA.
**Andrew Critch:**
So my BATNA today, if I didn't do this podcast, was going to be to write some things up, I was going to do some writing and I don't know what your BATNA was, but if this failed, maybe you would have just interviewed somebody else today. So if you want to maximize the probability that two people are going to choose to cooperate and execute a literally co-operative sequence of decisions, you want them to be able to find a plan that they both like more than their BATNA. And so we both liked this idea of the podcast today more than the other stuff we were going to do. So now we're doing it and if there's any what people call Pareto sub-optimality, meaning opportunities to improve the plan for you without making it worse for me, if there's any Pareto sub-optimality on the table, then there's a chance that we're below your BATNA needlessly. If we're in the midst of making a plan and we're crappy at planning together, we're bad at negotiations such that the plan we have is Pareto sub-optimal, meaning we could make it better for you without making it worse for me, or we could make it better for me without making it worse for you. That's Pareto sub-optimality.
**Andrew Critch:**
If it's Pareto sub-optimal, there's a risk that the plan is going to be below your BATNA and you're going to bail and it's below your BATNA needlessly. It's like we should just bump it up. And that way, if you treat your BATNA as a random variable, if I treat your BATNA or a mediator were to treat both of our BATNAs as random variables, there's a chance that those random variables are going to be below the best negotiated plan we have. So for me, Pareto optimization is related to or subservient to maximizing the probability that the negotiators will succeed in coming up with a sufficiently appealing cooperative plan that they choose to cooperate. And that's because I think A.I governance is going to require people to cooperate a lot.
**Andrew Critch:**
And negotiate a lot in the course of that. And so, now if you want to maximize entirely for cooperation and not for other important principles like, say, fairness, one thing you might inadvertently or you might intentionally do this or you might inadvertently do this, you might exploit people's differences in beliefs to implicitly have them bet against each other with the policy. So let's say I guess you're using Zencastr, so let's say I don't know much about Zencastr, but I think it's going to be fine because most software companies are reasonably careful with their data and you know Zencastr and you know all about them. And you happen to know that Zencastr is a terrible company that doesn't respect anybody's privacy.
**Andrew Critch:**
But you know that I don't know that. People have studied this dynamic, by the way, bargaining with asymmetric information. That's not new, but if we Pareto optimize, we end up with this plan to show up on Zencastr. And if you want me to do the podcast and I want to do it subject to privacy constraints and I don't know about them, I sign up to do it and then later I find out, oh, no my privacy is being violated by Zencastr and someone downloaded the data and recorded my neighbor's conversations from tiny trace audio. And now their privacy has been violated, too, and I've been penalized for having incorrect beliefs about how Zencastr was going to turn out for me.
**Andrew Critch:**
And the single shot version of that is just making a bet. It's like we made a bet, it's kind of two bets at once. You bet that you would like the plan, I bet that I would like the plan. And I lost my part of that bet because Zencastr turned out bad.
**Andrew Critch:**
So the interesting thing is what happens when that bet suddenly becomes a continuous process that happens for the rest of forever, which is what you see in a A.I. System, that is Pareto ex ante, meaning before it runs, ex ante subjectively Pareto optimal to the people or the principals it's serving is that it will actually every time step for the rest of eternity, settles a little bet between the principals who created it or who agreed to defer to it. And then if one of the principals had very inaccurate beliefs about what the A.I. system was going to observe, that principal's priority, the weight that it gets in the system's judgment goes down and down and down because every second it's losing a bet for how much control that principal is going to have over the A.I., or how much the A.I. Is going to choose to serve that principal's values.
**Andrew Critch:**
And so in the same way I could lose one bet with you over how good Zencastr is going to be, I could actually lose a whole series of bets with you every second about how Zencastr is going to turn out. And if you got really accurate beliefs about the world, you're going to win all those bets and our cooperation is going to end up great for you and worse for me.
**Andrew Critch:**
But it was my willingness to bet on my own false beliefs that caused me to cooperate with you in the first place. And if I had known, if I hadn't been deluded, as to Zencastr's ethics, I might have just not done the podcast. And maybe that's ethically the right thing to turn out to do. But with A.I., I worry that fragmentation could be quite bad if it leads to war, or even just standards - there's physical wars, and then there are standards wars where companies are just fighting over what standards are going to be important because they're fragmenting. And I think that can cause a lot of chaos and waste a lot of attention. It could even, if it's physical wars, actually get people killed. If countries are fighting over A.I. technology in the way that you might see companies fighting over oil as a resource. So I guess I want to point out an important trade off.
**Andrew Critch:**
This paper points out a trade off between fairness and cooperation, which is that cooperation ex ante rewards people with more accurate beliefs upon entering into the cooperation. And it's unfair to the people who had wrong misconceptions of what was going on. So your original question is why should we use this? Well in cooperation- [crosstalk 00:32:11]
**Daniel Filan:**
The original question, I think, was why doesn't the Harsanyi Aggregation Theorem apply?
**Andrew Critch:**
Right. And why or why should we use the belief updating rule instead or something?
**Daniel Filan:**
Yeah, something like that.
**Andrew Critch:**
And it's more like, well we could say in a meta problem of fairness and cooperation being two different principals you want to serve, you're trying to invent a negotiation framework that's pretty good for cooperation and pretty good for fairness. I would say Harsanyi's approach is Pareto sub-optimal because you can get more cooperation. But you should also be adding fairness as a constraint to Harsanyi. So I don't know the answer yet of what I would personally, subscribe to as the right way. I'm not so sure that no one will ever find a way that I'll look at and say, that's actually the right way, let's do it. I'm not so anti-realist about that moral judgment, but I don't have a strong view on it right now.
**Daniel Filan:**
Okay, so there are a few questions I could ask from there. First of all, I'll ask a quick technical question, in this theorem, you assume that policies can be stochastic. Can you say a little bit about what exactly you mean by that assumption? Because I think it's slightly different than what readers might think.
**Andrew Critch:**
Oh, I just mean that at every time step, you can randomize what you're going to do. So the A.I. system every time step is like flipping a coin and it's policy is just what the weight of that coin is. And it can also randomize at the outset if it wants, it can choose a random seed at the beginning to choose between two different random policies. So it has a memory in a sense. There's a few different ways of formalizing it. One is it generates a random seed at the beginning of time and then remembers that seed for the rest of time. Or you could just have that it can remember everything that it has previously done, including that initial coin flip. And that's the formalization that I adopt just because it proves a stronger theorem.
**Daniel Filan:**
Okay. Yeah, so discussing this theorem. So if I think about institutions where people can be rewarded or punished based on what they know, it seems we already have a few of those that people are broadly okay with. So, for instance, the stock market I have in my time-[crosstalk 00:35:38]
**Andrew Critch:**
I take exception to the claim that people are broadly okay with the stock market, but I'll agree that the stock market is, in fact still happening.
**Daniel Filan:**
Yeah, I think people are okay with some individual trades. If you and I trade in equity, the person who knows more about the value of that equity in the future has an edge on that trade. I guess I haven't seen polling, my assumption is that most people are okay with the idea that I can trade in equity with you, but maybe you don't think that.
**Andrew Critch:**
Well. Claims about what most people think are okay are a little bit dangerous. And I'm a little bit uncomfortable making them.
**Daniel Filan:**
How are they dangerous?
**Andrew Critch:**
Well, they can force people into equilibria that they didn't want to be in. So, you're in a room full of 30 people and somebody says, "Well, we're all clearly okay with the meetings being at 6:00 A.M every week. Right?" And there's this brief pause. And then, now, the meetings are at 6:00 A.M every week. And a bunch of people objected, but they didn't know everybody else would have objected. So they didn't object. And so when you say everybody agrees, blah. I'm if I can think of some people who don't agree with blah, I'm hesitant to just get on the everyone agrees with blah train because I'm oppressing that view if I do it. So I'm not going to get on board with the everybody's okay with the stock market claim, and I might not even get on board with the everyone's okay with asymmetric information trades, although I would agree more people are okay with that.
**Andrew Critch:**
I could imagine a future where, education is a human right right now, I can imagine a future where informed trade is a right and not thinking that's ridiculous. And it would create a lot more work for the economy to do, to produce information for people whenever they enter trade. But I think that's a tenable position. I know people who think that bargaining without transparency is just bad and wrong. And I'm like "yeah, I don't know". I don't want to close the book on that by just saying everybody agrees with it already.
**Daniel Filan:**
Yeah, I guess. To follow this tangent a bit, so I think one argument against ensuring that every trade is transparent is sometimes it might impose a really high communicative overhead, for instance- [crosstalk 00:38:39]
**Andrew Critch:**
It takes a lot of economic work. Exactly.
**Daniel Filan:**
Yeah. And especially where it's like, suppose we're going to have a podcast today. You know more potentially about what you'd be like on a podcast than I do and what if that knowledge is implicit. It's based on your experience of you talking to people over a lifetime that I don't have. It seems hard. I'm not even sure I know what it would look like for that trade to be fully informed- [crosstalk 00:39:11]
**Andrew Critch:**
I mean I'm definitely comfortable saying some people sometimes are okay with asymmetric information trades and I was okay with this one. And I guess you were too. And, you know, yeah.
**Daniel Filan:**
All right.
**Andrew Critch:**
But I think that was just a prelude to some other claim you were going to make, and I think you can still make that subsequent claim without couching it in the "everybody agrees that asymmetric information trades are fine."
**Daniel Filan:**
Yeah, I guess the claim that I might make is we have institutions that have asymmetric information trades or at least trades where participants believe different things when they make the trade and those institutions seem like they work. [crosstalk 00:40:14 ]
**Andrew Critch:**
Yeah, I don't want to undercut that but flag for minor potential disagreements, but go on.
**Daniel Filan:**
Yeah. So if I think about the stock market and the mating market, these are two, well, one of them is more like a market than the other, but there are two cases where people make asymmetric information trades or at least trades with different beliefs. The stock market, as far as I can tell, seems to successfully serve the purpose of predicting itself in the future and the mating market seems- [crosstalk 00:41:02]
**Andrew Critch:**
Sorry what do you mean by mating market?
**Daniel Filan:**
I mean the market by which, humans pair off and become romantic partners and maybe they pair bond for life.
**Andrew Critch:**
Okay.
**Daniel Filan:**
The mating market seems roughly successful in getting a pretty large majority, but not everybody, a romantic partner eventually, and importantly, both of these seem stable institutions.
**Andrew Critch:**
Yeah.
**Daniel Filan:**
It seems they are above people, probably they're not literally above everybody's BATNA, but they're above most people's BATNA. They're roughly doing what they're trying to do and we're not seeing really big revolts against them.
**Andrew Critch:**
I mean, so there have been revolts against the stock market.
**Daniel Filan:**
Yeah.
**Andrew Critch:**
And so. I guess that's that's my counterpoint.
**Daniel Filan:**
Yeah, that's that's a decent counterpoint. Yeah. So which revolts are you thinking of specifically? When I think of revolts in that class, I'm not sure I can think of ones that were actually a stock market- [crosstalk 00:42:30]
**Andrew Critch:**
What's Occupy Wall Street? Right. Let's just take the meme we are the 99%. What happens in a world where 1% of people acquire the resources necessary to make the best predictions about where other resources are going to go? And gain so much advantage from that, that they just dominate the exchange of resources to the point of accumulating most of the resources under the control of a small minority of people.
**Andrew Critch:**
Now is that good or is that bad? Well, it has properties, right? It rewards effort that makes people and institutions better at prediction. So there's an incentive there to get better at prediction. But it also is a little incestuous in the sense that these people and institutions are just protecting each other. It's like you said, the stock market is good at predicting itself. But I mean, what is the stock market doing? It's ensuring efficiency of trade among the owners of a very large number of business activities. And is that good?
**Andrew Critch:**
Well, I don't know, maybe constantly changing the ownership of very large, powerful entities creates a diffusion of responsibility where you can just rotate in and out board members and CEOs if things go wrong. And so it's not clear that the stock market is a good thing for all the people who didn't end up in control of it.
