Posts
Comments
Thanks for the post, I agree with the main points.
There is another claim on causality one could make, which would be: LLMs cannot reliably act in the world as robust agents since by acting in the world, you change the world, leading to a distributional shift from the correlational data the LLM encountered during training.
I think that argument is correct, but misses an obvious solution: once you let your LLM act in the world, simply let it predict and learn from the tokens that it receives in response. Then suddenly, the LLM does not model correlational, but actual causal relationships.
Agreed.
I think the most interesting part was that she made a comment that one way to predict a mind is to be a mind, and that that mind will not necessarily have the best of all of humanity as its goal. So she seems to take inner misalignment seriously.
40 min podcast with Anca Dragan who leads safety and alignment at google deepmind: https://youtu.be/ZXA2dmFxXmg?si=Tk0Hgh2RCCC0-C7q
To clarify: are you saying that since you perceive Chris Olah as mostly intrinsically caring about understanding neural networks (instead of mostly caring about alignment), you conclude that his work is irrelevant to alignment?
I can see that research into proof assistants might lead to better techniques for combining foundation models with RL. Is there anything more specific that you imagine? Outside of math there are very different problems because there is no easy to way to synthetically generate a lot of labeled data (as opposed to formally verifiable proofs).
Not much more specific! I guess from a certain level of capabilities onward, one could create labels with foundation models that evaluate reasoning steps. This is much more fuzzy than math, but I still guess a person who created a groundbreaking proof assistant would be extremely valuable for any effort that tries to make foundation models reason reliably. And if they’d work at a company like google, then I think their ideas would likely diffuse even if they didn’t want to work on foundation models.
Thanks for your details on how someone could act responsibly in this space! That makes sense. I think one caveat is that proof assistant research might need enormous amounts of compute, and so it’s unclear how to work on it productively outside of a company where the ideas would likely diffuse.
I think the main way that proof assistant research feeds into capabilies research is not through the assistants themselves, but by the transfer of the proof assistant research to creating foundation models with better reasoning capabilities. I think researching better proof assistants can shorten timelines.
- See also Demis' Hassabis recent tweet. Admittedly, it's unclear whether he refers to AlphaProof itself being accessible from Gemini, or the research into AlphaProof feeding into improvements of Gemini.
- See also an important paragraph in the blogpost for AlphaProof: "As part of our IMO work, we also experimented with a natural language reasoning system, built upon Gemini and our latest research to enable advanced problem-solving skills. This system doesn’t require the problems to be translated into a formal language and could be combined with other AI systems. We also tested this approach on this year’s IMO problems and the results showed great promise."
https://www.washingtonpost.com/opinions/2024/07/25/sam-altman-ai-democracy-authoritarianism-future/
Not sure if this was discussed at LW before. This is an opinion piece by Sam Altman, which sounds like a toned down version of "situational awareness" to me.
The news is not very old yet. Lots of potential for people to start freaking out.
One question: Do you think Chinchilla scaling laws are still correct today, or are they not? I would assume these scaling laws depend on the data set used in training, so that if OpenAI found/created a better data set, this might change scaling laws.
Do you agree with this, or do you think it's false?
https://x.com/sama/status/1813984927622549881
According to Sam Altman, GPT-4o mini is much better than text-davinci-003 was in 2022, but 100 times cheaper. In general, we see increasing competition to produce smaller-sized models with great performance (e.g., Claude Haiku and Sonnet, Gemini 1.5 Flash and Pro, maybe even the full-sized GPT-4o itself). I think this trend is worth discussing. Some comments (mostly just quick takes) and questions I'd like to have answers to:
- Should we expect this trend to continue? How much efficiency gains are still possible? Can we expect another 100x efficiency gain in the coming years? Andrej Karpathy expects that we might see a GPT-2 sized "smart" model.
- What's the technical driver behind these advancements? Andrej Karpathy thinks it is based on synthetic data: Larger models curate new, better training data for the next generation of small models. Might there also be architectural changes? Inference tricks? Which of these advancements can continue?
- Why are companies pushing into small models? I think in hindsight, this seems easy to answer, but I'm curious what others think: If you have a GPT-4 level model that is much, much cheaper, then you can sell the service to many more people and deeply integrate your model into lots of software on phones, computers, etc. I think this has many desirable effects for AI developers:
- Increase revenue, motivating investments into the next generation of LLMs
- Increase market-share. Some integrations are probably "sticky" such that if you're first, you secure revenue for a long time.
- Make many people "aware" of potential usecases of even smarter AI so that they're motivated to sign up for the next generation of more expensive AI.
- The company's inference compute is probably limited (especially for OpenAI, as the market leader) and not many people are convinced to pay a large amount for very intelligent models, meaning that all these reasons beat reasons to publish larger models instead or even additionally.
- What does all this mean for the next generation of large models?
- Should we expect that efficiency gains in small models translate into efficiency gains in large models, such that a future model with the cost of text-davinci-003 is massively more capable than today's SOTA? If Andrej Karpathy is right that the small model's capabilities come from synthetic data generated by larger, smart models, then it's unclear to me whether one can train SOTA models with these techniques, as this might require an even larger model to already exist.
- At what point does it become worthwhile for e.g. OpenAI to publish a next-gen model? Presumably, I'd guess you can still do a lot of "penetration of small model usecases" in the next 1-2 years, leading to massive revenue increases without necessarily releasing a next-gen model.
- Do the strategies differ for different companies? OpenAI is the clear market leader, so possibly they can penetrate the market further without first making a "bigger name for themselves". In contrast, I could imagine that for a company like Anthropic, it's much more important to get out a clear SOTA model that impresses people and makes them aware of Claude. I thus currently (weakly) expect Anthropic to more strongly push in the direction of SOTA than OpenAI.
I went to this event in 2022 and it was lovely. Will come again this year. I recommend coming!
Thanks for the answer!
But basically, by "simple goals" I mean "goals which are simple to represent", i.e. goals which have highly compressed representations
It seems to me you are using "compressed" in two very different meanings in part 1 and 2. Or, to be fairer, I interpret the meanings very differently.
