Thoughtdump on why I’m interested in computational mechanics:

one concrete application to natural abstractions from here: tl;dr, belief structures generally seem to be fractal shaped. one major part of natural abstractions is trying to find the correspondence between structures in the environment and concepts used by the mind. so if we can do the inverse of what adam and paul did, i.e. ‘discover’ fractal structures from activations and figure out what stochastic process they might correspond to in the environment, that would be cool

… but i was initially interested in reading compmech stuff not with a particular alignment relevant thread in mind but rather because it seemed broadly similar in directions to natural abstractions.

re: how my focus would differ from my impression of current compmech work done in academia: academia seems faaaaaar less focused on actually trying out epsilon reconstruction in real world noisy data. CSSR is an example of a reconstruction algorithm. apparently people did compmech stuff on real-world data, don’t know how good, but effort-wise far too less invested compared to theory work

would be interested in these reconstruction algorithms, eg what are the bottlenecks to scaling them up, etc.

tangent: epsilon transducers seem cool. if the reconstruction algorithm is good, a prototypical example i’m thinking of is something like: pick some input-output region within a model, and literally try to discover the hmm model reconstructing it? of course it’s gonna be unwieldly large. but, to shift the thread in the direction of bright-eyed theorizing …

the foundational Calculi of Emergence paper talked about the possibility of hierarchical epsilon machines, where you do epsilon machines on top of epsilon machines and for simple examples where you can analytically do this, you get wild things like coming up with more and more compact representations of stochastic processes (eg data stream → tree → markov model → stack automata → … ?)

this … sounds like natural abstractions in its wildest dreams? literally point at some raw datastream and automatically build hierarchical abstractions that get more compact as you go up

haha but alas, (almost) no development afaik since the original paper. seems cool

and also more tangentially, compmech seemed to have a lot to talk about providing interesting semantics to various information measures aka True Names, so another angle i was interested in was to learn about them.

eg crutchfield talks a lot about developing a right notion of information flow—obvious usefulness in eg formalizing boundaries?

many other information measures from compmech with suggestive semantics—cryptic order? gauge information? synchronization order? check ruro1 and ruro2 for more.

Epsilon machine (and MSP) construction is most likely computationally intractable [I don’t know an exact statement of such a result in the literature but I suspect it is true] for realistic scenarios.

Scaling an approximate version of epsilon reconstruction seems therefore of prime importance. Real world architectures and data has highly specific structure & symmetry that makes it different from completely generic HMMs. This must most likely be exploited.

The calculi of emergence paper has inspired many people but has not been developed much. Many of the details are somewhat obscure, vague. I also believe that most likely completely different methods are needed to push the program further. Computational Mechanics’ is primarily a theory of hidden markov models—it doesn’t have the tools to easily describe behaviour higher up the Chomsky hierarchy. I suspect more powerful and sophisticated algebraic, logical and categorical thinking will be needed here. I caveat this by saying that Paul Riechers has pointed out that actually one can understand all these gadgets up the Chomsky hierarchy as infinite HMMs which may be analyzed usefully just as finite HMMs.

The still-underdeveloped theory of epsilon transducers I regard as the most promising lens on agent foundations. This is uncharcted territory; I suspect the largest impact of computational mechanics will come from this direction.

Your point on True Names is well-taken. More basic examples than gauge information, synchronization order are the triple of quantites entropy rate h, excess entropy E and Crutchfield’s statistical/forecasting complexity C. These are the most important quantities to understand for any stochastic process (such as the structure of language and LLMs!)

Tl;dr, agents selected to perform robustly in various local interventional distributions must internally represent something isomorphic to a causal model of the variables upstream of utility, for it is capable of answering all causal queries for those variables.

Thm 1: agents achieving optimal policy (util max) across various local interventions must be able to answer causal queries for all variables upstream of the utility node

Thm 2: relaxation of above to nonoptimal policies, relating regret bounds to the accuracy of the reconstructed causal model

the proof is constructive—an algorithm that, when given access to regret-bounded-policy-oracle wrt an environment with some local intervention, queries them appropriately to construct a causal model

one implication is an algorithm for causal inference that converts black box agents to explicit causal models (because, y’know, agents like you and i are literally that aforementioned ‘regret-bounded-policy-oracle‘)

These selection theorems could be considered the converse of the well-known statement that given access to a causal model, one can find an optimal policy. (this and its relaxation to approximate causal models is stated in Thm 3)

Thm 1 / 2 is like a ‘causal good regulator‘ theorem.

gooder regulator theorem is not structural—as in, it gives conditions under which a model of the regulator must be isomorphic to the posterior of the system—a black box statement about the input-output behavior.

theorem is limited. only applies to cases where the decision node is not upstream of the environment nodes (eg classification. a negative example would be an mdp). but authors claim this is mostly for simpler proofs and they think this can be relaxed.

theorem is limited. only applies to cases where the decision node is not upstream of the environment nodes

I think you can drop this premise and modify the conclusion to “you can find a causal model for all variables upstream of the utility and not downstream of the decision.”

tl;dr, goal directedness of a policy wrt a utility function is measured by its min distance to one of the policies implied by the utility function, as per the intentional stance—that one should model a system as an agent insofar as doing so is useful.

Details

how is “policies implied by the utility function” operationalized? given a value u, we define a set containing policies of maximum entropy (of the decision variable, given its parents in the causal bayes net) among those policies that attain the utility u.

then union them over all the achievable values of u to get this “wide set of maxent policies,” and define goal directedness of a policy π wrt a utility function U as the maximum (negative) cross entropy between π and an element of the above set. (actually we get the same result if we quantify the min operation over just the set of maxent policies achieving the same utility as π.)

Intuition

intuitively, this is measuring: “how close is my policy π to being ‘deterministic,’ while ‘optimizing U at the competence level u(π)’ and not doing anything else ‘deliberately’?”