**Andrew Critch:**
And I mean, I worked in finance, so I think the stock market does some good and I don't think it's all bad and I wouldn't erase the stock market right now if I had an anarchy button, but I think it has some problems and I think many people agree that it has some problems. And one of the problems is that it just heaps resources on to people who are good at predicting it or institutions that are good at predicting it. It really leaves out people who don't have those big powerful institutions behind them to help them make their financial decisions.
**Daniel Filan:**
I mean. So this feature of the stock market doesn't generalize to other things we're talking about - the one that I'm about to say -
**Andrew Critch:**
Okay, sorry.
**Daniel Filan:**
Yeah, but I mean, the stock market does have this feature where you can just buy the whole market. And then if you're more willing to wait for resources than the market, quote, unquote, is you can just buy the market, wait and then eventually get reasonably wealthy from doing that, right?
**Andrew Critch:**
You're just saying the stock market has- [crosstalk 00:45:47] stock market value has gone up over time and index funds help protect you from the adverse selection of choosing which things to own?
**Daniel Filan:**
Yeah, I guess I'm saying that, even if I don't know much about individual stocks or I have very little information about what companies are doing what or where profit is being generated. I can buy an index fund and- [crosstalk 00:46:11]
**Andrew Critch:**
But you will not become a billionaire by buying index funds.
**Daniel Filan:**
That seems correct, unless I'm a 999 millionaire to start with.
**Andrew Critch:**
Yeah. I didn't mean to doxx you as a non billionaire there. I just know you're not listed on any of the public registries of billionaires. You could still be a billionaire.
**Daniel Filan:**
You don't know how much my shell companies have. Yeah. And so I guess the analogy is, if we made A.I. that literally rewarded people who knew stuff. It would reward some kind of insider trading or fooling others or something, and this would not be a stable situation. Is that a summary of what you think?
**Andrew Critch:**
That is a thing that I think yeah, I wouldn't say it's a summary, but it's a thing that I believe. Yeah. And it's not really addressed by the NRL paper, it's more like the NRL paper is pointing out - NRLP, negotiable reinforcement learning. It's pointing out if you Pareto optimize for cooperation or if you just Pareto optimize ex ante, you get this bet settling outcome. And the paper doesn't really explore very much about the unfairness of that outcome. There's a little bit in there, but that's more of a future work type of thing that I hope people will think about.
**Daniel Filan:**
Okay. Yeah, so speaking of that. Yeah, I guess you wrote this paper because you thought that it would have consequences in the world. Or I gather that you did.
**Andrew Critch:**
Yeah.
**Daniel Filan:**
Yeah. Can you say a bit more? What things do you hope will happen because you wrote this paper?
**Andrew Critch:**
Yeah, I mean, just proximally, I hope researchers and A.I. students, faculty, industry folks who are building sequential decision making systems will take an interest in differences in belief. As an interesting bearing on what happens with the system and I hope that they can see, wow, there's something different about how belief bears on the system from the way preference bears on the system. Or how belief ought to bear on the system versus how preference ought to bear on the system. Preferences, you just leave them as they are or you try not to disturb them. Whereas beliefs, you have this opportunity to share information and update each other's beliefs, for example, that's completely missing from the paper. I'd to see people designing mechanisms for individuals and institutions who govern powerful A.I systems to share information with each other. So it evens the playing field. So they all have the same information. There might be a small benefit to the institution that had more information at first or something to sell their information to other institutions. But I'm hoping we don't lock the entire future into some technological equilibrium that really disenfranchised a massive amount of people or a massive amount of different value systems that just didn't manage to have a say on what A.I. does.
**Andrew Critch:**
And to an extent, there's a lot of people thinking about fairness and accountability in A.I., and then transparency at least to engineers. So there's a cluster, fairness, accountability, transparency that really appeals to me here. And so, you know, maybe if people with those interests could think more about differences in belief and how that's going to play out in sequential decision making, policies are going to run for a long time, how should that play out? Yeah.
**Daniel Filan:**
Okay, yeah, so I guess a follow-on from that, I see this as similar to the social choice literature and I guess if you're hoping, and I'm not sure that you actually said this, quite, but if it's true that you're hoping that this research will eventually help facilitate bargaining and thinking about, okay how do we actually get people to cooperate over the creation of powerful artificial intelligences, do you think the existing social choice literature has analogously done a good job at fostering cooperation?
**Andrew Critch:**
Well, that's interesting. I have wondered this. And I don't know, I mean, Aumann and Schelling and people like that were commissioned by RAND Corporation to try and and devise nuclear disarmament protocols. And they tried and they admittedly failed. Their writings on this say, look, we tried, we couldn't come up with anything and they seemed to have earnestly tried. I don't think they were lazy. I mean maybe they were, they didn't seem that way to me. Maybe they're just brilliant and they can have good ideas while being lazy. But, so in a sense, I'm going to say no. Like, there were things that the world called on mathematicians and game theorists and decision theorists to figure out that they didn't figure out.
**Andrew Critch:**
And I think we're not done making that call. That happened during the Cold War, right? So it was time to make that call and some of the greatest minds came together to think about disarmament. How do you gradually deescalate a threatening situation between nations? But there hasn't been that much work since, there hasn't been that flurry of brilliance in the area of how to foster peace and cooperation as there was back then in the Cold War. And I hope, I think with the advent of increasingly capable A.I. technology, we're going to see more and more brilliant people taking an interest in how to maintain peace and harmony in the world with that much capability. So I'm half making a prediction and half making a bid that says let's revisit these foundational questions about how to achieve cooperation and see if we can do better than the 70's.
**Daniel Filan:**
Yeah.
**Andrew Critch:**
Yeah.
**Daniel Filan:**
Yeah. It's interesting that we stopped because it's not as if we don't currently live in a world where many countries have a lot of nuclear weapons or many countries disagree about who gets what bit of land. Right?
**Andrew Critch:**
It's true, it's not as if we don't live in that world.
**Daniel Filan:**
So another question on consequences. So at NeurIPS in 2020 papers are supposed to have a broader impact statement and the broader impact statement is supposed to include how could this research have a negative consequence if there's a plausible way. Suppose your research ended up making the world worse, by means other than just opportunity costs. There was something else that people could have done that was great, but instead they paid attention to yours, it actively made the world worse. How do you think that would have happened?
**Andrew Critch:**
Yeah, I mean, I guess I've alluded to it, right? Someone just grabs the formula from negotiable reinforcement learning and just runs it and then a bunch of people end up unwittingly signed on to a protocol that's exciting to all of them at start, because according to their own individual beliefs, it looks great. But some of the people's beliefs about how it's going to go down are wrong. And then they end up getting really screwed over. And I don't want that to happen. So do I have a theorem for how not to do it or what's the correct balance between getting everybody together versus making sure everybody's signed on to something fair. No, I don't, but that's another potential future work. Maybe there's an interesting boundary there between fairness and unity to be hugged.
**Andrew Critch:**
But, yeah, if it goes wrong, but that doesn't seem likely. So, it's just one idea who's going to use this, but maybe it's the mode of the distribution of unlikely ways this idea could end up having a large negative impact.
**Daniel Filan:**
Okay. So speaking of consequences, the paper has been out for a while. How's the reception been?
**Andrew Critch:**
Yeah, there have been a bunch of people who came up to me to try to take an interest in it. It seems like what happens is, it seems like there are little pockets of interest in it, but that are isolated or not even pockets.There hasn't been a whole lab of people who all got interested in it. And so there have been, person from this group person from group, "Oh, this is interesting". But I think it's hard for people to stay motivated on projects when the rest of their working environment is not adequately obsessed with it or something, so there hasn't been a flurry of follow up work on it and maybe now's the time for that. Maybe today, maybe this podcast, I also have more free time now. I just finished up a giant document. Maybe now is the time that we can try to get a cluster of people working together to solve the next problems in what I would call machine implementable social choice. But that hasn't happened yet. We'll see.
**Daniel Filan:**
Right. So speaking of that, do you have any final things you'd like to say? Or if people are interested in following your research, how should they do so?
**Andrew Critch:**
Well, I guess I don't have a Twitter account, if that's what you're asking. But thanks for asking. And the easiest way to notice if I write something is just [subscribe to me on Google Scholar](https://scholar.google.com/citations?user=F3_yOXUAAAAJ&hl=en&oi=ao). You can make Google Scholar alerts that just tell you when someone publishes a paper. So that should work. And I have a website, but that's a more active attention-intensive way of keeping track of what I do compared to Google Scholar. And maybe I'll get a Twitter account someday, it won't be this year, I think. Maybe, I could lose that bet, but it won't be in the next several months, that's for sure.
**Daniel Filan:**
Okay, well, thanks for talking with me Andrew. And to the listeners, thanks for listening and I hope you'll join us again. |
9e88ea91-1e92-4694-8f61-ce51a7473e5c | trentmkelly/LessWrong-43k | LessWrong | Can we safely automate alignment research?
(This is the fifth essay in a series that I’m calling “How do we solve the alignment problem?”. I’m hoping that the individual essays can be read fairly well on their own, but see this introduction for a summary of the essays that have been released thus far, and for a bit more about the series as a whole.
Podcast version (read by the author) here, or search for "Joe Carlsmith Audio" on your podcast app.
See also here for video and transcript of a talk on this topic that I gave at Anthropic in April 2025. And see here for slides.)
1. Introduction
In my last essay, I argued that we should try extremely hard to use AI labor to improve our civilization’s capacity to handle the alignment problem – a project I called “AI for AI safety.” In this essay, I want to look in more detail at an application of “AI for AI safety” that I view as especially important: namely, automating alignment research. In particular: I want to try to get clearer about the different ways that automating alignment research can fail, and what we can do about them.
I’m especially interested in ways that problems evaluating alignment research might pose a barrier to automating it. Here, one of my key points is that some types of alignment research are easier to evaluate than others. In particular, in my opinion, we should be especially optimistic about automating alignment research that we can evaluate via some combination of (a) empirical feedback loops and (b) formal methods. And if we can succeed in this respect, we can do a huge amount of that type of alignment research to help us safely automate the rest (what I call “conceptual alignment research”).
Overall: I think we have a real shot at safely automating alignment research. But:
* Evaluation: We need to figure out how to adequately evaluate automated alignment research – and especially, alignment research we can’t test empirically or check using formal methods;
* Scheming: We need to either (a) avoid/prevent AIs actively scheming |
603bfcbc-b6ce-4320-a10f-7493e966ee5d | trentmkelly/LessWrong-43k | LessWrong | Bring Back the Sabbath
Epistemic Status: Several months of experimentation
Previously: Choices are Bad, Choices Are Really Bad, Complexity Is Bad, Play in Easy Mode, Play in Hard Mode, Out to Get You, Slack
For More Thoughts After: Sabbath Commentary
Alternate Take (Endorsed): Sabbath Hard and Go Home.
Slack is life. It is under attack. We must fight for it.
Choices Are Bad. Really Bad. We need a break.
Complexity Is Bad. We need a break.
Work is exhausting. We need a break.
Relaxation is hard. Our attempts fail or backfire.
The modern world is Out to Get You. We need a break.
We need time for ourselves. Time that is truly our own.
Without setting aside such time, that won’t happen. Even when you take time, you’ll be continuously choosing to take time, and… well, whoops.
Modern life made the problem worse, but the problem is ancient. The ancients had an answer.
We need rules. We need ritual.
We need the Sabbath.
Cabin in the Woods
The parallels of my and Ben Hoffman’s Sabbath realizations are striking.
A few months ago, like Ben, I needed a break. My job puts me under constant pressure. My weekends weren’t refreshing me. Like Ben, I experimented with camping. Like Ben, I had no spare battery, and left my phone off. I read The Great Transformation. I had meant to do that for weeks. I loved the world leaving me alone. Like Ben, I could relax, slow down, think.
I wasn’t worried about things I could be doing – I couldn’t do them.
Could I get this without the trip? Friends had started hosting Friday night dinners. What about the whole thing? What if we brought back the Sabbath?