I try to make my view of things more concrete to explain:
Compressed representations: A representation is a function from observations of the world state (or sequences of such observations) into a representation space of "features". That this is "compressed" means (a) that in , only a small number of features are active at any given time and (b) that this small number of features is still sufficient to predict/act in the world.
Goals building on compressed representations: A goal is a (maybe linear) function from the representation space into the real numbers. The goal "likes" some features and "dislikes" others. (Or if it is not entirely linear, then it may like/dislike some simple combinations/compositions of features)
It seems to me that in part 2 of your post, you view goals as compositions . Part 1 says that is highly compressed. But it's totally unclear to me why the composition should then have the simplicity properties you claim in part 2, which in my mind don't connect with the compression properties of as I just defined them.
A few more thoughts:
- The notion of "simplicity" in part seems to be about how easy it is to represent a function -- i.e., the space of parameters with which the function is represented is simple in your part 2.
- The notion of "compression" in part 1 seems to be about how easy it is to represent an input -- i.e., is there a small number of features such that their activation tells you the important things about the input?
- These notions of simplicity and compression are very different. Indeed, if you have a highly compressed representation as in part 1, I'd guess that necessarily lives in a highly complex space of possible functions with many parameters, thus the opposite of what seems to be going on in part 2.
This is largely my fault since I haven't really defined "representation" very clearly, but I would say that the representation of the concept of a dog should be considered to include e.g. the neurons representing "fur", "mouth", "nose", "barks", etc. Otherwise if we just count "dog" as being encoded in a single neuron, then every concept encoded in any neuron is equally simple, which doesn't seem like a useful definition.
(To put it another way: the representation is the information you need to actually do stuff with the concept.)
I'm confused. Most of the time, when seeing a dog, most of what I need is actually just to know that it is a "dog", so this is totally sufficient to do something with the concept. E.g., if I walk on the street and wonder "will this thing bark?", then knowing "my dog neuron activates" is almost enough.
I'm confused for a second reason: It seems like here you want to claim that the "dog" representation is NOT simple (since it contains "fur", "mouth", etc.). However, the "dog" representation needs lots of intelligence and should thus come along with compression, and if you equate compression and simplicity, then it seems to me like you're not consistent. (I feel a bit awkward saying "you're not consistent", but I think it's probably good if I state my honest state of mind at this moment).
To clarify my own position, in line with my definition of compression further above: I think that whether representation is simple/compressed is NOT a property of a single input-output relation (like "pixels of dog gets mapped to dog-neuron being activated"), but instead a property of the whole FUNCTION that maps inputs to representations. This function is compressed if for any given input, only a small number of neurons in the last layer activate, and if these can be used (ideally in a linear way) for further predictions and for evaluating goal-achievement.
I agree that most people who say they are hedonic utilitarians are not 100% committed to hedonic utilitarianism. But I still think it's very striking that they at least somewhat care about making hedonium. I claim this provides an intuition pump for how AIs might care about squiggles too.
Okay, I agree with this, fwiw. :) (Though I may not necessarily agree with claims about how this connects to the rest of the post)
Thanks for the post!
a. How exactly do 1 and 2 interact to produce 3?
I think the claim is along the lines of "highly compressed representations imply simple goals", but the connection between compressed representations and simple goals has not been argued, unless I missed it. There's also a chance that I simply misunderstand your post entirely.
b. I don't agree with the following argument:
Decomposability over space. A goal is decomposable over space if it can be evaluated separately in each given volume of space. All else equal, a goal is more decomposable if it's defined over smaller-scale subcomponents, so the most decomposable goals will be defined over very small slices of space—hence why we're talking about molecular squiggles. (By contrast, you can't evaluate the amount of higher-level goals like "freedom" or "justice" in a nanoscale volume, even in principle.)
The classical ML-algorithm that evaluates features separately in space is a CNN. That doesn't mean that features in CNNs look for tiny structures, though: The deeper into the CNN you are, the more complex the features get. Actually, deep CNNs are an example of what you describe in argument 1: The features in later layers of CNNs are highly compressed, and may tell you binary information such as "is there a dog", but they apply to large parts of the input image.
Therefore, I'd also expect that what an AGI would care about are ultimately larger-scale structures since the AGI's features will nontrivially depend on the interaction of larger parts in space (and possibly time, e.g. if the AGI evaluates music, movies, etc.).
c. I think this leaves the confusion why philosophers end up favoring the analog of squiggles when they become hedonic utilitarians. I'd argue that the premise may be false since it's unclear to me how what philosophers say they care about ("henonium") connects with what they actually care about (e.g., maybe they still listen to complex music, build a family, build status through philosophical argumentation, etc.)
You should all be using the "Google Scholar PDF reader extension" for Chrome.
Features I like:
- References are linked and clickable
- You get a table of contents
- You can move back after clicking a link with Alt+left
Screenshot:
I guess (but don't know) that most people who downvote Garrett's comment overupdated on intuitive explanations of singular learning theory, not realizing that entire books with novel and nontrivial mathematical theory have been written on it.
I do all of these except 3, and implementing a system like 3 is among my deprioritized things in my ToDo-list. Maybe I should prioritize it.
Yes the first! Thanks for the link!
I really enjoyed reading this post! It's quite well-written. Thanks for writing it.
The only critique is that I would have appreciated more details on how the linear regression parameters are trained and what exactly the projection is doing. John's thread is a bit clarifying on this.
One question: If you optimize the representation in the residual stream such that it corresponds to a particular chosen belief state, does the transformer than predict the next token as if in that belief state? I.e., does the transformer use the belief state for making predictions?
MATS mentorships are often weekly, but only for limited time, unlike PhD programs that offer mentorship for several years. These years are probably often necessary to develop good research taste.
(Fwiw, I don’t remember problems with stipend payout at seri mats in the winter program. I was a winter scholar 2022/23.)