“close” / “deterministic” ~ large negative CE means small CE(π,πmaxent)=H(π)+KL(π||πmaxent)

“not doing anything else deliberately’” ~ because we’re quantifying over maxent policies. the policy is maximally uninformative/uncertain, the policy doesn’t take any ‘deliberate’ i.e. low entropy action, etc.

“at the competence level u(π)” ~ … under the constraint that it is identically competent to π

and you get the nice property of the measure being invariant to translation / scaling of U.

obviously so, because a policy is maxent among all policies achieving u on U iff that same policy is maxent among all policies achieving au+b on aU+b, so these two utilities have the same “wide set of maxent policies.”

Critiques

I find this measure problematic in many places, and am confused whether this is conceptually correct.

one property claimed is that the measure is maximum for uniquely optimal / anti-optimal policy.

it’s interesting that this measure of goal-directedness isn’t exactly an ~increasing function of u(π), and i think it makes sense. i want my measure of goal-directedness to, when evaluated relative to human values, return a large number for both aligned ASI and signflip ASI.

… except, going through the proof one finds that the latter property heavily relies on the “uniqueness” of the policy.

My policy can get the maximum goal-directedness measure if it is the only policy of its competence level while being very deterministic. It isn’t clear that this always holds for the optimal/anti-optimal policies or always relaxes smoothly to epsilon-optimal/anti-optimal policies.

Relatedly, the fact that the quantification is only happening over policies of the same competence level, which feels problematic.

minimum for uniformly random policy (this would’ve been a good property, but unless I’m mistaken I think the proof for the lower bound is incorrect, because negative cross entropy is not bounded below.)

honestly the maxent motivation isn’t super clear to me.

not causal. the reason you need causal interventions is because you want to rule out accidental agency/goal-directedness, like a rock that happens to be the perfect size to seal a water bottle—does your rock adapt when I intervene to change the size of the hole? discovering agents is excellent in this regards.

… except, going through the proof one finds that the latter property heavily relies on the “uniqueness” of the policy. My policy can get the maximum goal-directedness measure if it is the only policy of its competence level while being very deterministic. It isn’t clear that this always holds for the optimal/anti-optimal policies or always relaxes smoothly to epsilon-optimal/anti-optimal policies.

Yeah, uniqueness definitely doesn’t always hold for the optimal/anti-optimal policy. I think the way MEG works here makes sense: if you’re following the unique optimal policy for some utility function, that’s a lot of evidence for goal-directedness. If you’re following one of many optimal policies, that’s a bit less evidence—there’s a greater chance that it’s an accident. In the most extreme case (for the constant utility function) every policy is optimal—and we definitely don’t want to ascribe maximum goal-directedness to optimal policies there.

With regard to relaxing smoothly to epsilon-optimal/anti-optimal policies, from memory I think we do have the property that MEG is increasing in the utility of the policy for policies with greater than the utility of the uniform policy, and decreasing for policies with less than the utility of the uniform policy. I think you can prove this via the property that the set of maxent policies is (very nearly) just Boltzman policies with varying temperature. But I would have to sit down and think about it properly. I should probably add that to the paper if it’s the case.

minimum for uniformly random policy (this would’ve been a good property, but unless I’m mistaken I think the proof for the lower bound is incorrect, because negative cross entropy is not bounded below.)

Thanks for this. The proof is indeed nonsense, but I think the proposition is still true. I’ve corrected it to this.

Thanks for writing this up! Having not read the paper, I am wondering if in your opinion there’s a potential connection between this type of work and comp mech type of analysis/point of view? Even if it doesn’t fit in a concrete way right now, maybe there’s room to extend/modify things to combine things in a fruitful way? Any thoughts?

EDIT: I no longer think this setup is viable, for reasons that connect to why I think Critch’s operationalization is incomplete and why boundaries should ultimately be grounded in Pearlian Causality and interventions. Check update.

I believe there’s nothing much in the way of actually implementing an approximation of Critch’s boundaries^{[1]} using deep learning.

Recall, Critch’s boundaries are:

Given a world (markovian stochastic process) Wt, map its values W (vector) bijectively using f into ‘features’ that can be split into four vectors each representing a boundary-possessing system’s Viscera, Active Boundary, Passive Boundary, and Environment.

Then, we characterize boundary-ness (i.e. minimal information flow across features unmediated by a boundary) using two mutual information criterion each representing infiltration and exfiltration of information.

And a policy of the boundary-posessing system (under the ‘stance’ of viewing the world implied by f) can be viewed as a stochastic map (that has no infiltration/exfiltration by definition) that best approximates the true Wt dynamics.

The interpretation here (under low exfiltration and infiltration) is that f can be viewed as a policy taken by the system in order to perpetuate its boundary-ness into the future and continue being well-described as a boundary-posessing system.

All of this seems easily implementable using very basic techniques from deep learning!

Bijective feature map are implemented using two NN maps each way, with an autoencoder loss.

Mutual information is approximated with standard variational approximations. Optimize f to minimize it.

(the interpretation here being—we’re optimizing our ‘stance’ towards the world in a way that best views the world as a boundary-possessing system)

After you train your ‘stance’ using the above setup, learn the policy using an NN with standard SGD, with fixed f.

A very basic experiment would look something like:

Test the above setup on two cellular automata (e.g., GoL, Lenia, etc) systems, one containing just random ash, and the other some boundary-like structure like noise-resistant glider structures found via optimization (there are a lot of such examples in the Lenia literature).^{[2]}

Then (1) check if the infiltration/exfiltration values are lower for the latter system, and (2) do some interp to see if the V/A/P/E features or the learned policy NN have any interesting structures.

I’m not sure if I’d be working on this any time soon, but posting the idea here just in case people have feedback.