Tradition makes rules easier to justify and explain, to others and yourself. These rules were time tested. I could take them and make them my own.
I thought about the components. Which made sense? What rules would let me cut the enemy, and relax?
Return of the Ritual
Rituals need clear beginnings and endings.
Sabbath begins with candles. One lights two candles, and recites a blessing.
For the e |
19c0aef8-101c-45f5-a181-0fa7655a858f | awestover/filtering-for-misalignment | Redwood Research: Alek's Filtering Results | id: post1090
[Metadata: crossposted from https://tsvibt.blogspot.com/2023/01/the-voyage-of-novelty.html . First completed January 17, 2023. This essay is more like research notes than exposition, so context may be missing, the use of terms may change across essays, and the text might be revised later; only the versions at tsvibt.blogspot.com are definitely up to date.] Novelty is understanding that is new to a mind, that doesn't readily correspond or translate to something already in the mind. We want AGI in order to understand stuff that we haven't yet understood. So we want a system that takes a voyage of novelty: a creative search progressively incorporating ideas and ways of thinking that we haven't seen before. A voyage of novelty is fraught: we don't understand the relationship between novelty and control within a mind. What takes the voyage, from where to where? The voyage is taken by some AI system, from not understanding what we want it to understand, to understanding all that. A schematic: Novelty Novelty is structure that is new to a mind. A mind acquires novelty via creativity. Endosystemic novelty is novelty that fits in with the preexisting mind in ways analogous to how preexisting elements of the mind fit in with the mind; e.g., a new sub-skill such as "opening a car door", or a new word. Diasystemic novelty is novelty (new structure) that involves through-going changes to the mind; e.g. a new heuristic like "investigate boundaries" applied automatically across many contexts. At the extreme end of diasystemic novelty are cognitive realms distinct from the cognitive realm currently occupied by the mind. Structure hand-holdable by humanity Where does the voyage of novelty start? It starts wherever the mind is. But where can it safely start? A lot of structure has the property that it is easy enough for humans to render non-dangerous a mind's exploration (discovery, creation) and exercise (application, expression) of that structure. Humanity can "hold a mind's hand" through its voyage of novelty through such structure. Some overlapping reasons that structure can be hand-holdable, with examples: The structure just doesn't contribute much to a mind having large effects on the world. For example, a mind might know about some obscure species of grass that no human knows about; a human can't understand the mind's thoughts in absolutely full detail, at least without zerself learning about the species of grass. But it just doesn't matter that much, assuming the grass isn't a source of extremely potent toxin or something. The structure is easy to intervene on to limit the extent of its influence over the rest of the mind. For example, one can limit the information flow between graspable elements. (However, this sort of hand-holding is very fraught, for example because of conceptual Doppelgängers : the rest of the mind can just reconstitute the supposedly boxed structure in a Doppelgänger outside of the box.) The structure is easy for humans to understand fully, including all its implications for and participation in thinking and action, so that it doesn't contribute inscrutably to the mind's activity. See gemini modeling . For example, the mental model [coffee cup] is hand-holdable, as humans have it. This is a central example of hand-holdability: the mind (hypothetically, at least) in question does the same things, thinks the same way, with respect to coffee cups, as do humans; what's in common between the mind's activity in different contexts with respect to coffee cups, is also in common with humans. (It's not clear whether and how this is possible in practice, for one thing due to essential provisionality : the way a mind thinks about something can always change and expand, so that [the structure that currently plays the analogous role in the mind, as was played by the structure that was earlier mostly coincident with some structure in a human], as a path through the space of structure, will diverge from the human's structure.) The structure is somehow by nature limited in power or bounded in scope of application. For example, AlphaZero is hand-holdable because it just plays a narrow kind of game. See KANSI . A non-example would be the practice of science. Since science involves reconfiguring one's thoughts to describe an unboundedly expanding range of phenomena, it's not delimited--that is, it isn't easily precircumscribed or comprehended . What are some other ways structure can be hand-holdable? This is not a very clear, refined idea. One major issue is that hand-holdability is maybe unavoidably a property of the whole mind. For example, the hypothetical ideas that constitute the knowledge of how to make a nuclear weapon with spare junk in your garage could be individually hand-holdable, but together are not hand-holdable: we don't know how to enforce a ban on weapons you can manufacture that easily. For another example, humans have some theory of mind about other humans; at the human level of intelligence, most people's theory of mind is hand-holdable, but at a much higher level of intelligence would not be hand-holdable, and there'd be profounder and more widespread interpersonal manipulation than there is in our world. The theory of mind at a much higher level of intelligence would involve a lot of genuine novelty, rather than just scaling up the exercise of existing structure, but at some point the boundaries would blur between novelty and just scaling up. Another vagueness, is the idea of changing hand-holdability. If a human comes to deeply understand a new idea, that should be considered to potentially dynamically render that concept hand-holdable: now the human can comprehend another mind's version of that idea. On the other hand, I'd want to say that architecting an AI system so that it avoids trampling on other agents's goals does not count as rendering that AI's creativity hand-holdable. It seems intuitively like it's the AI doing the work of interfacing with the novelty to render non-dangerous (to us) the AI's discovery and exercise of that novelty. Human understanding In the schematic, a single human's knowledge is represented as a reticular near-subset of structure hand-holdable by humanity. The set protrudes a little outside of what's hand-holdable. The endosystemic protrusion could be for example the instructions and mechanical skills to make a garage nuclear weapon. The diasystemic protrusion could be for example the intuitive know-how to short-circuit people's critical law-like reason, in favor of banding together to scapegoat someone. The set of structure hand-holdable by humanity is not covered by the understanding of humans. For example, no human understands why exactly AlphaZero (or Stockfish) makes some move in a chess game, but those programs are hand-holdable. The dark matter shadow of hand-holdable structure "Understanding" here doesn't mean explicit understanding, and should maybe be thought of as abbreviating "possessed structure" ( "possess" from the same root as "power" and "potent", maybe meaning "has power over" or "incorporates into one's power"). One can understand (i.e. possess structure) in a non-reflective or not fully explicit way: I know how to walk, without knowing how I walk. If we don't know how I walk, other than by gemini modeling , then how I walk is dark matter : we know there is some structure that generates my walking behavior, and we can model it by calling on our own version of it, but we don't know what it is and how it works. The schematic shows the dark matter "shadow" of structure that's hand-holdable by humanity: the collection of structure that generates that hand-holdable structure. A dark matter shadow isn't unique; there's more than one way to make a mind that possesses a given collection of structure. Of the structure in the dark matter shadow of hand-holdable structure, some is endosystemic. We have concepts that we don't have words for, but that we use in the same way we use concepts which we do have words for, and that we build other hand-holdable structure on top of. (Though strictly speaking concepts can constitute diasystemic novelty, depending on how they're integrated into the mind; this is an unrefinedness in the endo-/dia-systemic distinction.) Some is diasystemic, for example whatever processes humans use to form discrete concepts. The schematic shows much more diasystemic than endosystemic dark matter. That's an intuitive guess, supported by the general facts that dark matter tends to be generatorward (generators being necessary for the observed structure, and also harder to see and understand), and that generators tend to be diasystemic. Of the structure in the dark matter shadow of hand-holdable structure, some is itself hand-holdable. For example, at least some unlexicalized concepts should be easy for us to handle. Imagine not knowing the modern, refined mathematical concept of "set" or "natural number", but still being able to deal intuitively, fluently with another mind's conduct with respect to (ordinary, small) sets, such as identity under rearrangement and equinumerosity under bijection (see The Simple Truth ). Some is not hand-holdable, e.g. theory of mind as mentioned above. Dark matter has a tendency to be not hand-holdable, even though humans possess a sufficient collection of dark matter structure to generate much of hand-holdable structure. That's because dark matter is not understood, and because dark matter is more generatorward and therefore more powerful. The dark matter shadow of structure hand-holdable by humanity may be easy to scale up , and therefore difficult to hand-hold. Desired structure We want an AI in order to possess structure we don't already possess. So there's some target collection of structure we'd like the AI to acquire. This probably includes both endosystemic novelty, e.g. implementable designs for nanomachines that we can understand, and also diasystemic novelty, e.g. some structuring of thought that produces such designs quickly. More precisely, that's novelty which is endosystemic or diasystemic (or parasystemic) for the AI. The novel understanding that the AI gains which we don't already possess is, for us, ectosystemic novelty, unless we follow along with the AI's voyage. Voyage The schematic again for the voyage of novelty: So, as an aspirational trajectory: an AI system could start with structure that is easily understood or bounded by humans; then the AI system gains novelty through creativity; the new structure includes the structure that generates human understanding, concepts new to humanity, and forms of thought new to humanity; the new structure might include very alien cognitive realms; until the AI has some target collection of structure. Having an AI follow such a trajectory is fraught , since we don't know how to interface with novelty and we don't know what determines the effects of a mind . Thanks to Sam Eisenstat for morally related conversations. Thanks to Vivek Hebbar for a related conversation. |
988fc9b1-e1dc-4af4-9edd-1684e0ccf1d4 | trentmkelly/LessWrong-43k | LessWrong | If only we had taller been
|
da592ad6-958e-44ba-b01c-38c78ceb3bda | trentmkelly/LessWrong-43k | LessWrong | Californians, tell your reps to vote yes on SB 1047!
See the event page here.
Hello Californians!
We need you to help us fight for SB 1047, a landmark bill to help set a benchmark for AI safety, decrease existential risk, and promote safety research. This bill has been supported by some of the world’s leading AI scientists and the Center of AI Safety, and is extremely important for us to pass. As Californians, we have a unique opportunity to inspire other states to follow suit.
SB 1047 has a hearing in the Assembly Appropriations Committee scheduled for August 15th. Unfortunately, due to misinformation and lobbying by big tech companies, the bill risks getting watered down or failing to advance. This would be a significant blow against safety and would continue the “race to the bottom” in AI capabilities without any guardrails.
We need you to do the following to save the bill. This will take no more than 5 minutes:
* Email the Chair (Buffy Wicks, D) and Vice-Chair (Kate Sanchez, R) of the Appropriations Committee, and ask them to advance the bill without watering it down.
* Buffy Wicks: assemblymember.wicks@assembly.ca.gov
* Kate Sanchez: assemblymember.sanchez@assembly.ca.gov
* Email templates and examples can be found here.
* Call the offices of the Chair and Vice-Chair in support of the bill. The best day to do this is August 13 or 14.
* Buffy Wicks (D): (916) 319-2014
* Kate Sanchez (R): (916) 319-2071
* These calls can be very short (~1 minute) and straightforward. See here for more guidance.
This document has additional information about the bill and other ways to help. [But some of the dates are wrong! This post is up-to-date.]
Please try to get this done as soon as possible, and let us know if you need any help. Your voice matters, and it is urgent that we push this before it’s too late.
Thank you so much for your support! |
3c163298-45a8-4caf-babc-456f07e4ca7d | trentmkelly/LessWrong-43k | LessWrong | Robustness & Evolution [MLAISU W02]
Welcome to this week’s ML Safety Report where we talk about robustness in machine learning and the human-AI dichotomy. Stay until the end to check out several amazing competitions you can participate in today.
Watch this week's MLAISU on YouTube or listen to it on Spotify.
Robust Models
Robustness is a crucial aspect of ensuring the safety of machine learning systems. A robust model is better able to adapt to new datasets and is less likely to be confused by unusual inputs. By ensuring robustness, we can prevent sudden misalignments caused by malfunction.
To test the robustness of models, we use adversarial attacks. These are inputs specially made to confuse the model and can help us create defense methods against these. There are many libraries for adversarial example generation in computer vision but the new attack method TextGrad creates adversarial examples automatically for text as well. It works under the two constraints of 1) text being much more discrete than images and therefore harder to modify without being obvious and 2) still ensuring fluent text, i.e. making the attacks hard to see for a human. You can see many more text attacks in the aptly named TextAttack library.