This is very helpful, thanks! Actually, the post includes several sections, including in the appendix, that might be more interesting to many readers than the grant recommendations themselves. Maybe it would be good to change the title a bit so that people also expect other updates.
Thanks for the reply!
As I show in the examples in DSLT1, having degenerate Fisher information (i.e. degenerate Hessian at zeroes) comes in two essential flavours: having rank-deficiency, and having vanishing second-derivative (i.e. ). Precisely, suppose is the number of parameters, then you are in the regular case if can be expressed as a full-rank quadratic form near each singularity,
Anything less than this is a strictly singular case.
So if , then is a singularity but not a strict singularity, do you agree? It still feels like somewhat bad terminology to me, but maybe it's justified from the algebraic-geometry--perspective.
Zeta Functions in Singular Learning Theory
In this shortform, I very briefly explain my understanding of how zeta functions play a role in the derivation of the free energy in singular learning theory. This is entirely based on slide 14 of the SLT low 4 talk of the recent summit on SLT and Alignment, so feel free to ignore this shortform and simply watch the video.
The story is this: we have a prior , a model , and there is an unknown true distribution . For model selection, we are interested in the evidence of our model for a data set , which is given by
where is the empirical KL divergence. In fact, we are interested in selecting the model that maximizes the average of this quantity over all data sets. The average is then given by
where is the Kullback-Leibler divergence.
But now we have a problem: how do we compute this integral? Computing this integral is what the free energy formula is about.
The answer: by computing a different integral. So now, I'll explain the connection to different integrals we can draw.
Let
which is called the state density function. Here, is the Dirac delta function. For different , it measures the density of states (= parameter vectors) that have . It is thus a measure for the "size" of different level sets. This state density function is connected to two different things.
Laplace Transform to the Evidence
First of all, it is connected to the evidence above. Namely, let be the Laplace transform of . It is a function given by
In first step, we changed the order of integration, and in the second step we used the defining property of the Dirac delta. Great, so this tells us that ! So this means we essentially just need to understand .
Mellin Transform to the Zeta Function
But how do we compute ? By using another transform. Let be the Mellin transform of . It is a function (or maybe only defined on part of ?) given by
Again, we used a change in the order of integration and then the defining property of the Dirac delta. This is called a Zeta function.
What's this useful for?
The Mellin transform has an inverse. Thus, if we can compute the zeta function, we can also compute the original evidence as
Thus, we essentially changed our problem to the problem of studying the zeta function To compute the integral of the zeta function, it is then useful to perform blowups to resolve the singularities in the set of minima of , which is where algebraic geometry enters the picture. For more on all of this, I refer, again, to the excellent SLT low 4 talk of the recent summit on singular learning theory.
Thanks for the answer! I think my first question was confused because I didn't realize you were talking about local free energies instead of the global one :)
As discussed in the comment in your DSLT1 question, they are both singularities of since they are both critical points (local minima).
Oh, I actually may have missed that aspect of your answer back then. I'm confused by that: in algebraic geometry, the zero's of a set of polynomials are not necessarily already singularities. E.g., in , the zero set consists of the two axes, which form an algebraic variety, but only at is there a singularity because the derivative disappears.
Now, for the KL-divergence, the situation seems more extreme: The zero's are also, at the same time, the minima of , and thus, the derivative disappears at every point in the set . This suggests every point in is singular. Is this correct?
So far, I thought "being singular" means the effective number of parameters around the singularity is lower than the full number of parameters. Also, I thought that it's about the rank of the Hessian, not the vanishing of the derivative. Both perspectives contradict the interpretation in the preceding paragraph, which leaves me confused.
The uninteresting answer is that SLT doesn't care about the prior (other than its regularity conditions) since it is irrelevant in the limit.
I vaguely remember that there is a part in the MDL book by Grünwald where he explains how using a good prior such as Jeffrey's prior somewhat changes asymptotic behavior for , but I'm not certain of that.
Thanks also for this post! I enjoy reading the sequence and look forward to post 5 on the connections to alignment :)
At some critical value , we recognise a phase transition as being a discontinuous change in the free energy or one of its derivatives, for example the generalisation error .
"Discontinuity" might suggest that this happens fast. Yet, e.g. in work on grokking, it actually turns out that these "sudden changes" happen over a majority of the training time (often, the x-axis is on a logarithmic scale). Is this compatible, or would this suggest that phenomena like grokking aren't related to the phase transitions predicted by SLT?
There is, however, one fundamentally different kind of "phase transition" that we cannot explain easily with SLT: a phase transition of SGD in time, i.e. the number gradient descent steps. The Bayesian framework of SLT does not really allow one to speak of time - the closest quantity is the number of datapoints , but these are not equivalent. We leave this gap as one of the fundamental open questions of relating SLT to current deep learning practice.
As far as I know, modern transformers are often only trained once on each data sample, which should close the gap between SGD time and the number of data samples quite a bit. Do you agree with that perspective?
In general, it seems to me that we're probably most interested in phase transitions that happen across SGD time or with more data samples, whereas phase transitions related to other hyperparameters (for example, varying the truth as in your examples here) are maybe less crucial. Would you agree with that?
Would you expect that most phase transitions in SGD time or the number of data samples are first-order transitions (as is the case when there is a loss-complexity tradeoff), or can you conceive of second-order phase transitions that might be relevant in that context as well?
Which altered the posterior geometry, but not that of since (up to a normalisation factor).
I didn't understand this footnote.
but the node-degeneracy and orientation-reversing symmetries only occur under precise configurations of the truth.
Hhm, I thought that these symmetries are about configurations of the parameter vector, irrespective of whether it is the "true" vector or not.
Are you maybe trying to say the following? The truth determines which parameter vectors are preferred by the free energy, e.g. those close to the truth. For some truths, we will have more symmetries around the truth, and thus lower RLCT for regions preferred by the posterior.
We will use the label weight annihilation phase to refer to the configuration of nodes such that the weights all point into the centre region and annihilate one another.