I think research on boundaries—both conceptual work and developing practical algorithms for approximating them & schemes involving them—are quite important for alignment for reasons discussed earlier in my shortform.

Ultimately we want our setup to detect boundaries that aren’t just physically contiguous chunks of matter, like informational boundaries, so we want to make sure our algorithm isn’t just always exploiting basic locality heuristics.

I can’t think of a good toy testbed (ideas appreciated!), but one easy thing to try is to just destroy all locality by mapping the automata lattice (which we were feeding as input) with the output of a complicated fixed bijective map over it, so that our system will have to learn locality if it turns out to be a useful notion in its attempt at viewing the system as a boundary.

I don’t see much hope in capturing a technical definition that doesn’t fall out of some sort of game theory, and even the latter won’t directly work for boundaries as representation of respect for autonomy helpful for alignment (as it needs to apply to radically weaker parties).

Boundaries seem more like a landmark feature of human-like preferences that serves as a test case for whether toy models of preference are reasonable. If a moral theory insists on tiling the universe with something, it fails the test. Imperative to merge all agents fails the test unless the agents end up essentially reconstructed. And with computronium, we’d need to look at the shape of things it’s computing rather than at the computing substrate.

I think it’s plausible that the general concept of boundaries can possibly be characterized somewhat independently of preferences, but at the same time have boundary-preservation be a quality that agents mostly satisfy (discussion here. very unsure about this). I see Critch’s definition as a first iteration of an operationalization for boundaries in the general, somewhat-preference-independent sense.

But I do agree that ultimately all of this should tie back to game theory. I find Discovering Agents most promising in this regards, though there are still a lot of problems—some of which I suspect might be easier to solve if we treat systems-with-high-boundaryness as a sort of primitive for the kind-of-thing that we can associate agency and preferences with in the first place.

There are two different points here, boundaries as a formulation of agency, and boundaries as a major component of human values (which might be somewhat sufficient by itself for some alignment purposes). In the first role, boundaries are an acausal norm that many agents end up adopting, so that it’s natural to consider a notion of agency that implies boundaries (after the agent had an opportunity for sufficient reflection). But this use of boundaries is probably open to arbitrary ruthlessness, it’s not respect for autonomy of someone the powers that be wouldn’t sufficiently care about. Instead, boundaries would be a convenient primitive for describing interactions with other live players, a Schelling concept shared by agents in this sense.

The second role as an aspect of values expresses that the agent does care about autonomy of others outside game theoretic considerations, so it only ties back to game theory by similarity, or through the story of formation of such values that involved game theory. A general definition might be useful here, if pointing AIs at it could instill it into their values. But technical definitions don’t seem to work when you consider what happens if you try to protect humanity’s autonomy using a boundary according to such definitions. It’s like machine translation, the problem could well be well-defined, but impossible to formally specify, other than by gesturing at a learning process.

I no longer think the setup above is viable, for reasons that connect to why I think Critch’s operationalization is incomplete and why boundaries should ultimately be grounded in Pearlian Causality and interventions.

(Note: I am thinking as I’m writing, so this might be a bit rambly.)

The world-trajectory distribution is ambiguous.

Intuition: Why does a robust glider in Lenia intuitively feel like a system possessing boundary? Well, I imagine various situations that happen in the world (like bullets) and this pattern mostly stays stable in face of them.

Now, notice that the measure of infiltration/exfiltration depends on ϕ∈Δ(Wω), a distribution over world history. Infil(ϕ):=Aggt≥0MutWω∼ϕ((Vt+1,At+1);Et∣(Vt,At,Pt))

So, for the above measure to capture my intuition, the approximate Markov condition (operationalized by low infil & exfil) must consider the world state Wω that contains the Lenia pattern with it avoiding bullets.

Remember, W is the raw world state, no coarse graining. So ϕ is the distribution over the raw world trajectory. It already captures all the “potentially occurring trajectories under which the system may take boundary-preserving-action.” Since everything is observed, our distribution already encodes all of “Nature’s Intervention.” So in some sense Critch’s definition is already causal (in a very trivial sense), by the virtue of requiring a distribution over the raw world trajectory, despite mentioning no Pearlian Causality.

Issue: Choice of ϕ

Maybe there is some canonical true ϕ for our physical world that minds can intersubjectively arrive at, so there’s no ambiguity.

But when I imagine trying to implement this scheme on Lenia, there’s immediately an ambiguity as to which distribution (representing my epistemic state on which raw world trajectories that will “actually happen”) we should choose:

Perhaps a very simple distribution: assigning uniform probability over world trajectories where the world contains nothing but the glider moving in a random direction with some initial point offset.

I suspect many stances other the one factorizing the world into gliders would have low infil/exfil, because the world is so simple. This is the case of “accidental boundary-ness.”

Perhaps something more complicated: various trajectories where e.g., the Lenia patterns encounters bullets, evolves with various other patterns, etc.

This I think rules out “accidental boundary-ness.”

I think the latter works. But now there’s a subjective choice of the distribution, and what are the set of possible/realistic “Nature’s Intervention”—all the situations that can ever be encountered by the system under which it has boundary-like behaviors—that we want to implicitly encode into our observational distribution. I don’t think it’s natural for ϕ assign much probability to a trajectory whose initial conditions are set in a very precise way such that everything decays into noise. But this feels quite subjective.

Hints toward a solution: Causality

I think the discussion above hints at a very crucial insight:

ϕ must arise as a consequence of the stable mechanisms in the world.

Suppose the world of Lenia contains various stable mechanisms like a gun that shoots bullets at random directions, scarce food sources, etc.