In the paper “(Certified!!) Adversarial Robustness for Free!” (yes, that is it’s name), they find a new method for making image models more robust against different attacks without training their own model during defense but using off-the-shelf models, something other papers have not achieved. Additionally, they do this and get the highest average certified defense rate against the competition.
Additionally, Li, Li & Xie investigate how to defend against the simple attack of writing a weird sentence in front of the prompt that can significantly confuse models in question-answering (QA) settings. They then extend this to the image-text domain as well and modify an image prompt to confuse during QA.
With these specific cases, is there not a way for us to generally test for examples |
d89f4ca0-8f7d-442a-8ee9-da0a16e25018 | trentmkelly/LessWrong-43k | LessWrong | Why Selective Breeding is a Bad Way to do Genetic Engineering
A Brief Intro
During any conversation about genetic engineering, people inevitably bring up worries about eugenics movements of the past and often use the cruelty, bad science, and objective failure of these efforts as an example of why we shouldn't ever try anything remotely related again. In this short post, I'm going to summarize why I think selective breeding of humans is bad both from a moral perspective and ineffective as a means of improving human genes.
Selective breeding at its core involves taking organisms that score well on some test of desirable traits and enabling them to reproduce at higher rates than organisms that score poorly. Despite its many flaws, this technique has lead to amazing gains in both agriculture and animal husbandry, and allowed domesticated corn crops to undergo this incredible transformation over the last few thousand years.
But despite the amazing performance on crops, there are reasons this technique would not work very well on humans.
Humans are Slow Breeders
Every generation that you selectively breed an organism you get some gain in a particular trait. The faster reproduction happens, the faster you see improvements in the trait(s) under selection. Humans are extremely slow-breeding animals. Though humans are capable of reproducing sometime in early adolescence, most humans today opt to wait until their 20's to 30's to have children. This is a very very long time if you want to do selective breeding.
Selective Breeding Leads to an Undesirable Reduction in Genetic Diversity
Genetic diversity is valuable. Because selective breeding can only work by throwing entire organisms out of the gene pool, it naturally ends up reducing a lot of desirable genetic diversity. Even organisms that don't score well overall will still have many good genes. With selective breeding, there is no way to keep this valuable genetic diversity unless one were to select the best X% from every lineage in a population.
Selective Breeding Creates a S |
f557774d-face-420a-8165-b1412ad006fc | trentmkelly/LessWrong-43k | LessWrong | Meetup : Rationality Potluck
Discussion article for the meetup : Rationality Potluck
WHEN: 25 May 2017 06:30:00PM (-0400)
WHERE: 1191 Avenue Hope, Montreal
Eric Chisholm from the Vancouver Rationalist Community is staying at the Macroscope this week. You're invited to come say hi, and talk with other rationality enthusiasts! Feel free to invite friends.
Bring food and/or beverage if possible. Vegan food will be available.
Eric Chisholm, an alumnus from the Center for Applied Rationality, will present the Double Crux technique: a technique for resolving disagreement.
Facebook Event : https://goo.gl/f8Uwfg
Discussion article for the meetup : Rationality Potluck |
24d09810-7e3d-4b7b-b4ac-22e85b18041c | trentmkelly/LessWrong-43k | LessWrong | Doing Nothing Utility Function
One of the questions I've heard asked is "how do you design a utility function that would make the AI do nothing?" That is, how could we put a pause button on an AI so that we could pause it if we wanted to? I had an idea about how one could do this, and am sure it has already been thought of, so I'm curious to know why this doesn't end up working.
Why can't we just write something like:
If (paused) and not (taking actions): utility=PredictedUtilityIfNotPaused
If not (paused): utility=NormalUtilityFunction |
16685957-1d24-4e7d-a316-f578cb577b46 | trentmkelly/LessWrong-43k | LessWrong | Staggering Hunters
Summary: A stag hunt is a situation where people try to coordinate on doing a big, important task but have reasons to individually prefer a safe and comfortable task. Staggering Hunters is a gamified variation of the Stag Hunt.
Tags: Medium, Experimental
Purpose: Practice coordinating with people, even through communication difficulties.
Materials: A bunch of “coins” (which can be small change or poker chips or beads or any other collection of small objects) and a pile of six sided dice (which can be any similar random number generator.)
Announcement Text: The Stag Hunt is a bit of game theory, similar to the more popularly known Prisoner's Dilemma. The idea is that a group of hunters are going out, and they can either cooperate to hunt Stag for a large benefit that only works if everyone chooses it, or individually hunt Rabbit for a small benefit that works even if not everyone agrees. Cooperation is harder than we think; not everybody is always on the same page about what we’re trying to do. Staggering Hunters is a gamified version of this! The LessWrong post here is suggested reading, though you don’t have to have read it to show up and play. We’re going to play the game, then discuss how it went and try again to see if we can do better.
Description: Put the pile of coins and dice in the middle of the table, in reach of as many players as possible.
Explain the following rules:
1. Definitions: There’s a Pile of Coin and a Pile of dice. Those are like the bank in board games, when we add Coin we can do it from the Pile. The Pot is different; at the end of a round everyone will divide up the Coin in the Pot. We’re going to play for three rounds.
2. Setup: Everyone rolls a die in secret, then takes Coin from the central pile for what they rolled. (You rolled a 5, take 5 coins.) Keep your current Coin supply secret.
3. Each round, do the following:
1. Silence: Roll a die. If it’s 4, 5 or 6 you can talk, if not, stay silent this round. (This includes |
f2311b85-306d-49d9-8a45-8652f6c8b659 | trentmkelly/LessWrong-43k | LessWrong | Happiness Engineering
On the same day that Lukeprog posted How to be happy, Scott Adams made a similar post on the Dilbert Blog, Happiness Engineering.
I'm always skeptical when receiving life advice from successful people, because their advice is biased towards taking too much risk, because successful people are selected for having been lucky. But Scott's list doesn't raise any red flags with me, and is admirably concise. |
746363e1-0132-4c03-9f54-0fc146866982 | trentmkelly/LessWrong-43k | LessWrong | Ms. Blue, meet Mr. Green
Blue and Green are references to the MTG Color Wheel personalities. (Very fun reading, but not required to understand this post.)
Ms. Blue: Hi, Mr. Green! Heard you’ve been traveling recently. Where have you been?
Mr. Green: Oh, I’m so glad you asked! I just came back from Bali. I went to a three weeks long workshop, where they taught us to align our chakras and channel cosmic energy into our third eye. I’ve gone to different astral planes, became one with the universe, and communed with god.
Ms. Blue: *Slowly walks backwards, takes out her phone, and starts dialing the local psychiatric hospital.*
Mr. Green: Hey, I’m not crazy!
----------------------------------------
I’m sure most of you have met or heard someone who used this kind of language. I’ve heard it called woo-woo or woo-speak. I’m going to refer to it as “mystical language” here, but I’m open to a better name. For most you, this language is probably an instant cue to tune out the listener. They’re obviously out of touch with reality and what they are saying is such garbage, that it’s not even worth listening to. I certainly used to have that exact reaction. In this post, I want to show how to approach and understand this language, how to learn it, and how you might even derive some benefit from it.
Understanding the language
> “Consciousness is a quantum leap from the physical to the non-physical dimension.” --Sadhguru
I think one particularly annoying thing to Ms. Blue and her tribe is how mystical language often borrows words and concepts from physics, e.g. quantum, energy, dimensions, reality, and so on. In physics, there is a clear separation of the map and the territory. The territory is the source of truth; it’s the thing that doesn’t go away when you close your eyes. Everyone can look at it and agree on what they’re seeing. The map is the model we have of the world. If this model is good, it can make accurate predictions about the territory.
The mystical language has something similar. |
215e4373-60ff-4abc-b097-e60d7ebecad7 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | AGI alignment results from a series of aligned actions
Before we start: AGI, TAI, HLMI and related concepts would ideally all fall within this conceptual framework. The nuances (whether it’s transformative or machine-based etc.) are irrelevant in this framework. Same for “safety”. I use AGI throughout for my own convenience.
+++
TL;DR: this post aims to develop a conceptual framework general enough to fit with every reader's understanding of how to achieve AGI safety: i.e. AGI alignment as a series of actions. The hope is to gradually make it more concrete to aid EA community communication and predictions in this cause area.
+++
I have found many of my conversations about AGI safety as a cause area to be impeded by misunderstandings, even within the EA and long-termist community and even with people with whom I tend to reach similar broad conclusions. For example, even with EAs who agree with the broad conclusion that “AI governance matters for the catastrophic and existential riskiness of AI”, there will be confusion about a possible implication that “international standard-setting organisations matter for catastrophic and existential riskiness of AI.” This confusion “downstream” in the reasoning makes me think either or both of us do(es)n’t really get the broad conclusion upstream. Most EAs with whom I have talked about AGI safety have also vastly different beliefs about the same statements (even though they use the same definitions). There also seems to be no common understanding of what statements are important to have beliefs about, and no common framework to even begin to collectively identify cruxes from.[1]
Beyond the notions of “AGI”, ”timeline”, ”comprehensive AI services”, ”takeoff” and ”likelihood of AGI occurring/being developed”, etc. which have been talked about a bit more, there aren’t widely agreed upon dissolutions into more tangible concepts, let alone subsequent analysis that could help anticipation of future events and course-correcting interventions. We also seem to have common notions of “US-China race dynamics”, “lonewolf trajectories”, … yet, even though I have a clear idea of what *I* mean by it, I am unsure others understand it the same way. These notions function well enough for informal communication, but for the sake of anticipating future developments, I want to ask: what do you mean by “China and the US are racing for AI supremacy”? What does a world where “China” and “the US” are not “racing” for “AI supremacy” look like? As there are dozens of these statements, I think there might be economies of scale in developing a common framework to parse them through.
Common understanding should hopefully aid a convergence in anticipation. I don’t imagine we will be able to confidently predict the exact trajectory of AGI any time soon, but convergence in anticipation helps AGI safety-concerned people coordinate. There is work that no one would have bothered doing without common understanding that it is useful and why it is useful. For example, GCRI’s [survey of AGI projects](https://gcrinstitute.org/2020-survey-of-artificial-general-intelligence-projects-for-ethics-risk-and-policy/) would not have taken place without some common understanding that, ceteris paribus, the intention to develop AGI-like technologies correlates positively with assigned likelihood of developing AGI and that, therefore, to a certain extent, we should try to know of existing AGI projects.
Many EAs' decisions (career choice, research questions, or philanthropic funding) hinge on this common understanding and anticipation. I too think that it is hard to plan for affecting something which you cannot anticipate properly and good anticipation relies on a genuine understanding of the underlying mechanisms and parameters driving the results. Finally, I would like to subject my understanding to others’ views - I have now more time to do that: I can interact productively with the occasional reader and don’t fear as much that it would waste anyone’s time.
Before reaching more directly useful predictions, I hope we can reach a framework general enough to encompass every reader’s existing ideas about how AGI safety might happen i.e. that no-one feels like a key concept on this is not captured by the framework below. If we can achieve this, we would then be better able to communicate about, compare and coordinate existing approaches or ideas for interventions (e.g. “technical AGI safety”, “hardware control” and “AI Governance”), and scrutinize new ones.
I start with 10 concepts that I have found useful across conversations. I attempt to explain, rather than define, the concepts. I spare few words for the concepts I believe are relatively straightforward. You can help by answering these questions:
* Do these concepts suffice to capture (at a high level of abstraction) your understanding of AGI safety as a cause area? I.e. do we all agree on that abstract description?
+ How should the concepts be explained to better capture what you mean?
* If not, could you comment and explain the missing concept(s)? How does it relate to these 10 concepts?
* If yes, feel free to contribute: explain your own understanding, hone these concepts' explanation and perhaps identify sub-concepts that could help make this framework more granular. For example, I haven’t included the notion of “relevance” of an actor, because it brings us distractingly close to questions where there is no agreement (e.g. whether governments, military and academia matter, whether UK matters more than the EU, etc.). I will do so in comments later if these first 10 concept are uncontroversial.