It seems to me that in the other phase, the weights also annihilate each other, so the "non-weight annihilation phase" is a somewhat weird terminology. Or did I miss something?
The weight annihilation phase is never preferred by the posterior
I think there is a typo and you meant .
Thanks Liam also for this nice post! The explanations were quite clear.
The property of being singular is specific to a model class , regardless of the underlying truth.
This holds for singularities that come from symmetries where the model doesn't change. However, is it correct that we need the "underlying truth" to study symmetries that come from other degeneracies of the Fisher information matrix? After all, this matrix involves the true distribution in its definition. The same holds for the Hessian of the KL divergence.
Both configurations, non-weight-annihilation (left) and weight-annihilation (right)
What do you mean with non-weight-annihilation here? Don't the weights annihilate in both pictures?
In particular, it is the singularities of these minimum-loss sets — points at which the tangent is ill-defined — that determine generalization performance.
To clarify: there is not necessarily a problem with the tangent, right? E.g., the function has a singularity at because the second derivative vanishes there, but the tangent is define. I think for the same reason, some of the pictures may be misleading to some readers.
- A model, , parametrized by weights , where is compact;
Why do we want compactness? Neural networks are parameterized in a non-compact set. (Though I guess usually, if things go well, the weights don't blow up. So in that sense it can maybe be modeled to be compact)
The empirical Kullback-Leibler divergence is just a rescaled and shifted version of the negative log likelihood.
I think it is only shifted, and not also rescaled, if I'm not missing something.
But these predictions of "generalization error" are actually a contrived kind of theoretical device that isn't what we mean by "generalization error" in the typical ML setting.
Why is that? I.e., in what way is the generalization error different from what ML people care about? Because real ML models don't predict using an updated posterior over the parameter space? (I was just wondering if there is a different reason I'm missing)
Thanks for the answer mfar!
Yeah I remember also struggling to parse this statement when I first saw it. Liam answered but in case it's still not clear and/or someone doesn't want to follow up in Liam's thesis, is a free variable, and the condition is talking about linear dependence of functions of .
Consider a toy example (not a real model) to help spell out the mathematical structure involved: Let so that and . Then let and be functions such that and .. Then the set of functions is a linearly dependent set of functions because .
Thanks! Apparently the proof of the thing I was wondering about can be found in Lemma 3.4 in Liam's thesis. Also thanks for your other comments!
Thanks for the answer Liam! I especially liked the further context on the connection between Bayesian posteriors and SGD. Below a few more comments on some of your answers:
The partition function is equal to the model evidence , yep. It isn’t equal to (I assume is fixed here?) but is instead expressed in terms of the model likelihood and prior (and can simply be thought of as the “normalising constant” of the posterior),
and then under this supervised learning setup where we know , we have . Also note that this does “factor over ” (if I’m interpreting you correctly) since the data is independent and identically distributed.
I think I still disagree. I think everything in these formulas needs to be conditioned on the -part of the dataset. In particular, I think the notation is slightly misleading, but maybe I'm missing something here.
I'll walk you through my reasoning: When I write or , I mean the whole vectors, e.g., . Then I think the posterior compuation works as follows:
That is just Bayes rule, conditioned on in every term. Then, because from alone you don't get any new information about the conditional (A more formal way to see this is to write down the Bayesian network of the model and to see that and are d-separated). Also, conditioned on , is independent over data points, and so we obtain
So, comparing with your equations, we must have Do you think this is correct?
Btw., I still don't think this "factors over ". I think that
The reason is that old data points should inform the parameter , which should have an influence on future updates. I think the independence assumption only holds for the true distribution and the model conditioned on .
If you expand that term out you find that
because the second integral is the first central moment of a Gaussian. The derivative of the prior is irrelevant.
Right. that makes sense, thank you! (I think you missed a factor of , but that doesn't change the conclusion)
Thanks also for the corrected volume formula, it makes sense now :)
Thanks for this nice post! I fight it slightly more vague than the first post, but I guess that is hard to avoid when trying to distill highly technical topics. I got a lot out of it.
Fundamentally, we care about the free energy because it is a measure of posterior concentration, and as we showed with the BIC calculation in DSLT1, it tells us something about the information geometry of the posterior.
Can you tell more about why it is a measure of posterior concentration (It gets a bit clearer further below, but I state my question nonetheless to express that this statement isn't locally clear to me here)? I may lack some background in Bayesian statistics here. In the first post, you wrote the posterior as
and it seems like you want to say that if free energy is low, then the posterior is more concentrated. If I look at this formula, then low free energy corresponds to high , meaning the prior and likelihood have to "work quite a bit" to ensure that this expression overall integrates to . Are you claiming that most of that work happens very localized in a small parameter region?
Additionally, I am not quite sure what you mean with "it tells us something about the information geometry of the posterior", or even what you mean by "information geometry" here. I guess one answer is that you showed in post 1 that the Fisher information matrix appears in the formula for the free energy, which contains geometric information about the loss landscape. But then in the proof, you regarded that as a constant that you ignored in the final BIC formula, so I'm not sure if that's what you are referring to here. More explicit references would be useful to me.
Since there is a correspondence
we say the posterior prefers a region when it has low free energy relative to other regions of .
Note to other readers (as this wasn't clear to me immediately): That correspondence holds because one can show that
Here, is the global partition function.
The Bayes generalisation loss is then given by
I believe the first expression should be an expectation over .
It follows immediately that the generalisation loss of a region is
I didn't find a definition of the left expression.
So, the region in that minimises the free energy has the best accuracy-complexity tradeoff. This is the sense in which singular models obey Occam's Razor: if two regions are equally accurate, then they are preferred according to which is the simpler model.
Purposefully naive question: can I just choose a region that contains all singularities? Then it surely wins, but this doesn't help us because this region can be very large.
So I guess you also want to choose small regions. You hinted at that already by saying that should be compact. But now I of course wonder if sometimes just all of lies within a compact set.
There are two singularities in the set of true parameters,
which we will label as and respectively.