We want ϕ to describe distributions that the boundary system will “actually” experience in some sense. I want the “Lenia pattern dodges bullet” world trajectory to be considered, because there is a plausible mechanism in the world that can cause such trajectories to exist. For similar reasons, I think the empty world distributions are impoverished, and a distribution containing trajectories where the entire world decays into noise is bad because no mechanism can implement it.

Thus, unless you have a canonical choice of ϕ, a better starting point would be to consider the abstract causal model that encodes the stable mechanisms in the world, and using Discovering Agents-style interventional algorithms that operationalize the notion “boundaries causally separate environment and viscera.”

Well, because of everything mentioned above on how the causal model informs us on which trajectories are realistic, especially in the absence of a canonical ϕ. It’s also far more efficient, because the knowledge of the mechanism informs the algorithm of the precise interventions to query the world for, instead of having to implicitly bake them in ϕ.

There are still a lot more questions, but I think this is a pretty clarifying answer as to how Critch’s boundaries are limiting and why DA-style causal methods will be important.

I think the update makes sense in general, isn’t there however some way mutual information and causality is linked? Maybe it isn’t strong enough for there to be an easy extrapolation from one to the other.

Also I just wanted to drop this to see if you find it interesting, kind of on this topic? Im not sure its fully defined in a causality based way but it is about structure preservation.

Yeah I’d like to know if there’s a unified way of thinking about information theoretic quantities and causal quantities, though a quick literature search doesn’t show up anything interesting. My guess is that we’d want separate boundary metrics for informational separation and causal separation.

why bayesnets and markovnets? factorized cognition, how to do efficient bayesian updates in practice, it’s how our brain is probably organized, etc. why would anyone want to study this subject if they’re doing alignment research? explain philosophy behind them.

simple examples of bayes nets. basic factorization theorems (the I-map stuff and separation criterion)

tangent on why bayes nets aren’t causal nets, though Zack M Davis had a good post on this exact topic, comment threads there are high insight

how inference is basically marginalization (basic theorems of: a reduced markov net represents conditioning, thus inference upon conditioning is the same as marginalization on a reduced net)

why is marginalization hard? i.e. NP-completeness of exact and approximate inference worst-case what is a workaround? solve by hand simple cases in which inference can be greatly simplified by just shuffling in the order of sums and products, and realize that the exponential blowup of complexity is dependent on a graphical property of your bayesnet called the treewidth

exact inference algorithms (bounded by treewidth) that can exploit the graph structure and do inference efficiently: sum-product / belief-propagation

approximate inference algorithms (works in even high treewidth! no guarantee of convergence) - loopy belief propagation, variational methods, etc

connections to neuroscience: “the human brain is just doing belief propagation over a bayes net whose variables are the cortical column” or smth, i just know that there is some connection

Any thoughts on how to customize LessWrong to make it LessAddictive? I just really, really like the editor for various reasons, so I usually write a bunch (drafts, research notes, study notes, etc) using it but it’s quite easy to get distracted.

You could use the ad & content blocker uBlock Origin to zap any addictive elements of the site, like the main page feed or the Quick Takes or Popular Comments. Then if you do want to access these, you can temporarily turn off uBlock Origin.

Incidentally, uBlock Origin can also be installed on mobile Firefox, and you can manually sync its settings across devices.

moments of microscopic fun encountered while studying/researching:

Quantum mechanics call vector space & its dual bra/ket because … bra-c-ket. What can I say? I like it—But where did the letter ‘c’ go, Dirac?

Defining cauchy sequences and limits in real analysis: it’s really cool how you “bootstrap” the definition of Cauchy sequences / limit on real using the definition of Cauchy sequences / limit on rationals. basically:

(1) define Cauchy sequence on rationals

(2) use it to define limit (on rationals) using rational-Cauchy

(3) use it to define reals

(4) use it to define Cauchy sequence on reals

(5) show it’s consistent with Cauchy sequence on rationals in both directions

a. rationals are embedded in reals hence the real-Cauchy definition subsumes rational-Cauchy definition

b. you can always find a rational number smaller than a given real number hence a sequence being rational-Cauchy means it is also real-Cauchy)

(6) define limit (on reals)

(7) show it’s consistent with limit on rationals

(8) … and that they’re equivalent to real-Cauchy

(9) proceed to ignore the distinction b/w real-Cauchy/limit and their rational counterpart. Slick!

Maybe he dropped the “c” because it changes the “a” phoneme from æ to ɑː and gives a cleaner division in sounds: “brac-ket” pronounced together collides with “bracket” where “braa-ket” does not.

Any advice on reducing neck and shoulder pain while studying? For me that’s my biggest blocker to being able to focus longer (especially for math, where I have to look down at my notes/book for a long period of time). I’m considering stuff like getting a standing desk or doing regular back/shoulder exercises. Would like to hear what everyone else’s setups are.

Train skill of noticing tension and focus on it. Tends to dissolve. No that’s not so satisfying but it works. Standing desk can help but it’s just not that comfortable for most.

I still have lots of neck and shoulder tension, but the only thing I’ve found that can reliably lessen it is doing some hard work on a punching bag for about 20 minutes every day, especially hard straights and jabs with full extension.

(Quality: Low, only read when you have nothing better to do—also not much citing)

30-minute high-LLM-temp stream-of-consciousness on “How do we make mechanistic interpretability work for non-transformers, or just any architectures?”

We want a general way to reverse engineer circuits

e.g., Should be able to rediscover properties we discovered from transformers

Concrete Example: we spent a bunch of effort reverse engineering transformer-type architectures—then boom, suddenly some parallel-GPU-friendly-LSTM architecutre turns out to have better scaling properties, and everyone starts using it. LSTMs have different inductive biases, like things in the same layer being able to communicate multiple times with each other (unlike transformers), which incentivizes e.g., reusing components (more search-y?).