+++10 concepts+++
**1. Series of actions**
Regardless of its alignment, technological trajectory, and speed of takeoff, AGI will occur following a series of numerous relevant actions by one or more human being(s), whether deliberate or not. For the AGI safety cause area, a series of actions is relevant if it leads to AGI.[2] Even “lone wolf hacker”/”AGI in my parents’ garage” pathways will require some actions by more than one human being, be it for procurement of specific hardware, access to education, development of software capabilities, access to data, and so on. Some series of actions are long, meandrous, and involve many human actors (e.g., for-profit research & development of technologies for synchronisation of distributed comprehensive services over several decades); some are more straightforward (e.g., skunkworks project with “test-ready AGI code” as a deliverable); and all are hypothetical. Some result in gradual achievement of AGI-like capabilities embedded in a system, others result in achievement of AGI-like capabilities at the push of a button. Importantly, many actions along the way might be entirely disconnected from the intent to create AGI. Individuals concerned about whether AGI is safe would prefer a series of actions that outputs safe AGI.
**2. Actors**
Each action along these pathways is carried out by humans. Even if some automated intermediaries are involved, the intermediaries are built by humans. These humans are therefore actors in the series of actions. We can consider a simplified model of humans as having objective functions factoring in rewards and costs; sensors, memory and actuators; and a significant amount of noise and design flaws in sensing, input processing and actuation.
**3. Factors**
Factors can be action-specific (affecting a single action) or actor-specific (affecting all that actor’s actions). Many factors come into play for any action - from one’s physiological state and direct environmental stimuli to moral psychological configuration and more sophisticated systems shaping one’s beliefs, attitudes, and skills, such as education, (geo)political situation, and philosophical stance. Concretely, whether Jane presses “Cmd+r” or hesitates half a second longer and remembers to fix an error on the impact regularizer - or whatever the relevant intermediary action may be - is based on these many other factors.[3] These factors are therefore important determinants of actions undertaken by actors.
**4. Interventions**
I also argue these various factors can be altered by the actions of other humans with different objectives, belief systems, etc. (e.g., a colleague working on the impact regularizer, a product manager urging that development proceeds more quickly, an actor lauding software perfectionism in the latest blockbuster). To make the series of actions converge towards the safe AGI outcome, interventions to alter these factors are needed along the series. Some interventions alter only one action’s determining factors (e.g., a colleague talking to Jane just before she presses “Cmd+r”), while other interventions alter many actions’, through their influence on actors’ factors (e.g., a PR campaign promoting Stuart Russell’s Human Compatible book). Similarly, some interventions alter factors significantly (e.g., manager bilaterally pressuring Jane for performance improvement and delivery 1 hour before the action at hand), while other interventions alter factors only marginally or at a distance (e.g., a conversation with Jane on alignment, robustness and assurance at a conference 2 years before the action).
**5. Uncertainty**
Moreover, we are uncertain about the series of actions that could output safe vs. unsafe AGI and which are the most influential factors for shaping the relevant series of actions. There are many factors that pertain to many actions in many different possible series, and there are many interventions that may cut in both directions.
**6. Strategic efficiency**
A more efficient strategy is a portfolio of interventions that collectively more significantly affects a greater number of more relevant actions towards a safer outcome, using less resources as the given comparable (add “weakly” before every comparative if you prefer ≤ instead of <, and any subset of these comparatives being strictly defined is sufficient to gain efficiency).
**7. Strategic effectiveness**
An effective strategy converts available resources (including through collecting extra resources if needed) into confidence that AGI will be safe.
**8. Strategic research**
Under this framework, research activities pursuing safe AGI are useful when they reduce uncertainty and enable the development of more efficient strategies to positively influence actions (e.g., answering what series of actions are more or less likely, what factors are most important, and what interventions are most significant).
**9. Direct work**
Direct work activities are the implementation of strategies and associated interventions. They are useful when they increase confidence that the outcome of the current series of actions is safety-enhancing. For example, developing elements of solutions to the AGI alignment problem would increase confidence that many series of actions result in safe AGI.
**10. Waste**
For individuals solely concerned about whether AGI is safe, if strategic research doesn’t reduce uncertainty or identify efficient strategies, or direct work output doesn’t increase confidence in the outcome being safe, it is wasteful.
Footnotes:
[1]: Please do share if you know of any. If there are, it seems they are not being referred to by AGI safety-concerned individuals in the EA community, so I’ll edit this post to further disseminate them.
[2]: While this framework has been developed with AI governance in mind, it can be applied to a broad range of cause areas and subareas where the objective is to intervene to influence “target actions” (e.g. the decision of whether to donate 10% of one’s income, the decision to impose a different tax rate on meat products, the decision to make a grant for foresight research, …)
[3]: This example of action immediately preceding takeoff is chosen for the sake of simplicity. I don’t want to make AGI development (or alignment) sound “easy”, but any other more sophisticated AGI-generating action would require further assumptions. I am very interested in understanding better the likely ultimate actions to counter them, but this exploration is beyond the scope of this post. |
ad303296-8ae9-4934-8a5e-17ee1d47e4a4 | trentmkelly/LessWrong-43k | LessWrong | Game Analysis Index
This post links to this blog’s posts discussing game design, balance, economics and related topics, as well as any strategy posts. It does not contain new content.
Much of the blog is relevant to gaming, but these are the explicitly on-topic posts.
ETERNAL SEQUENCE
Eternal, and Hearthstone Economy versus Magic Economy
The Eternal Grind
Eternal: The Exit Interview
ARTIFACT / CARD REBALANCING SEQUENCE
Review: Artifact
Artifact Embraces Card Balance Changes
Card Collection and Ownership
Card Balance and Artifact
Card Rebalancing, Card Oversupply and Economic Considerations in Digital Card Games
Advantages of Card Rebalancing
Disadvantages of Card Rebalancing
GAME REVIEWS (INCLUDING THOSE LISTED ABOVE)
All games reviewed are recommended, we don’t generally waste time on unworthy games.
Persona 5: Spoiler-Free Review
Review: Artifact
Review: Slay the Spire
Octopath Traveler: Spoiler-Free Review
MAGIC STRATEGY
Deck Guide: Burning Drakes
|
d0581b16-b13a-470a-8f10-5d7fddc20334 | trentmkelly/LessWrong-43k | LessWrong | Proofs Section 1.2 (Mixtures, Updates, Pushforwards)
The previous proofs are here.
Proposition 5: If Bmin⊆Ma(X), then the condition "there is a λ⊙ where, ∀(λμ,b)∈Bmin:λ≤λ⊙" is equivalent to "there is a compact C s.t. Bmin⊆C"
Proof sketch: One direction is immediate from the Compactness Lemma. For showing that just a bound on the λ values suffices to be contained in a compact set, instead of a bound on the λ and b values to invoke the Compactness Lemma, we use a proof by contradiction where we can get a bound on the b values of the minimal points from just a bound on the λ values.
Proof: In one direction, assume there's a compact C s.t. Bmin⊆C, and yet there's no upper-bounding λ⊙ on the λ values. This is impossible by the Compactness Lemma, since (λμ)+(1)=λμ+(1)=λμ(1)=λ.
In the other direction, assume there's a λ⊙ bound on λ for the minimal points. Fix some arbitrary (λμ,b)∈Bmin for the rest of the proof. Now, we will show that all minimal points (λ′μ′,b′)∈Bmin have λ′≤λ⊙, and b′≤λ⊙+b, letting us invoke the Compactness Lemma to get that everything is in a suitable compact set C. The first bound is obvious. Since λ′ came from a minimal point, it must have λ⊙ as an upper bound.
For the other one, by contradiction, let's assume that there's a minimal point (λ′μ′,b′) where b′>λ⊙+b. Then, we can write (λ′μ′,b′) as: (λμ,b)+(−λμ,λ⊙)+(λ′μ′,b′−λ⊙−b)
The first component, (λμ,b) is our fixed minimal point of interest. The second component is an sa-measure, because λ⊙−λ≥0, due to the λ⊙ upper bound on the λ value of minimal points. The third component is also a nonzero sa-measure, because λ′ is nonnegative (it came from a minimal point), and by assumption, b′>λ⊙+b. Hang on, we wrote a minimal point (λ′μ′,b′) as another minimal point (λμ,b), plus two sa-measures (one of which is nonzero), so (λ′μ′,b′) can't be minimal, and we have a contradiction.
Therefore, all (λ′μ′,b′)∈Bmin have b′≤λ⊙+b. Now that we have bounds on λ and b for minimal points, we can invoke the Compactness Lemma to conclude that everything is in a compact |
536d18c7-85dd-44ee-af3c-b3eed842c33d | trentmkelly/LessWrong-43k | LessWrong | Meetup : Logical Counterfactuals, Tel Aviv
Discussion article for the meetup : Logical Counterfactuals, Tel Aviv
WHEN: 05 August 2015 03:00:00AM (+0300)
WHERE: 98 Yigal Alon Street, Tel Aviv
Vladimir Slepnev, who is visiting Tel Aviv, will be giving a talk on logical counterfactuals. We'll meet at 19.00 on Tuesday, August 4 Google, 98 Yigal Alon Street, Tel Aviv, 29th floor (not the Google Campus Floor). We will then move to another room. Vladimir has done advanced research, together with MIRI, on reflective decision theory and related topics. You can read some of his work at cousin_it on LessWrong and under his own name at Intelligent Agent Foundations. If you have Facebook, please confirm at the FB Event so we can get a sense of the number of attendees https://www.facebook.com/events/108233322857656/
Discussion article for the meetup : Logical Counterfactuals, Tel Aviv |
3ebe1b3f-b95d-4501-81ea-129b29fd798f | trentmkelly/LessWrong-43k | LessWrong | Harry Potter and the Methods of Rationality discussion thread, part 7
Update: Discussion has moved on to a new thread.
The load more comments links are getting annoying (at least if you're not logged in), so it's time for a new Harry Potter and the Methods of Rationality discussion thread. We're also approaching the traditional 500-comment mark, but I think that hidden comments provide more appropriate joints to carve these threads at. So as of chapter 67, this is the place to share your thoughts about Eliezer Yudkowsky's Harry Potter fanfic.
The first 5 discussion threads are on the main page under the harry_potter tag. Threads 6 and on (including this one) are in the discussion section using its separate tag system. Also: one, two, three, four, five, six. The fanfiction.net author page is the central author-controlled HPMOR clearinghouse with links to the RSS feed, pdf version, TV Tropes pages, fan art, and more, and AdeleneDawner has kept an archive of Author's Notes.
As a reminder, it's often useful to start your comment by indicating which chapter you are commenting on.
Spoiler Warning: this thread is full of spoilers. With few exceptions, spoilers for MOR and canon are fair game to post, without warning or rot13. More specifically:
> You do not need to rot13 anything about HP:MoR or the original Harry Potter series unless you are posting insider information from Eliezer Yudkowsky which is not supposed to be publicly available (which includes public statements by Eliezer that have been retracted).
>
> If there is evidence for X in MOR and/or canon then it's fine to post about X without rot13, even if you also have heard privately from Eliezer that X is true. But you should not post that "Eliezer said X is true" unless you use rot13. |
b4510abf-7d09-43df-a1f6-841a5c6870f9 | trentmkelly/LessWrong-43k | LessWrong | How to Visualize Bayesianism
Major spoilers for planecrash (Book 2) and for Eliezer's Masculine Mongoose #3.