Possible correction: one of those points isn't a singularity, but a regular loss-minimizing point (as you also clarify further below).
Let's consider a one parameter model with KL divergence defined by
on the region with uniform prior
The prior seems to do some work here: if it doesn't properly support the region with low RLCT, then the posterior cannot converge there. I guess a similar story might a priori hold for SGD, where how you initialize your neural network might matter for convergence.
How do you think about this? What are sensible choices of priors (or network initializations) from the SLT perspective?
Also, I find it curious that in the second example, the posterior will converge to the lowest loss, but SGD would not since it wouldn't "manage to get out of the right valley", I assume. This seems to suggest that the Bayesian view of SGD can at most be true in high dimensions, but not for very low-dimensional neural networks. Would you agree with that, or what is your perspective?
Thank you for this wonderful article! I read it fairly carefully and have a number of questions and comments.
where the partition function (or in Bayesian terms the evidence) is given by
Should I think of this as being equal to , and would you call this quantity ? I was a bit confused since it seems like we're not interested in the data likelihood, but only the conditional data likelihood under model .
And to be clear: This does not factorize over because every data point informs and thereby the next data point, correct?
The learning goal is to find small regions of parameter space with high posterior density, and therefore low free energy.
But the free energy does not depend on the parameter, so how should I interpret this claim? Are you already one step ahead and thinking about the singular case where the loss landscape decomposes into different "phases" with their own free energy?
there is almost sure convergence as to a constant that doesn't depend on , [5]
I think the first expression should either be an expectation over , or have the conditional entropy within the parantheses.
- In the realisable case where , the KL divergence is just the euclidean distance between the model and the truth adjusted for the prior measure on inputs,
I briefly tried showing this and somehow failed. I didn't quite manage to get rid of the integral over . Is this simple? (You don't need to show me how it's done, but maybe mentioning the key idea could be useful)
A regular statistical model class is one which is identifiable (so implies that ), and has positive definite Fisher information matrix for all .
The rest of the article seems to mainly focus on the case of the Fisher information matrix. In particular, you didn't show an example of a non-regular model where the Fisher information matrix is positive definite everywhere.
Is it correct to assume models which are merely non-regular because the map from parameters to distributions is non-injective aren't that interesting, and so you maybe don't even want to call them singular? I found this slightly ambiguous, also because under your definitions further down, it seems like "singular" (degenerate Fisher information matrix) is a stronger condition then "strictly singular" (degenerate Fisher information matrix OR non-injective map from parameters to distributions).
It can be easily shown that, under the regression model, is degenerate if and only the set of derivatives
is linearly dependent.
What is in this formula? Is it fixed? Or do we average the derivatives over the input distribution?
Since every true parameter is a degenerate singularity[9] of , it cannot be approximated by a quadratic form.
Hhm, I thought having a singular model just means that some singularities are degenerate.
One unrelated conceptual question: when I see people draw singularities in the loss landscape, for example in Jesse's post, they often "look singular": i.e., the set of minimal points in the loss landscape crosses itself. However, this doesn't seem to actually be the case: a perfectly smooth curve of loss-minimizing points will consist of singularities because in the direction of the curve, the derivative does not change. Is this correct?
We can Taylor expand the NLL as
I think you forgot a in the term of degree 1.
In that case, the second term involving vanishes since it is the first central moment of a normal distribution
Could you explain why that is? I may have missed some assumption on or not paid attention to something.
In this case, since for all , we could simply throw out the free parameter and define a regular model with parameters that has identical geometry , and therefore defines the same input-output function, .
Hhm. Is the claim that if the loss of the function does not change along some curve in the parameter space, then the function itself remains invariant? Why is that?
Then the dimension arises as the scaling exponent of , which can be extracted via the following ratio of volumes formula for some :
This scaling exponent, it turns out, is the correct way to think about dimensionality of singularities.
Are you sure this is the correct formula? When I tried computing this by hand it resulted in , but maybe I made a mistake.
General unrelated question: is the following a good intuition for the correspondence of the volume with the effective number of parameters around a singularity? The larger the number of effective parameters around , the more blows up around in all directions because we get variation in all directions, and so the smaller the region where is below . So contributes to this volume. This is in fact what it does in the formulas, by being an exponent for small .
So, in this case the global RLCT is , which we will see in DSLT2 means that the posterior is most concentrated around the singularity .
Do you currently expect that gradient descent will do something similar, where the parameters will move toward singularities with low RLCT? What's the state of the theory regarding this? (If this is answered in later posts, feel free to just refer to them)
Also, I wonder whether this could be studied experimentally even if the theory is not yet ready: one could probably measure the RLCT around minimal loss points by measuring volumes, and then just check whether gradient descent actually ends up in low-RLCT regions. Maybe this is what you do in later posts. If this is the case, I wonder whether I should be surprised or not: it seems like the lower the RLCT, the larger the number of (fractional) directions where the loss is minimal, and so the larger the basin. So for purely statistical reasons, one may end up in such a region instead of isolated loss-minimizing points of high RLCT.
https://twitter.com/ai_risks/status/1664323278796898306?s=46&t=umU0Z29c0UEkNxkJx-0kaQ
Apparently Bill Gates signed.
Stating the obvious: Do we expect that Bill Gates will donate money to prevent the extinction from AI?
It's great to see Yoshua Bengio and other eminent AI scientists like Geoffrey Hinton actively engage in the discussion around AI alignment. He evidently put a lot of thought into this. There is a lot I agree with here.
Below, I'll discuss two points of disagreement or where I'm surprised by his takes, to highlight potential topics of discussion, e.g. if someone wants to engage directly with Bengio.
- Most of the post is focused on the outer alignment problem -- how do we specify a goal aligned with our intent -- and seems to ignore the inner alignment problem -- how do we ensure that the specified goal is optimized for.