Formalize:

You have task X. You train a model A with inductive bias I_A. You also train a model B with inductive bias I_B. Your mechanistic interpretability techniques work well on deciphering A, but not B. You want your mechanistic interpretability techniques to work well for B, too.

Proposal: Communication channel

Train a Transformer on task X

Existing Mechanistic interpretability work does well on interpreting this architecture

Somehow stitch the LSTM to the transformer (?)

I’m trying to get at to the idea of “interface conversion,” that by the virtue of SGD being greedy, it will try to convert the outputs of transformer-friendly types

Now you can better understand the intermediate outputs of the LSTM by just running mechanistic interpretability on the transformer layers whose input are from the LSTM

(I don’t know if I’m making any sense here, my LLM temp is > 1)

Proposal: approximation via large models?

Train a larger transformer architecture to approximate the smaller LSTM model (either just input output pairs, or intermediate features, or intermediate features across multiple time-steps, etc):

the basic idea is that a smaller model would be more subject to following its natural gradient shaped by the inductive bias, while larger model (with direct access to the intermediate outputs of the smaller model) would be able to approximate it despite not having as much inductive bias incentive towards it.

probably false but illustrative example: Train small LSTM on chess. By the virtue of being able to run serial computation on same layers, it focuses on algorithms that have repeating modular parts. In contrast, a small Transformer would learn algorithms that don’t have such repeating modular parts. But instead, train a large transformer to “approximate” the small LSTM—it should be able to do so by, e.g., inefficiently having identical modules across multiple layers. Now use mechanistic interpretability on that.

Proposal: redirect GPS?

Thane’s value formation picture says GPS should be incentivized to reverse-engineer the heuristics because it has access to inter-heuristic communication channel. Maybe, in the middle of training, gradually swap different parts of the model with those that have different inductive biases, see GPS gradually learn to reverse-engineer those, and mechanistically-interpret how GPS exactly does that, and reimplement in human code?

Proposal: Interpretability techniques based on behavioral constraints

e.g., Discovering Latent Knowledge without Supervision, putting constraints?

How to do we “back out” inductive biases, just given e.g., architecture, training setup? What is the type signature?

I am curious as to how often the asymptotic results proven using features of the problem that seem basically practically-irrelevant become relevant in practice.

Like, I understand that there are many asymptotic results (e.g., free energy principle in SLT) that are useful in practice, but i feel like there’s something sus about similar results from information theory or complexity theory where the way in which they prove certain bounds (or inclusion relationship, for complexity theory) seem totally detached from practicality?

joint source coding theorem is often stated as why we can consider the problem of compression and redundancy separately, but when you actually look at the proof it only talks about possibility (which is proven in terms of insanely long codes) and thus not-at-all trivial that this equivalence is something that holds in the context of practical code-engineering

complexity theory talks about stuff like quantifying some property over all possible boolean circuits of a given size which seems to me considering a feature of the problem just so utterly irrelevant to real programs that I’m suspicious it can say meaningful things about stuff we see in practice

as an aside, does the P vs NP distinction even matter in practice? we just … seem to have very good approximation to NP problems by algorithms that take into account the structures specific to the problem and domains where we want things to be fast; and as long as complexity methods doesn’t take into account those fine structures that are specific to a problem, i don’t see how it would characterize such well-approximated problems using complexity classes.

Wigderson’s book had a short section on average complexity which I hoped would be this kind of a result, and I’m unimpressed (the problem doesn’t sound easier—now how do you specify the natural distribution??)

One result to mention in computational complexity is the PCP theorem which not only gives probabilistically checkable proofs but also gives approximation case hardness. Seems deep but I haven’t understood the proof yet.

Great question. I don’t have a satisfying answer. Perhaps a cynical answer is survival bias—we remember the asymptotic results that eventually become relevant (because people develop practical algorithms or a deeper theory is discovered) but don’t remember the irrelevant ones.

Existence results are categorically easier to prove than explicit algorithms. Indeed, classical existence may hold (the former) while intuitioinistically (the latter) might not. We would expect non-explicit existence results to appear before explicit algorithms.

One minor remark on ‘quantifying over all boolean algorithms’. Unease with quantification over large domains may be a vestige of set-theoretic thinking that imagines types as (platonic) boxes. But a term of a for-all quantifier is better thought of as an algorithm/ method to check the property for any given term (in this case a Boolean circuit). This doesn’t sound divorced from practice to my ears.

as an aside, does the P vs NP distinction even matter in practice?

Yes, it does, for several reasons:

It basically is necessary to prove P != NP to get a lot of other results to work, and for some of those results, proving P != NP is sufficient.

If P != NP (As most people suspect), it fundamentally rules out solving lots of problems generally and quickly without exploiting structure, and in particular lets me flip the burden of proof to the algorithm maker to explain why their solution to a problem like SAT is efficient, rather than me having to disprove the existence of an efficient algorithm.

It’s either by exploiting structure, somehow having a proof that P=NP, or relying on new physics models that enable computing NP-complete problems efficiently, and the latter 2 need very, very strong evidence behind them.

This in particular applies to basically all learning problems in AI today.

It explains why certain problems cannot be reasonably solved optimally, without huge discoveries, and the best examples are travelling salesman problems for inability to optimally solve, as well as a whole lot of other NP-complete problems. There are also other NP problems where there isn’t a way to solve them efficiently at all, especially if FPT != W[1] holds.

Also a note that we also expect a lot of NP-complete problems to also not be solvable by fast algorithms even in the average case, which basically means it’s likely to be very relevant quite a lot of the time, so we don’t have to limit ourselves to the worst case either.

I recently learned about metauni, and it looks amazing. TL;DR, a bunch of researchers give out lectures or seminars on Roblox—Topics include AI alignment/policy, Natural Abstractions, Topos Theory, Singular Learning Theory, etc.