How Bayesians Lie; How to Lie to Bayesians
> Pyrofessor groaned out loud. “This is why I can’t stand his kind of cognitive augment,” she said. "He can’t just refuse to admit his identity like a normal fucking meta. No, the Goose has to make a big deal out of trying to act exactly like a real human in his shoes. Not because he’s trying to hide who he is. He knows we all know. He’s just being a fucking priss about his interpretation of the mask code. He thinks that if you knowingly behave according to a likelihood function that you can probabilistically distinguish from the likelihood function of a normal, you might as well hang a sign on your forehead. So he acts all ostentatiously precise about his interpretation of Bruce Kent, in order to sniff about how the rest of us are getting it wrong. And he does that knowing all you admiring numbskulls are completely oblivious to how he’s behaving on the augment-to-augment level. God, I hate Bayesians, they’re often right in principle but do they have to be such fucking snobs about it -“
>
> --Eliezer, Masculine Mongoose #3
> Keltham is constantly tracking the Conspiracy world in his mind. That's part of this. He's living in both worlds simultaneously and distinctly and unhesitatingly. There's no pause in him about whether or not the Conspiracy is real, for purposes of accusing Carissa of being in on it within the Conspiracy world. Keltham steps all the way mentally into the world where the Conspiracy is just a thing and Carissa is just part of it, and then in that world when Sevar suddenly vanished away 'to the bathroom' obviously she was up to something in response to his own lecture and obviously the other students' questions were meant as a distraction.
>
> Asmodia sees the game now, has seen the game, even without the enhancement spells she remembers.
>
> Cheliax can't rely on what anything 'looks like', they can't ask if |
b505ce6c-b55d-4c0a-b2b3-0ba6bb28dace | trentmkelly/LessWrong-43k | LessWrong | Omission vs commission and conservation of expected moral evidence
Consequentialism traditionally doesn't distinguish between acts of commission or acts of omission. Not flipping the lever to the left is equivalent with flipping it to the right.
But there seems one clear case where the distinction is important. Consider a moral learning agent. It must act in accordance with human morality and desires, which it is currently unclear about.
For example, it may consider whether to forcibly wirehead everyone. If it does so, they everyone will agree, for the rest of their existence, that the wireheading was the right thing to do. Therefore across the whole future span of human preferences, humans agree that wireheading was correct, apart from a very brief period of objection in the immediate future. Given that human preferences are known to be inconsistent, this seems to imply that forcible wireheading is the right thing to do (if you happen to personally approve of forcible wireheading, replace that example with some other forcible rewriting of human preferences).
What went wrong there? Well, this doesn't respect "conversation of moral evidence": the AI got the moral values it wanted, but only though the actions it took. This is very close to the omission/commission distinction. We'd want the AI to not take actions (commission) that determines the (expectation of the) moral evidence it gets. Instead, we'd want the moral evidence to accrue "naturally", without interference and manipulation from the AI (omission). |
10225c95-bbbc-4554-8f81-5ff98f2faa08 | trentmkelly/LessWrong-43k | LessWrong | Rhythmic Baby Toys
In another debatably educational application of my rhythm stage setup, Nora really likes playing with my bass pedals. She especially likes them when I have them set in a particular way:
* When you tap them they make a noise
* When you tap them with a consistent beat they make them more complex noise.
* If your beat isn't consistent enough they go back to the simpler noise until your tempo is clear again.
Here's a demo:
This has me wondering whether some sort of toy like this could be good at teaching kids a sense of consistent rhythm? The initial noise is enough to make the toy fun, but then if you can be consistent enough you get something more fun. You could probably start with Somewhat relaxed threshold for "consistent enough", and then tighten it over time as they got better?
Comment via: facebook |
35b36001-2b03-47aa-ae22-83abfa8fa080 | trentmkelly/LessWrong-43k | LessWrong | Bragging Thread February 2015
Your job, should you choose to accept it, is to comment on this thread explaining the most awesome thing you've done this month. You may be as blatantly proud of yourself as you feel. You may unabashedly consider yourself the coolest freaking person ever because of that awesome thing you're dying to tell everyone about. This is the place to do just that.
Remember, however, that this isn't any kind of progress thread. Nor is it any kind of proposal thread. This thread is solely for people to talk about the awesome things they have done. Not "will do". Not "are working on". Have already done. This is to cultivate an environment of object level productivity rather than meta-productivity methods.
So, what's the coolest thing you've done this month?
(Previous Bragging Thread) |
c306208b-2cda-41c0-a534-694ba9e1e916 | trentmkelly/LessWrong-43k | LessWrong | Carl Shulman On Dwarkesh Podcast June 2023
This is basically just an exposition of the content of Dwarkesh Patel, I thought it might be useful to have a slightly edited full version here. Full credit to Dwarkesh and Carl.
I find it easier to ingest the information this way and it is a bit quicker too. I did not find the transcripts of Dwarkesh until I had already packaged the subtitles of the videos into one pdf with some formatting. Then I thought I might as well use all the links he provided. So Dwarkesh has really done all the work here except some minor formatting and packaging, if he is not cool with this I will take this down obviously.
But as momentous as this interview seemed to me it might be useful for other people too as easily accessible text so here is the full interview in such a format.
Original Full Interviews can be found here:
Original Transcripts:
https://www.dwarkeshpatel.com/p/carl-shulman
https://www.dwarkeshpatel.com/p/carl-shulman-2
Carl Shulman: Intelligence Explosion, Primate Evolution, Robot Doublings, & Alignment
Intelligence Explosion
D Today I have the pleasure of speaking with Carl Shulman. Many of my former guests, and this is not an exaggeration, have told me that a lot of their biggest ideas have come directly from Carl especially when it has to do with the intelligence explosion and its impacts. So I decided to go directly to the source and we have Carl today on the podcast. He keeps a super low profile but is one of the most interesting intellectuals I've ever encountered and this is actually his second podcast ever. We're going to go deep into the heart of many of the most important ideas that are circulating right now directly from the source. Carl is also an advisor to the Open Philanthropy project which is one of the biggest funders on causes having to do with AI and its risks, not to mention global health and well being. And he is a research associate at the Future of Humanity Institute at Oxford. So Carl, it's a huge pleasure to have you on the pod |
8cd84022-8283-40d7-b16f-6a662fd09ec3 | trentmkelly/LessWrong-43k | LessWrong | Robert Caro And Mechanistic Models In Biography
I absolutely adore Robert Caro’s books: The Power Broker, which is a massive and meticulous study of the life of Robert Moses, the New York Park Commissioner who influenced urban and park design the most in the 20th century, for better and for worse; and The Years of Lyndon Johnson, a series of 4 books (of 5 planned) on the life of former US president Lyndon B. Johnson, remembered both for passing some of the biggest civil rights legislations since the end of the Civil War, and for massively increasing the American involvement in Vietnam.
But it took me some time to figure out why I love his books so much.
I’ve already discussed in a previous blog post why I see biography as so interesting: at its best, the genre builds deep, detailed, and insightful theories of people’s life, of whole periods of histories.
Caro definitely fits the bill: his attention to detail (“Turn Every Page” was the advice he got as a young reporter, and he took it seriously) and commitment to detailed interviews, historical methods, and living where the action happened, impresses even trained historians. In fact, he was a key inspiration behind my original blog post.
Yet he more I read and reread him, the more I feel that what puts him apart is what kind of theory he builds.
For you see, there are two extremes in the spectrum of theoretical models: the phenomenological and the mechanistic.
A phenomenological theory focuses heavily on description and compression: it defines some key concepts, elements, variables, and then links them together. Classical examples include Thermodynamics (often called phenomenological thermodynamics), Newtonian Mechanics and Macroeconomics
Whereas a mechanistic theory (also called gears-level model in rationalists circles) starts from some more basic and fundamental elements, and rederive the behavior of the system under consideration from these building blocks. The mechanistic analogous (or attempts) for the previous phenomenological theories are Statistica |
562ab608-ddd9-420a-b779-09166f3c2507 | trentmkelly/LessWrong-43k | LessWrong | Yes, a blog.
When I recommend LessWrong to people, their gut reaction is usually "What? You think the best existing philosophical treatise on rationality is a blog?"
Well, yes, at the moment I do.
"But why is it not an ancient philosophical manuscript written by a single Very Special Person with no access to the massive knowledge the human race has accumulated over the last 100 years?"
Besides the obvious? Three reasons: idea selection, critical mass, and helpful standards for collaboration and debate.
Idea selection.
Ancient people came up with some amazing ideas, like how to make fire, tools, and languages. Those ideas have stuck around, and become integrated in our daily lives to the point where they barely seem like knowledge anymore. The great thing is that we don't have to read ancient cave writings to be reminded that fire can keep us warm; we simply haven't forgotten. That's why more people agree that fire can heat your home than on how the universe began.
Classical philosophers like Hume came up with some great ideas, too, especially considering that they had no access to modern scientific knowledge. But you don't have to spend thousands of hours reading through their flawed or now-uninteresting writings to find their few truly inspiring ideas, because their best ideas have become modern scientific knowledge. You don't need to read Hume to know about empiricism, because we simply haven't forgotten it... that's what science is now. You don't have to read Kant to think abstractly about Time; thinking about "timelines" is practically built into our language nowadays.
See, society works like a great sieve that remembers good ideas, and forgets some of the bad ones. Plenty of bad ideas stick around because they're viral (self-propagating for reasons other than helpfulness/verifiability), so you can't always trust an idea just because it's old. But that's how any sieve works: it narrows your search. It keeps the stuff you want, and throws away some of the bad stuff so |
0b7c8031-daca-4ce4-b204-7ae704734dd9 | trentmkelly/LessWrong-43k | LessWrong | Log-odds are better than Probabilities
[This is a cross-post from my blog at aizi.substack.com. I'm sure someone has made a point like this before, but I don't know any specific instances and I wanted to give my take on it.]
At my previous job I worked on ML classifiers, and I learned a useful alternative way to think about probabilities which I want to share. I’m referring to log-odds aka logits, where a probability p is represented by logit(p):=log(p/(1-p))[1].
I claim that, at least for Bayesian updates and binary prediction, it can be better to think in terms of log-odds than probabilities, and this post is laying out that case.
Log-odds simplifies Bayesian calculations
Do you do Bayesian updates in your head? I didn’t, in part because the classic Bayes formula is kinda bad to work with:
\[P(H|E)= \frac{P(H)P(E|H)}{P(E)}\]
The first problem is that you need to know P(E), the chance that E is true at all. But the value of P(E) should be irrelevant since we know we live in a timeline where E is true! Of course you can use a formula like this to hide P(E) but at a complexity cost:
\[P(H|E)=\frac{P(H)P(E|H)}{P(E|H)P(H)+P(E \neg H)P(\neg H)}\]
For me, this calculation requires too many operations and cached numbers to do easily in my head.
But more importantly, these formula don’t emphasize how P(H) was updated. Sure, you can say P(H) is being multiplied by P(E|H)/P(E), but that number isn’t really comparable across priors. For instance, if P(E|H)/P(E)=2, that’s a small update if your prior is P(H)=.1 (taking you from 10% to 20%), a huge update if P(H)=.5 (taking you from a coinflip to certainty), and impossible for P(H)>.5. So “P(E|H)/P(E)=2” isn’t a meaningful intermediary calculation.
Now let’s compare the log-odds version. I’ll write L(H) for logit(P(H)):
\[\begin{eqnarray*} L(H|E) &=& \log\left( \frac{P(H|E)}{P(\neg H | E)}\right)\\ &=& \log\left( \frac{\left(\frac{P(H)P(E|H)}{\cancel{P(E)}} \right)}{\left( \frac{P(\neg H)P(E| \neg H)}{\cancel{P(E)}}\right)}\right)\\ &=&\log(\left(\frac{P(H |
40d4fec8-a32c-4fe5-b182-a9fd89fd9e2a | trentmkelly/LessWrong-43k | LessWrong | A Voting Puzzle, Some Political Science, and a Nerd Failure Mode
In grade school, I read a series of books titled Sideways Stories from Wayside School by Louis Sachar, who you may know as the author of the novel Holes which was made into a movie in 2003. The series included two books of math problems, Sideways Arithmetic from Wayside School and More Sideways Arithmetic from Wayside School, the latter of which included the following problem (paraphrased):
The students have Mrs. Jewl's class have been given the privilege of voting on the height of the school's new flagpole. She has each of them write down what they think would be the best hight for the flagpole. The votes are distributed as follows:
* 1 student votes for 6 feet.
* 1 student votes for 10 feet.
* 7 students vote for 25 feet.
* 1 student votes for 30 feet.
* 2 students vote for 50 feet.
* 2 students vote for 60 feet.