- E.g., he makes an example of us telling the AI to fix climate change, after which the AI wipes out humanity since that fixes climate change more effectively than respecting our implicit constraints of which the AI has no knowledge. In fact, I think language models show that there may be quite some hope that AI models will understand our implicit intent. Under that view, the problem lies at least as much in ensuring that the AI cares.
- He also extensively discusses the wireheading problem of entities (e.g., humans, corporations, or AI systems) that try to maximize their reward signal. I think we have reasons to believe that wireheading isn't as much of a concern: inner misalignment will cause the agent to have some other goal than the precise maximization of the reward function, and once the agent is situationally aware, it has incentives to keep its goals from changing by gradient descent.
- He does discuss the fact that our brains reward us for pleasure and avoiding pain, which is misaligned with the evolutionary goal of genetic fitness. In the alignment community, this is most often discussed as an inner alignment issue between the "reward function" of evolution and the "trained agent" being our genomes. However, his discussion highlights that he seems to view it as an outer alignment issue between evolution and our reward signals in the brain, which shape our adult brains through in-lifetime learning. This is also the viewpoint in Brain-Like-AGI Safety, as far as I remember, and also seems related to viewpoints discussed in shard theory.
- "In fact, over two decades of work in AI safety suggests that it is difficult to obtain AI alignment [wikipedia], so not obtaining it is clearly possible."
- I agree with the conclusion, but I am surprised by the argument. It is true that we have seen over two decades of alignment research, but the alignment community has been fairly small all this time. I'm wondering what a much larger community could have done.
Yoshua Bengio was on David Krueger's PhD thesis committee, according to David's CV.
After filling out the form, I could click on "see previous responses", which allowed me to see the responses of all other people who have filled out the form so far.
That is probably not intended?
I disagree with this. I think the most useful definition of alignment is intent alignment. Humans are effectively intent-aligned on the goal to not kill all of humanity. They may still kill all of humanity, but that is not an alignment problem but a problem in capabilities: humans aren't capable of knowing which AI designs will be safe.
The same holds for intent-aligned AI systems that create unaligned successors.
Has this already been posted? I could not find the post.
For what it's worth, I think this comment seems clearly right to me, even if one thinks the post actually shows misalignment. I'm confused about the downvotes of this (5 net downvotes and 12 net disagree votes as of writing this).
Now to answer our big question from the previous section: I can find some satisfying the conditions exactly when all of the ’s are independent given the “perfectly redundant” information. In that case, I just set to be exactly the quantities conserved under the resampling process, i.e. the perfectly redundant information itself.
In the original post on redundant information, I didn't find a definition for the "quantities conserved under the resampling process". You name this F(X) in that post.
Just to be sure: is your claim that if F(X) exists that contains exactly the conserved quantities and nothing else, then you can define like this? Or is the claim even stronger and you think such can always be constructed?
Edit: Flagging that I now think this comment is confused. One can simply define as the conditional, which is a composition of the random variable and the function
When I converse with junior folks about what qualities they’re missing, they often focus on things like “not being smart enough” or “not being a genius” or “not having a PhD.” It’s interesting to notice differences between what junior folks think they’re missing & what mentors think they’re missing.
There may also be social reasons to give different answers depending on whether you are a mentor or mentee. I.e., answering "the better mentees were those who were smarter" seems like an uncomfortable thing to say, even if it's true.
(I do not want to say that this social explanation is the only reason that answers between mentors and mentees differed. But I do think that one should take it into account in one's models)
Thanks!
Then is a projection matrix, projecting into the span.
To clarify: for this, you probably need the basis to be orthonormal?
Summary
- Disagreements often focus on outputs even though underlying models produced those.
- Double Crux idea: focus on the models!
- Double Crux tries to reveal the different underlying beliefs coming from different perspectives on reality
- Good Faith Principle:
- Assume that the other side is moral and intelligent.
- Even if some actors are bad, you minimize the chance of error if you start with the prior that each new person is acting in good faith
- Identifying Cruxes
- For every belief A, there are usually beliefs B, C, D such that their believed truth supports belief A
- These are “cruxes” if them not being true would shake the belief in A.
- Ideally, B, C, and D are functional models of how the world works and can be empirically investigated
- If you know your crux(es), investigating it has the chance to change your belief in A
- For every belief A, there are usually beliefs B, C, D such that their believed truth supports belief A
- In Search of more productive disagreement
- Often, people obscure their cruxes by telling many supporting reasons, most of which aren’t their true crux.
- This makes it hard for the “opponent” to know where to focus
- If both parties search for truth instead of wanting to win, you can speed up the process a lot by telling each other the cruxes
- Often, people obscure their cruxes by telling many supporting reasons, most of which aren’t their true crux.
- Playing Double Crux
- Lower the bar: instead of reaching a shared belief, find a shared testable claim that, if investigated, would resolve the disagreement.
- Double Crux: A belief that is a crux for you and your conversation partner, i.e.:
- You believe A, the partner believes not A.
- You believe testable claim B, the partner believes not B.
- B is a crux of your belief in A and not B is a crux of your partner’s belief in not B.
- Investigating conclusively whether B is true may resolve the disagreement (if the cruxes were comprehensive enough)
- The Double Crux Algorithm
- Find a disagreement with another person (This might also be about different confidences in beliefs)
- Operationalize the disagreement (Avoid semantic confusions, be specific)
- Seek double cruxes (Seek cruxes independently and then compare)
- Resonate (Do the cruxes really feel crucial? Think of what would change if you believed your crux to be false)
- Repeat (Are there underlying easier-to-test cruxes for the double cruxes themselves?)
Summary
In this post, John starts with a very basic intuition: that abstractions are things you can get from many places in the world, which are therefore very redundant. Thus, for finding abstractions, you should first define redundant information: Concretely, for a system of n random variables X1, …, Xn, he defines the redundant information as that information that remains about the original after repeatedly resampling one variable at a time while keeping all the others fixed. Since there will not be any remaining information if n is finite, there is also the somewhat vague assumption that the number of variables goes to infinity in that resampling process.