I haven’t actually participated in any of their live events yet and only watched their videos, but they all look really interesting. I’m somewhat surprised that there hasn’t been much discussion about this on LW!

Discovering agents provide a genuine causal, interventionist account of agency and an algorithm to detect them, motivated by the intentional stance. I find this paper very enlightening from a conceptual perspective!

I’ve tried to think of problems that needed to be solved before we can actually implement this on real systems—both conceptual and practical—on approximate order of importance.

There are no ‘dynamics,’ no learning. As soon as a mechanism node is edited, it is assumed that agents immediately change their ‘object decision variable’ (a conditional probability distribution given its object parent nodes) to play the subgame equilibria.

Assumption of factorization of variables into ‘object’ / ‘mechanisms,’ and the resulting subjectivity. The paper models the process by which an agent adapts its policy given changes in the mechanism of the environment via a ‘mechanism decision variable’ (that depends on its mechanism parent nodes), which modulates the conditional probability distribution of its child ‘object decision variable’, the actual policy.

For example, the paper says a learned RL policy isn’t an agent, because interventions in the environment won’t make it change its already-learned policy—but that a human or a RL policy together with its training process is an agent, because it can adapt. Is this reasonable?

Say I have a gridworld RL policy that’s learned to get cheese (3 cell world, cheese always on left) by always going to the left. Clearly it can’t change its policy when I change the cheese distribution to favor right, so it seems right to call this not an agent.

Now, say the policy now has sensory access to the grid state, and correctly generalized (despite only being trained on left-cheese) to move in the direction where it sees the cheese, so when I change the cheese distribution, it adapts accordingly. I think it is right to call this an agent?

Now, say the policy is an LLM agent (static weight) on an open world simulation which reasons in-context. I just changed the mechanism of the simulation by lowering the gravity constant, and the agent observes this, reasons in-context, and adapts its sensorimotor policy accordingly. This is clearly an agent?

I think this is because the paper considers, in the case of the RL policy alone, the ‘object policy’ to be the policy of the trained neural network (whose induced policy distribution is definitionally fixed), and the ‘mechanism policy’ to be a trivial delta function assigning the already-trained object policy. And in the case of the RL policy together with its training process, the ‘mechanism policy’ is now defined as the training process that assigns the fully-trained conditional probability distribution to the object policy.

But what if the ‘mechanism policy’ was the in-context learning process by which it induces an ‘object policy’? Then changes in the environment’s mechanism can be related to the ‘mechanism policy’ and thus the ‘object policy’ via in-context learning as in the second and third example, making them count as agents.

Ultimately, the setup in the paper forces us to factorize the means-by-which-policies-adapt into mechanism vs object variables, and the results (like whether a system is to be considered an agent) depends on this factorization. It’s not always clear what the right factorization is, how to discover them from data, or if this is the right frame to think about the problem at all.

Implicit choice of variables that are convenient for agent discovery. The paper does mention that the algorithm is dependent in the choice of the variable, as in: if the node corresponding to the ‘actual agent decision’ is missing but its children is there, then the algorithm will label its children to be the decision nodes. But this is already a very convenient representation!

Prototypical example: Minecraft world with RL agents interacting represented as a coarse-grained lattice (dynamical Bayes Net?) with each node corresponding to a physical location and its property, like color. Clearly no single node here is an agent, because agents move! My naive guess is that in principle, everything will be labeled an agent.

So the variables of choice must be abstract variables of the underlying substrate, like functions over them. But then, how do you discover the right representation automatically, in a way that interventions in the abstract variable level can faithfully translate to actually performable interventions in the underlying substrate?

Given the causal graph, even the slightest satisfaction of the agency-criterion labels the nodes as decision / utility. No “degree-of-agency”—maybe by summing over the extent to which the independencies fail to satisfy?

Then different agents are defined as causally separated chunks (~connected component) of [set-of-decision-nodes / set-of-utility-nodes]. How do we accommodate hierarchical agency (like subagents), systems with different degrees of agency, etc?

The interventional distribution on the object/mechanism variables are converted into a causal graph using the obvious [perform-do()-while-fixing-everything-else] algorithm. My impression is that causal discovery doesn’t really work in practice, especially in noisy reality with a large number of variables via gazillion conditional independence tests.

The correctness proof requires lots of unrealistic assumptions, e.g., agents always play subgame equilibria, though I think some of this can be relaxed.

Complaint with Pugh’s real analysis textbook: He doesn’t even define the limit of a function properly?!

It’s implicitly defined together with the definition of continuity where ∀ϵ>0∃δ>0|x−x0|<δ⟹|f(x)−f(x0)|<ϵ, but in Chapter 3 when defining differentiability he implicitly switches the condition to 0<|x−x0|<δ without even mentioning it (nor the requirement that x0 now needs to be an accumulation point!) While Pugh has its own benefits, coming from Terry Tao’s analysis textbook background, this is absurd!

(though to be fair Terry Tao has the exact same issue in Book 2, where his definition of function continuity via limit in metric space precedes that of defining limit in general … the only redeeming factor is that it’s defined rigorously in Book 1, in the limited context of R)

*sigh* I guess we’re still pretty far from reaching the Pareto Frontier of textbook quality, at least in real analysis.

… Speaking of Pareto Frontiers, would anyone say there is such a textbook that is close to that frontier, at least in a different subject? Would love to read one of those.

Maybe you should email Pugh with the feedback? (I audited his honors analysis course in fall 2017; he seemed nice.)

As far as the frontier of analysis textbooks goes, I really like how Schröder Mathematical Analysis manages to be both rigorous and friendly: the early chapters patiently explain standard proof techniques (like the add-and-subtract triangle inequality gambit) to the novice who hasn’t seen them before, but the punishing details of the subject are in no way simplified. (One wonders if the subtitle “A Concise Introduction” was intended ironically.)