* 1 student votes for 65 feet.
* 3 students vote for 75 feet.
* 1 student votes for 80 feet, 6 inches.
* 4 students vote for 85 feet.
* 1 student votes for 91 feet.
* 5 students vote for 100 feet.
At first, Mrs. Jewls declares 25 feet the winning answer, but one of the students who voted for 100 feet convinces her there should be a runoff between 25 feet and 100 feet. In the runoff, each student votes for the height closest to their original answer. But after that round of voting, one of the students who voted for 85 feet wants their turn, so 85 feet goes up against the winner of the previous round of voting, and the students vote the same way, with each student voting for the height closest to their original answer. Then the same thing happens again with the 50 foot option. And so on, with each number, again and again, "very much like a game of tether ball."
Question: if this process continues until it settles on an answer that can't be beaten by any other answer, how tall will the new flagpole be?
Answer (rot13'd): fvkgl-svir srrg, orpnhfr gung'f gur zrqvna inyhr bs gur bevtvany frg bs ibgrf. Naq abj lbh xabj gur fgbel bs zl svefg rapbh |
4e36bcee-4cb0-4caa-aa94-286e78d33346 | trentmkelly/LessWrong-43k | LessWrong | A utility-maximizing varient of AIXI
Response to: Universal agents and utility functions
Related approaches: Hibbard (2012), Hay (2005)
Background
Here is the function implemented by finite-lifetime AI:
,
where is the number of steps in the lifetime of the agent, is the current step being computed, is the set of possible observations, is the set of possible actions, is a function that extracts a reward value from an observation, a dot over a variable represents that its value is known to be the true value of the action or observation it represents, underlines represent that the variable is an input to a probability distribution, and is a function that returns the probability of a sequence of observations, given a certain known history and sequence of actions, and starting from the Solomonoff prior. More formally,
,
where is the set of all programs, is a function that returns the length of a program in bits, and a program applied to a sequence of actions returns the resulting sequence of observations. Notice that the denominator is a constant, depending only on the already known , and multiplying by a positive constant does not change the argmax, so we can pretend that the denominator doesn't exist. If is a valid program, then any longer program with as a prefix is not a valid program, so .
Problem
A problem with this is that it can only optimize over the input it receives, not over aspects of the external world that it cannot observe. Given the chance, AI would hack its input channel so that it would only observe good things, instead of trying to make good things happen (in other words, it would wirehead itself). Anja specified a variant of AI in which she replaced the sum of rewards with a single utility value and made the domain of the utility function be the entire sequence of actions and observations instead of a single observation, like so:
.
This doesn't really solve the problem, because the utility function still only takes what the agent can see, rather than what is ac |
b74452a3-3c74-40ba-9afa-8e2458a3c242 | trentmkelly/LessWrong-43k | LessWrong | Taking it Private: Short Circuiting Demon Threads (working example)
This post is intended as a working example of how I think Demon Threads should be resolved. The gist of my suggestion is:
> Step 1. Make it easy and common to take a conversation private if someone is feeling annoyed/threatened/angry/etc (if it seems like the conversation is actually important. Meanwhile, also make it easier to tap out if the conversation doesn't seem like the best use of your time)
> Step 2. In private chat, two people do their best to communicate honestly, to notice when they are defensive, to productively find the truth as best they can. (I think this is much easier 1-on-1 than in public)
> Step 3. Someone writes a short summary of whatever progress they were able to make (and any major outstanding disagreements that remain), focusing primarily on what they learned and rather than "who's right."
> The summary should be something both parties endorse. Ideally they'd both sign off on it. If that trivial inconvenience would prevent you from actually writing the post, and you both generally trust each other, I think it's fine to make a good-faith effort to summarize and then correct each other if they missed some points.
> Writing such a summary needs to get you as much kudos / feel-good as winning an argument does.
> Step 4. The public conversation continues, with the benefit of whatever progress they made in private.
> Ideally, this means the public conversation gets to progress, without being as emotionally fraught, and every time something comes up that does feel fraught, you recurse to steps 1-3 again.
Qiaochu had a criticism of the Demon Thread article. I had said:
> Demon Threads are explosive, frustrating, many-tentacled conversations that feel important but aren’t.
He responded:
> I want to object to this framing, particularly the "but aren't." It's far from clear to me that demon threads are unimportant. It may seem like nothing much happened afterwards, but that could be due to everyone in the thread successfully canceling out e |
ffbec5bb-aa2c-4667-813f-c4922fa2f0fa | trentmkelly/LessWrong-43k | LessWrong | Sense, Denotation and Semantics
J. Y. Girard, et al. (1989). Proofs and types. Cambridge University Press, New York, NY, USA. (PDF)
I found introductory description of a number of ideas given in the beginning of this book very intuitively clear, and these ideas should be relevant to our discussion, preoccupied with the meaning of meaning as we are. Though the book itself is quite technical, the first chapter should be accessible to many readers.
From the beginning of the chapter:
> Let us start with an example. There is a standard procedure for multiplication, which yields for the inputs 27 and 37 the result 999. What can we say about that?
>
> A first attempt is to say that we have an equality
>
> 27 × 37 = 999
>
> This equality makes sense in the mainstream of mathematics by saying that the two sides denote the same integer and that × is a function in the Cantorian sense of a graph.
>
> This is the denotational aspect, which is undoubtedly correct, but it misses the essential point:
>
> There is a finite computation process which shows that the denotations are equal. It is an abuse (and this is not cheap philosophy — it is a concrete question) to say that 27 × 37 equals 999, since if the two things we have were the same then we would never feel the need to state their equality. Concretely we ask a question, 27 × 37, and get an answer, 999. The two expressions have different senses and we must do something (make a proof or a calculation, or at least look in an encyclopedia) to show that these two senses have the same denotation.
>
> Concerning ×, it is incorrect to say that this is a function (as a graph) since the computer in which the program is loaded has no room for an infinite graph. Hence we have to conclude that we are in the presence of a finitary dynamics related to this question of sense.
>
> Whereas denotation was modelled at a very early stage, sense has been pushed towards subjectivism, with the result that the present mathematical treatment of sense is more or less reduced |
9ac34354-8b8b-4459-81ef-7a8d52e55265 | trentmkelly/LessWrong-43k | LessWrong | Effective Altruism and Rationalist Philosophy Discussion Group
I've created a group on Facebook for rationalists and effective altruists to discuss philosophy (currently has about 120 members). To be clear, this isn't a group where the philosophy has to be related to effective altruism or rationality, but instead a group for Rationalists and Effective Altruists to discuss philosophy.
I know it's already possible to post about philosophy here, but people tend to be reluctant to post unless they've invested a huge amount of effort in writing it up, so it's useful to have somewhere else where there's a lower barrier. |
85b2d119-aaa0-4517-868a-f5ef3188b9e7 | trentmkelly/LessWrong-43k | LessWrong | Andrew Yang on "The Priests of the Decline"
In this post excerpted from his new book, Andrew Yang uses his experience in politics to explain why the US political system rarely accomplishes much. It's strikingly similar to another recent post someone shared from Dominic Cummings: Dominic Cummings : Regime Change #2: A plea to Silicon Valley - LessWrong.
I've copied some of the key sections below:
"I call this dynamic constructive institutionalism — a tendency among leaders to state publicly and even hold the belief that everything will work out, despite quantitative evidence to the contrary, coupled with an inability to actually address a given institution’s real problems...
Indeed, two groups that are especially prone to constructive institutionalism are those that we rely upon both to give us a sense of the problems and to solve them — journalists and politicians.
Journalists are typically trained to be impartial observers, which inhibits them from expressing emotion or opinion. They are supposed to calmly document and present the news. For many, there is an implicit perch of authority and stability. Unfortunately, this has also turned many into market-friendly automatons and cultural guardians who make pro forma gestures about decorum, virtue, and propriety while ignoring the disintegration of trust, the dissipating integrity of their own organizations, or the decline of the American way of life...
If journalists are conditioned to calmly document dispassionately, politicians are conditioned to invoke profundity, resilience, and the greater good at every turn. As a politician you’re like a totem or shaman. You show up to a gathering or charity event to speechify and elevate the proceeding: “Thank you for the incredible work that you’re doing. It’s so important.” Which it is, of course. Though it would be if you didn’t show up too.
You are meant to embody the concerns of the community. You listen patiently to all. You are present. If someone asks you a question, you answer it reassuringly. You express |
3be25afe-eae3-4c77-bc2a-b2df9c7ae7e7 | trentmkelly/LessWrong-43k | LessWrong | [SEQ RERUN] Making History Available
Today's post, Making History Available was originally published on 31 August 2007. A summary (taken from the LW wiki):
> It's easy not to take the lessons of history seriously; our brains aren't well-equipped to translate dry facts into experiences. But imagine living through the whole of human history - imagine watching mysteries be explained, watching civilizations rise and fall, being surprised over and over again - and you'll be less shocked by the strangeness of the next era.
Discuss the post here (rather than in the comments to the original post).
This post is part of the Rerunning the Sequences series, where we'll be going through Eliezer Yudkowsky's old posts in order so that people who are interested can (re-)read and discuss them. The previous post was Failing to Learn from History, and you can use the sequence_reruns tag or rss feed to follow the rest of the series.
Sequence reruns are a community-driven effort. You can participate by re-reading the sequence post, discussing it here, posting the next day's sequence reruns post, or summarizing forthcoming articles on the wiki. Go here for more details, or to have meta discussions about the Rerunning the Sequences series. |
55cf1828-7c17-4c48-a81a-8181bfda445a | trentmkelly/LessWrong-43k | LessWrong | Ungendered Spanish
Spanish has gramatical gender in a way English doesn't:
> una amiga ruidosa — a loud (female) friend
> un amigo ruidoso — a loud (male) friend
> unas amigas ruidosas — some loud (female) friends
> unos amigos ruidosos — some loud (not-all-female) friends
I remember when I was studying Spanish, learning the rule that even if you had a hundred girls and one boy you would use the male plural. My class all thought this was very sexist and unfair, but our teacher told us we were misapplying American norms and intuitions.
It's been interesting, ~twenty years later, following the development of gender-neutral ‑e:
> unes amigues ruidoses — some loud friends
(Some Spanish language sources, English language sources.)
Spanish gender-neutral ‑e has something in common with English singular they that makes me optimistic about it: a decent path through which it could become a standard and unremarkable part of the language. For they this has looked like:
* Existing long-standing use in someone and everyone constructions: someone lost their fork.
* Usage expands into more generic constructions: I hear you have a new lab partner; what are they like?
* Many non-binary people adopt it as their pronoun. People get practice referring to specific named individuals with it: Pat said they might be early.
* Usage expands into cases where the person's gender is not relevant: The person who gave me a ride home from the dance last night doesn't take care of their car.
* [prediction] Usage expands to where people use they unless they specifically want to emphasize gender.
Unlike the alternatives, amigos y amigas, amigxs, amig@s, and amig*s, gender-netural ‑e fits well with spoken Spanish. Reading articles from a while ago it seems strange to me that this wasn't seen as more of a priority before? Still there's now something of a path for it to enter the language as it's generally spoken:
* Existing long-standing use in words like estudiante, though still with gendered arti |
1c270cb2-e366-41f7-a42b-f56d8b22618b | trentmkelly/LessWrong-43k | LessWrong | How's the case for wearing googles for COVID-19 protection when in public transportation?
I'm in a high risk group and there's a higher likelihood of dying if I get infected then for the average person. I will reduce the amount of public transportation travel but being totally isolated won't be good for my health either.