The first main theorem says that this resampling process will not break the graphical structure of the original variables, i.e., if X1, …, Xn form a Markov random field or Bayesian network with respect to a graph, then the resampled variables will as well, even when conditioning on the abstraction of them. John’s interpretation is that you will still be able to make inferences about the world in a local way even if you condition on your high-level understanding (i.e., the information preserved by the resampling process)
The second main theorem applies this to show that any abstraction F(X1, …, Xn) that contains all the information remaining from the resampling process will also contain all the abstract summaries from the telephone theorem for all the ways that X1, …, Xn (with n going to infinity) could be decomposed into infinitely many nested Markov blankets. This makes F a supposedly quite powerful abstraction.
Further Thoughts
It’s fairly unclear how exactly the resampling process should be defined. If n is finite and fixed, then John writes that no information will remain. If, however, n is infinite from the start, then we should (probably?) expect the mutual information between the original random variable and the end result to also often be infinite, which also means that we should not expect a small abstract summary F.
Leaving that aside, it is in general not clear to me how F is obtained. The second theorem just assumes F and deduces that it contains the information from the abstract summaries of all telephone theorems. The hope is that F is low-dimensional and thus manageable. But no attempt is made to show the existence of a low-dimensional F in any realistic setting.
Another remark: I don’t quite understand what it means to resample one of the variables “in the physical world”. My understanding is as follows, and if anyone can correct it, that would be helpful: We have some “prior understanding” (= prior probability) about how the world works, and by measuring aspects in the world — e.g., patches full of atoms in a gear — we gain “data” from that prior probability distribution. When forgetting the data of one of the patches, we can look at the others and then use our physical understanding to predict the values for the lost patch. We then sample from that prediction.
Is that it? If so, then this resampling process seems very observer-dependent since there is probably no actual randomness in the universe. But if it is observer-dependent, then the resulting abstractions would also be observer-dependent, which seems to undermine the hope to obtain natural abstractions.
I also have a similar concern about the pencils example: if you have a prior on variables X1, …, Xn and you know that all of them will end up to be “objects of the same type”, and a joint sample of them gives you n pencils, then it makes sense to me that resampling them one by one until infinity will still give you a bunch of pencil-like objects, leading you to conclude that the underlying preserved information is a graphite core inside wood. However, where do the variables X1, …, Xn come from in the first place? Each Xi is already a high-level object and it is unclear to me what the analysis would look like if one reparameterized that space. (Thanks to Erik Jenner for mentioning that thought to me; there is a chance that I misrepresent his thinking, though.)
Summary
- Goal: Find motivation through truth-seeking rather than coercion or self-deception
- Ideally: the urges are aligned with the high-level goals
- Turn “wanting to want” into “want”
- If a person has simultaneously conflicting beliefs and desires, then one of those is wrong.
- [Comment from myself: I find this, as stated, not evidently true since desires often do not have a “ground truth” due to the orthogonality thesis. However, even if there is a conflict between subsystems, the productive way forward is usually to find a common path in a values handshake. This is how I interpret conflicting desires to be “wrong”]
- Understanding “shoulds”
- If you call some urges “lazy”, then you spend energy on a conflict
- If you ignore your urges, then part of you is not “focused” on the activity, making it less worthwhile
- Acknowledge your conflicting desires: “I have a belief that it’s good to run and I have a belief that it’s good to watch Netflix”
- The different parts aren’t right or wrong; they have tunnel vision, not seeing the value of the other desire
- Shoulds: When there is a default action, there is often a sense that you “should” have done something else. If you would have done this “something else”, then the default action becomes the “should” and the situation is reversed.
- View shoulds as “data” that is useful for making better conclusions
- The IDC Algorithm (with an example in the article)
- Recommendation: Do not tweak the structure of IDC before having tried it a few times
- Step 0: Find an internal disagreement
- Identify a “should” that’s counter to a default action
- Step 1: Draw two dots on a piece of paper and name them with the subagents representing the different positions
- Choose appropriate names/handles that don’t favor one side over the other
- Step 2: Decide who speaks first (it might be the side with more “urgency”)
- Say one thing embodied from that perspective
- Maybe use Focusing to check that the words resonate
- Step 3: Get the other side to acknowledge truth.
- Let it find something true in the statement or derived from it
- Step 4: The second side also adds “one thing”
- Be open in general about the means of communication of the sides; they may also scribble something, express a feeling, or …
- Step 5: Repeat
- Notes:
- It’s okay for some sides to “blow off steam” once in a while and not follow the rules; if so, correct that after the fact from a “moderation standpoint”
- You may write down “moderator interjections” with another color
- Eventually, you might realize the disagreement to be about something else.
- This can give clarity on the “internal generators” of conflict
- If so, start a new piece of paper with two new debaters
- Ideally, the different parts understand each other better, leading them to stop getting into conflict since they respect each other's values
Summary
- Focusing is a technique for bringing subconscious system 1 information into conscious awareness
- Felt sense: a feeling in the body that is not yet verbalized but may subconsciously influence behavior, and which carries meaning.
- The dominant factor in patient outcomes: does the patient remain uncertain, instead of having firm narratives?
- A goal of therapy is increased awareness and clarity. Thus, it is not useful to spend much time in the already known.
- The successful patient thinks and listens to information
- If the verbal part utters something, the patient will check with the felt senses to correct the utterance
- Listening can feel like “having something on the tip of your tongue”
- From felt senses to handles
- A felt sense is like a picture
- There’s lots of tacit, non-explicit information in it
- A handle is like a sketch of the picture that is true to it.
- Handles “resonate” with the felt sense
- The first attempt at a handle will often not resonate — then you need to iterate
- In the end, you might get a “click”, “release of pressure”, or “sense of deep rightness”
- The felt sense can change or disappear once “System 2 got the message”
- A felt sense is like a picture
- Advice and caveats
- The felt sense may also not be true — your system 1 may be biased.