I used to try out near-random search on ideaspace, where I made a quick app that spat out 3~5 random words from a dictionary of interesting words/concepts that I curated, and I spent 5 minutes every day thinking very hard on whether anything interesting came out of those combinations.

Of course I knew random search on exponential space was futile, but I got a couple cool invention ideas (most of which turned out to already exist), like:

infinite indoor rockclimbing: attach rocks to a vertical treadmill, and now you have an infinite indoor rock climbing wall (which is also safe from falling)! maybe add some fancy mechanism to add variations to the rocks + a VR headgear, I guess.

clever crypto mechanism design (in the spirit of CO2 Coin) to incentivize crowdsourcing of age-reduction molecule design animal trials from the public. (I know what you’re thinking)

You can probably do this smarter now if you wanted, with eg better GPT models.

Having lived ~19 years, I can distinctly remember around 5~6 times when I explicitly noticed myself experiencing totally new qualia with my inner monologue going “oh wow! I didn’t know this dimension of qualia was a thing.” examples:

hard-to-explain sense that my mind is expanding horizontally with fractal cube-like structures (think bismuth) forming around it and my subjective experience gliding along its surface which lasted for ~5 minutes after taking zolpidem for the first time to sleep (2 days ago)

getting drunk for the first time (half a year ago)

feeling absolutely euphoric after having a cool math insight (a year ago)

...

Reminds me of myself around a decade ago, completely incapable of understanding why my uncle smoked, being “huh? The smoke isn’t even sweet, why would you want to do that?” Now that I have [addiction-to-X] as a clear dimension of qualia/experience solidified in myself, I can better model their subjective experiences although I’ve never smoked myself. Reminds me of the SSC classic.

Also one observation is that it feels like the rate at which I acquire these is getting faster, probably because of increase in self-awareness + increased option space as I reach adulthood (like being able to drink).

Anyways, I think it’s really cool, and can’t wait for more.

Sunlight scattered by the atmosphere on cloudless mornings during the hour before sunrise inspires a subtle feeling (“this is cool, maybe even exciting”) that I never noticed till I started intentionally exposing myself to it for health reasons (specifically, making it easier to fall asleep 18 hours later).

More precisely, I might or might not have noticed the feeling, but if I did notice it, I quickly forgot about it because I had no idea how to reproduce it.

I have to get away from artificial light (streetlamps) (and from direct (yellow) sunlight) for the (blue) indirect sunlight to have this effect. Also, it is no good looking at a small patch of sky, e.g., through a window in a building: most or all of the upper half of my field of vision must be receiving this indirect sunlight. (The intrinsically-photosensitive retinal ganglion cells are all over the bottom half of the retina, but absent from the top half.)

To me, the fact that the human brain basically implements SSL+RL is very very strong evidence that the current DL paradigm (with a bit of “engineering” effort, but nothing like fundamental breakthroughs) will kinda just keep scaling until we reach point-of-no-return. Does this broadly look correct to people here? Would really appreciate other perspectives.

I mostly think “algorithms that involve both SSL and RL” is a much broader space of possible algorithms than you seem to think it is, and thus that there are parts of this broad space that require “fundamental breakthroughs” to access. For example, both AlexNet and differentiable rendering can be used to analyze images via supervised learning with gradient descent. But those two algorithms are very very different from each other! So there’s more to an algorithm than its update rule.

See also 2nd section of this comment, although I was emphasizing alignment-relevant differences there whereas you’re talking about capabilities. Other things include the fact that if I ask you to solve a hard math problem, your brain will be different (different weights, not just different activations / context) when you’re halfway through compared to when you started working on it (a.k.a. online learning, see also here), and the fact that brain neural networks are not really “deep” in the DL sense. Among other things.

Makes sense. I think we’re using the terms differently in scope. By “DL paradigm” I meant to encompass the kind of stuff you mentioned (RL-directing-SS-target (active learning), online learning, different architecture, etc) because they really seemed like “engineering challenges” to me (despite them covering a broad space of algorithms) in the sense that capabilities researchers already seem to be working on & scaling them without facing any apparent blockers to further progress, i.e. in need of any “fundamental breakthroughs”—by which I was pointing more at paradigm shifts away from DL like, idk, symbolic learning.

I have a slightly different takeaway. Yes techniques similar to current techniques will most likely lead to AGI but it’s not literally ‘just scaling LLMs’. The actual architecture of the brain is meaningfully different from what’s being deployed right now. So different in one sense. On the other hand it’s not like the brain does something completely different and proposals that are much closer to the brain architecture are in the literature (I won’t name them here...). It’s plausible that some variant on that will lead to true AGI. Pure hardware scaling obviously increases capabilities in a straightforward way but a transformer is not a generally intelligent agent and won’t be even if scaled many more OOMs.

(I think Steven Byrnes has a similar view but I wouldn’t want to misrepresent his views)

a transformer is not a generally intelligent agent and won’t be even if scaled many more OOMs

So far as I can tell, a transformer has three possible blockers (that would need to stand undefeated together): (1) in-context learning plateauing at a level where it’s not able to do even a little bit of useful work without changing model weights, (2) terrible sample efficiency that asks for more data than is available on new or rare/situational topics, and (3) absence of a synthetic data generation process that’s both sufficiently prolific and known not to be useless at that scale.

A need for online learning and terrible sample efficiency are defeated by OOMs if enough useful synthetic data can be generated, which the anemic in-context learning without changing weights might turn out to be sufficient for. This is the case of defeating (3), with others falling as a result.

Another possibility is that much larger multimodal transformers (there is a lot of video) might suffice without synthetic data if a model learns superintelligent in-context learning. SSL is not just about imitating humans, the problems it potentially becomes adept at solving are arbitrarily intricate. So even if it can’t grow further and learn substantially new things within its current architecture/model, it might happen to already be far enough along at inference time to do the necessary redesign on its own. This is the case of defeating (1), leaving it to the model to defeat the others. And it should help with (3) even at non-superintelligent levels.