I have a decent supply of surgical masks. How much percentage of the infection risk is do to having exposed eyes? Does it make sense to protect them with swimming googles? |
a5771cd5-f4f8-4337-ab96-4c34be62f097 | StampyAI/alignment-research-dataset/blogs | Blogs | Gender Bootstrappism
Gender Bootstrappism
--------------------
(I define "gender" to mean "the set of sociocultural characteristics historically associated with either sex")
(disclamer: this post may or may not be crippled by a terrible understanding of the topic at hand)
I'm not a Gender Abolitionist, I'm a Gender Bootstrappist. What I mean by this is that I think gender should become its own cultural notion, separate from sex, even though that's where it started historically. Gender should still be able to be partaken of by people, in whichever manner they want (including not at all). But, one important aspect, is that *if* gender as a broad cultural notion is to be preserved (for it is, after all, *[Very Culture](word-report-2.html)*), then there will be *social expectations*; I don't think that's avoidable, and I think a reasonable amount of social expectations can be had without falling into straight-up discrimination. Having expectations at all are what allow the *subversion* of expectations, and also expectations are pretty [humancore](word-report-2.html). |
b2818e7f-dd71-459b-a8fe-fc202159672b | LDJnr/LessWrong-Amplify-Instruct | LessWrong | "Game theory is the study of how rational actors interact to pursue incentives. It starts with the same questionable premises as economics: that everyone behaves rationally, that everyone is purely self-interested1, and that desires can be exactly quantified - and uses them to investigate situations of conflict and cooperation.Here we will begin with some fairly obvious points about decision trees, but by the end we will have the tools necessary to explain a somewhat surprising finding: that giving a US president the additional power of line-item veto may in many cases make the president less able to enact her policies. Starting at the beginning:The basic unit of game theory is the choice. Rational agents make choices in order to maximize their utility, which is sort of like a measure of how happy they are. In a one-person game, your choices affect yourself and maybe the natural environment, but nobody else. These are pretty simple to deal with:Here we visualize a choice as a branching tree. At each branch, we choose the option with higher utility; in this case, going to the beach. Since each outcome leads to new choices, sometimes the decision trees can be longer than this:Here's a slightly more difficult decision, denominated in money instead of utility. If you want to make as much money as possible, then your first choice - going to college or starting a minimum wage job right Now - seems to favor the more lucrative minimum wage job. But when you take Later into account, college opens up more lucrative future choices, as measured in the gray totals on the right-hand side. This illustrates the important principle of reasoning backward over decision trees. If you reason forward, taking the best option on the first choice and so on, you end up as a low-level manager. To get the real cash, you've got to start at the end - the total on the right - and then examine what choice at each branch will take you there.This is all about as obvious as, well, not hitting yourself on the head with a hammer, so let's move on to where it really gets interesting: two-player games.I'm playing White, and it's my move. For simplicity I consider only two options: queen takes knight and queen takes rook. The one chess book I've read values pieces in number of pawns: a knight is worth three pawns, a rook five, a queen nine. So at first glance, it looks like my best move is to take Black's rook. As for Black, I have arbitrarily singled out pawn takes pawn as her preferred move in the current position, but if I play queen takes rook, a new option opens up for her: bishop takes queen. Let's look at the decision tree:If I foolishly play this two player game the same way I played the one-player go-to-college game, I note that the middle branch has the highest utility for White, so I take the choice that leads there: capture the rook. And then Black plays bishop takes queen, and I am left wailing and gnashing my teeth. What did I do wrong?I should start by assuming Black will, whenever presented with a choice, take the option with the highest Black utility. Unless Black is stupid, I can cross out any branch that requires Black to play against her own interests. So now the tree looks like this:The two realistic options are me playing queen takes rook and ending up without a queen and -4 utility, or me playing queen takes knight and ending up with a modest gain of 2 utility.(my apologies if I've missed some obvious strategic possibility on this particular chessboard; I'm not so good at chess but hopefully the point of the example is clear.)This method of alternating moves in a branching tree matches both our intuitive thought processes during a chess game (“Okay, if I do this, then Black's going to do this, and then I'd do this, and then...”) and the foundation of some of the algorithms chess computers like Deep Blue use. In fact, it may seem pretty obvious, or even unnecessary. But it can be used to analyze some more complicated games with counterintuitive results.Art of Strategy describes a debate from 1990s US politics revolving around so-called “line-item veto” power, the ability to veto only one part of a bill. For example, if Congress passed a bill declaring March to be National Game Theory Month and April to be National Branching Tree Awareness Month, the President could veto only the part about April and leave March intact (as politics currently works, the President could only veto or accept the whole bill). During the '90s, President Clinton fought pretty hard for this power, which seems reasonable as it expands his options when dealing with the hostile Republican Congress.But Dixit and Nalebuff explain that gaining line-item veto powers might hurt a President. How? Branching trees can explain.Imagine Clinton and the Republican Congress are fighting over a budget. We can think of this as a game of sequential moves, much like chess. On its turn, Congress proposes a budget. On Clinton's turn, he either accepts or rejects the budget. A player “wins” if the budget contains their pet projects. In this game, we start with low domestic and military budgets. Clinton really wants to raise domestic spending (utility +10), and has a minor distaste for raised military spending (utility -5). Congress really wants to raise military spending (utility +10), but has a minor distaste for raised domestic spending (utility -5). The status quo is zero utility for both parties; if neither party can come to an agreement, voters get angry at them and they both lose 2 utility. Here's the tree when Clinton lacks line-item veto:For any particular Republican choice, Clinton will never respond in a way that does not maximize his utility, so the the Republicans reason backward and arrive at something like this:If Republicans are perfectly rational agents, they choose the second option, high domestic and high military spending, to give them their highest plausibly obtainable utility of 5.But what if Clinton has the line-item veto? Now his options look like this:If the Republicans stick to their previous choice of “high domestic and high military spending”, Clinton line-item vetoes the military spending, and we end up with a situation identical to the first choice: Clinton sitting on a pile of utility, and the Republicans wailing and gnashing their teeth. The Republicans need to come up with a new strategy, and their thought processes, based on Clinton as a utility-maximizer, look like this:Here Congress's highest utility choice is to propose low domestic spending (it doesn't matter if they give more money to the military or not as this will get line-item vetoed). Let's say they propose low domestic and low military spending, and Clinton accepts. The utilities are (0, 0), and now there is much wailing and gnashing of teeth on both sides (game theorists call this a gnash equilibrium. Maybe you've heard of it.)But now Clinton has a utility of 0, instead of a utility of 5. Giving him extra options has cost him utility! Why should this happen, and shouldn't he be able to avoid it?This happened because Clinton's new abilities affect not only his own choices, but those of his opponents (compare Schelling: Strategies of Conflict). He may be able to deal with this if he can make the Republicans trust him.In summary, simple sequential games can often be explored by reasoning backwards over decision trees representing the choices of the players involved. The next post will discuss simultaneous games and the concept of a Nash equilibrium.Footnotes:1: Game theory requires self-interest in that all players' are driven solely by their desire to maximize their own payoff in the game currently being played without regard to the welfare of other players or any external standard of fairness. However, it can also be used to describe the behavior of altruistic agents so long as their altruistic concerns are represented in the evaluation of their payoff."
The passage is from a user-contributed post on LessWrong.
Can you break this down into the top 5 most important points for me.
1. Game theory is a study of strategic interaction among rational players. It's based on the assumption that individuals act rationally and in their own self-interest to maximize their utility, or happiness. This concept is applied to both one-player and multi-player scenarios, with decision trees used to visualize the choices and potential outcomes.
2. The method of reasoning backward over decision trees is crucial in game theory. This means that to make the best decision, one should start from the end goal and work backwards to determine the best choice at each decision point. This method is used in chess and can be applied to more complex games or scenarios.
3. In two-player games, it's important to consider the other player's potential moves and their utility. This is demonstrated through the chess example, where the player must consider not only their own utility but also the opponent's potential responses and their utility.
4. Game theory can be applied to real-world scenarios, such as political decisions. The example of the "line-item veto" power illustrates how increasing options for one player (the President) can actually decrease their utility, due to the strategic responses of the other player (Congress).
5. The outcome of a game can be influenced by the players' abilities to trust each other and predict each other's actions. This is shown in the "line-item veto" example, where the President's utility could potentially be improved if the Republicans trust him. This highlights the importance of trust and predictability in strategic interactions.
Could you elaborate on how the concept of trust and predictability in strategic interactions can be quantified or incorporated into game theory models?
Incorporating trust and predictability into game theory models often involves the use of repeated games, Bayesian games, and signaling games.
1. Repeated Games: In these types of games, the same players interact with each other over multiple rounds. Trust can be built over time as players observe each other's actions and adjust their strategies accordingly. For example, in the iterated prisoner's dilemma, a player might choose to cooperate (an action that requires trust) if they observe that their opponent has cooperated in the past.
2. Bayesian Games: These games incorporate the concept of incomplete information, where players do not have perfect knowledge about other players' payoffs or strategies. Trust can be modeled as a player's belief about the likelihood that an opponent will take a certain action. These beliefs can be updated over time using Bayes' rule as more information becomes available.
3. Signaling Games: In these games, one player (the sender) sends a signal to another player (the receiver) about their type (e.g., their payoff function or strategy). The receiver then uses this signal to update their beliefs about the sender and decide on their action. Trust can be modeled as the receiver's belief in the truthfulness of the sender's signal.
Predictability can be incorporated into game theory models through the concept of equilibrium. In a Nash equilibrium, each player's strategy is the best response to the strategies of the other players, given that those players are also playing their best responses. This concept implies a certain level of predictability in players' actions, as no player has an incentive to deviate unilaterally from their strategy.
However, quantifying trust and predictability can be challenging due to their subjective nature. Different players may have different perceptions of trust and predictability, and these perceptions can change over time based on the players' experiences and observations.
How might the concepts of trust and predictability, as incorporated in game theory models, be applied to a real-world negotiation scenario, such as a peace treaty negotiation between two countries?
In a real-world negotiation scenario like a peace treaty negotiation between two countries, the concepts of trust and predictability play crucial roles and can be analyzed through the lens of game theory.
1. Repeated Interactions: Peace treaty negotiations often involve repeated interactions between the countries involved. Each round of negotiation allows the countries to observe each other's actions and adjust their strategies accordingly. For example, if Country A makes a concession in one round, and Country B reciprocates in the next round, trust is built. This is similar to the concept of repeated games in game theory.
2. Incomplete Information: In these negotiations, countries often do not have complete information about each other's intentions, capabilities, or bottom lines. They form beliefs about these unknowns and update these beliefs as more information becomes available. This is similar to the concept of Bayesian games. For example, if Country A reveals some of its demands or red lines, Country B can update its beliefs about Country A's intentions and adjust its negotiation strategy accordingly.
3. Signaling: Countries often send signals to each other during negotiations, either directly or indirectly. These signals can be about their intentions, their resolve, or their willingness to make concessions. The other country then uses these signals to update its beliefs and decide on its actions. This is similar to the concept of signaling games. For example, a public statement by Country A's leader about the country's commitment to peace can be seen as a signal to Country B.
4. Predictability and Equilibrium: In a negotiation, each country's strategy is shaped by its expectations of the other country's strategy. A stable outcome, or equilibrium, is reached when neither country has an incentive to deviate from its strategy given the other country's strategy. This concept of equilibrium brings a level of predictability to the negotiation process.
However, real-world negotiations are often more complex than game theory models, with many factors at play. Trust and predictability are important, but they are not the only considerations. Other factors, such as the countries' domestic politics, international pressures, and historical contexts, can also significantly influence the negotiation process and outcome. |
a24cc60d-b108-44f9-87cd-5de500d10692 | trentmkelly/LessWrong-43k | LessWrong | Rationality Activism: Open Thread
So as I read around the discussion section I keep coming across ideas (like this, this, this, or this) which all seem to be very related to the same topics:
* How do we use rationality to better contribute to the world?
* How do we spread rationality among other people?
Other posts talk about similar things.
I have been interested in this for a while now, and have gotten some great feedback.
But now I'm wondering how many groups, other than SIAI, are trying to do this. It seems like it would be silly to have something like this stagnate because of a simple coordination problem.
So if you are, please come forth and comment. If you're interested, do the same. Share what you know, learn from others, maybe maybe maybe get the ball rolling a bit more.
Note:
I changed the title (originally "Rationality Activism Groups") to reflect the more discussion-oriented nature of this thread. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.