- Tips:
- Choosing a topic: if you don’t have a felt sense to focus on, produce the utterance “Everything in my life is perfect right now” and see how system 1 responds. This will usually create a topic to focus on
- Get physically comfortable
- Don’t “focus” in the sense of effortful attention, but “focus” in the sense of “increase clarity”
- Hold space: don’t go super fast or “push”; silence in one’s mind is normal
- Stay with one felt sense at a time
- Always return to the felt sense, also if the coherent verbalized story feels “exciting”
- Don’t limit yourself to sensations in your body — there are other felt senses
- Try saying things out loud (both utterances and questions “to the felt sense”)
- Try to not “fall into” overwhelming felt senses; they can sometimes make the feeling a “subject” instead of an “object” to hold and talk with
- Going “meta” and asking what the body has to say about a felt sense can help with not getting sucked in
- Verbalizing “I feel rage” and then “something in me is feeling rage” etc. can progressively create distance to felt senses
- The Focusing Algorithm
- Select something to bring into focus
- Create space (get physically comfortable and drop in for a minute; Put attention to the body; ask sensations to wait if there are multiple; go meta if you’re overwhelmed)
- Look for a handle of the felt sense (Iterate between verbalizing and listening until the felt sense agrees; Ask questions to the felt sense; Take time to wait for responses)
Summary:
- Some things seem complicated/difficult and hard to do. You may also have uncertainties about whether you can achieve it
- E.g.: running a marathon; or improving motivation
- The resolve cycle: set a 5-minute timer and just solve the thing.
- Why does it work?
- You want to be actually trying, even when there is no immediate need.
- Resolve cycles are an easy way of achieving that (They make it more likely and less painful to invest effort)
- One mechanism of its success: when asking “Am I ready for this?” the answer is often “No”. But when there’s a five-minute timer, the answer becomes “Yes”, no matter how hard the problem itself may seem.
- You want to be actually trying, even when there is no immediate need.
- The Technique
- Choose a thing to solve — it can be small or big.
- Try solving it with a 5 minutes timer. Don’t defer parts of it to the future unless they are extremely easy (i.e., there is a TAP in place ensuring that they will get done)
- If it’s not solved: spend another 5 minutes brainstorming a list of five-minute actions to solve your problem
- Do one of the five-minute actions to get momentum
- Developing a “grimoire”:
- This section lists many questions to ask yourself in a resolve-cycle. These questions can help for generating useful ideas.
Note: I found this article in particular a bit hard to summarize, especially the section "The argument for CoZE". I find it hard to say what exactly it is telling me, and how it relates to the later sections.
Summary
- Comfort is a lack of pain, discomfort, negative emotions, fear, and anxiety, …
- Comfort often comes from experience
- There’s a gray area between comfort and discomfort that can be worth exploring
- Explore/Exploit Tradeoff:
- Should you exploit the current hill and climb higher there, or search for a new one?
- Problem: there is inherent uncertainty
- Exploration is risky for individuals; there is a strong bias toward known paths
- Argument for CoZE
- Try Things model — cheap, non-destabilizing experiments
- The text contains a list of questions that can generate a lot of the “things we might try”.
- Problem: We might be uncomfortable with them.
- How do we reason through them but also bring System 1 on board, which might have useful insights?
- Chesterton’s Fence
- In the story, someone destroys an ugly fence, only to then be attacked by an animal behind.
- Don’t destroy a barrier before you know exactly why it’s there.
- CoZE includes the lessons of Chesterton’s fence
- That’s why it’s called exploration instead of expansion
- This is the difference to exposure therapy
- When exploring the area around the fence, remain alert, attentive, receptive:
- Stay open to all outcomes: the fence shouldn’t be there, the fence is exactly where it should be, the fence should be further away or even closer to you…
- In the story, someone destroys an ugly fence, only to then be attacked by an animal behind.
- CoZE Algorithm:
- Choose an experience to explore (outside of the current action space, or somewhat blocked, maybe with a yum factor)
- Prepare to accept all worlds: both possibilities need to feel comfortable in your imagination
- Devise an experiment to “taste” the experience
- Try the experiment (Potentially with help of others)
- How do the body and mind react?
- How does the external world respond?
- Digest the experience
- Compare the experience to expectations
- Should you continue trying something like this? Do not force yourself
- Give your system 1 space
Summary
- Idea: use knowledge of how physiology influences the mind
- Mental shutdown:
- Stress → Mental shutdown (trouble thinking, making decisions, …)
- This leads to decisions that feel correct at the moment but are obviously flawed in hindsight
- Metacognitive failure: Part of what we lose is also our ability to notice the loss in abilities → Need objective “sobriety test”
- The automatic nervous system
- Sympathetic nervous system (SNS): accelerator; fight/flight/freeze, excitement
- Parasympathetic nervous system (PSNS): brakes; chill, open vulnerability, reflective…
- Relative arousal of these systems matters
- The qualities come bundled: not often do you have an “open, relaxed body posture” while also being extremely excited
- The ancestral need for survival guides our SNS-dominated responses, and they are not reflective
- Changing the state
- Algorithm idea:
- Check where you are on the SNS-PSNS spectrum
- Change your position at that spectrum “at will”
- Most of the skill lies in shifting toward PSNS, which is also more useful for our rationality
- Algorithm for moving toward PSNS:
- Notice that you are SNS-dominated (TAP: physical sign, perceived hostility, someone saying “calm down”)
- Open your Body posture
- Take low, slow, deep breaths: belly-dominated, exhale longer than inhale
- Get aware of the sensations in your feet
- Then expand that awareness to the whole body
- Take another low, slow, deep breath. Enjoy
- Algorithm idea:
- Metacognitive blindspots
- The againstness-resolving technique is one mechanism to resolve metacognitive blindspots (where our metacognition is not able to reflect on our cognition)
- General advice:
- if everyone sounds wrong/stupid/malicious etc., take seriously the idea that it’s actually you.
- Get external objective evidence on your blindspot (e.g., sobriety test for driving, noticing SNS-state)
- Assess your blindspots with the same tools you assess other’s blindspots
I share your confusion.