Failing that, RL demonstrates human level sample efficiency in increasingly non-toy settings, promising that saner amounts of useful synthetic data might suffice, defeating (2), though at this point it’s substantially not-a-transformer.

generating useful synthetic data and solving novel tasks with little correlation with training data is the exact issue here. Seems straightforwardly true that a transformer arcthiecture doesn’t do that?

I don’t know what superintelligent in-context learning is—I’d be skeptical that scaling a transformer a further 3 OOMS will suddenly make it do tasks that are very far from the text distribution it is trained on, indeed solutions to tasks that are not even remotely in the internet text data like building a recursively self-improving agent (if such a thing is possible...)? Maybe I’m misunderstanding what you’re claiming here.

Not saying it’s impossible, just seems deeply implausible. ofc LLMs being so impressive was also a prior implausible but this seems another OOM of implausibility bits if that makes sense?

generating useful synthetic data and solving novel tasks with little correlation with training data is the exact issue here. Seems straightforwardly true that a transformer arcthiecture doesn’t do that?

I’m imagining some prompts to generate reasoning, inferred claims about the world. You can’t generate new observations about the world, but you can reason about the observations available so far, and having those inferred claims in the dataset likely helps, that’s how humans build intuition about theory. If an average a 1000 inferred claims are generated for every naturally observed statement (or just those on rare/new/situational topics), that could close the gap of sample efficiency with humans. Might take the form of exercises or essays or something.

If this is all done with prompts, using a sufficiently smart order-following chatbot, then it’s straightforwardly just a transformer, with some superficial scaffolding. If this can work, it’ll eventually appear in distillation literature, though I’m not sure if serious effort to check was actually made with current SOTA LLMs, to pre-train exclusively on synthetic data that’s not too simplistically prompted. Possibly you get nothing for a GPT-3 level generator, and then something for GPT-4+, because reasoning needs to be good enough to preserve contact with ground truth. From Altman’s comments I get the impression that it’s plausibly the exact thing OpenAI is hoping for.

I don’t know what superintelligent in-context learning is

In-context learning is capability to make use of novel data that’s only seen in a context, not in pre-training, to do tasks that make use of this novel data, in ways that normally would’ve been expected to require it being seen in pre-training. In-context learning is a model capability, it’s learned. So its properties are not capped by those of the hardcoded model training algorithm, notably in principle in-context learning could have higher sample efficiency (which might be crucial for generating a lot of synthetic data out of a few rare observations). Right now it’s worse in most respects, but that could change with scale without substantially modifying the transformer architecture, which is the premise of this thread.

By superintelligent in-context learning I mean the capabilities of in-context learning significantly exceeding those of humans. Things like fully comprehending a new paper without changing any model weights, becoming able to immediately write the next one in the same context window. I agree that it’s not very plausible, and probably can’t happen without sufficiently deep circuits, which even deep networks don’t seem to normally develop. But it’s not really ruled out by anything that’s been tried so far. Recent stuff on essentially pre-trainingwith somefrozen weights without losing resulting performance suggests a trend of increasing feasible model size for given compute. So I’m not sure this can’t be done in a few years. Then there’s things like memory transformers, handing a lot more data than a context to a learned learning capability.

I wonder if the following is possible to study textbooks more efficiently using LLMs:

Feed the entire textbook to the LLM and produce a list of summaries that increases in granularity and length, covering all the material in the textbook just at a different depth (eg proofs omitted, further elaboration on high-level perspectives, etc)

The student starts from the highest-level summary, and gradually moves to the more granular materials.

When I study textbooks, I spend a significant amount of time improving my mental autocompletion, like being able to familiarize myself with the terminologies, which words or proof-style usually come in which context, etc. Doing this seems to significantly improve my ability to read eg long proofs, since I can ignore all the pesky details (which I can trust my mental autocompletion to later fill in the details if needed) and allocate my effort in getting a high-level view of the proof.

Textbooks don’t really admit this style of learning, because the students don’t have prior knowledge of all the concept-dependencies of a new subject they’re learning, and thus are forced to start at the lowest-level and make their way up to the high-level perspective.

Perhaps LLMs will let us reverse this direction, instead going from the highest to the lowest.

What’s a good technical introduction to Decision Theory and Game Theory for alignment researchers? I’m guessing standard undergrad textbooks don’t include, say, content about logical decision theory. I’ve mostly been reading posts on LW but as with most stuff here they feel more like self-contained blog posts (rather than textbooks that build on top of a common context) so I was wondering if there was anything like a canonical resource providing a unified technical / math-y perspective on the whole subject.

i absolutely hate bureaucracy, dumb forms, stupid websites etc. like, I almost had a literal breakdown trying to install Minecraft recently (and eventually failed). God.

I think what’s so crushing about it, is that it reminds me that the wrong people are designing things, and that they wont allow them to be fixed, and I can only find solace in thinking that the inefficiency of their designs is also a sign that they can be defeated.

God, I wish real analysis was at least half as elegant as any other math subject — way too much pathological examples that I can’t care less about. I’ve heard some good things about constructivism though, hopefully analysis is done better there.

Yeah, real analysis sucks. But you have to go through it to get to delightful stuff— I particularly love harmonic and functional analysis. Real analysis is just a bunch of pathological cases and technical persnicketiness that you need to have to keep you from steering over a cliff when you get to the more advanced stuff. I’ve encountered some other subjects that have the same feeling to them. For example, measure-theoretic probability is a dry technical subject that you need to get through before you get the fun of stochastic differential equations. Same with commutative algebra and algebraic geometry, or point-set topology and differential geometry.

Constructivism, in my experience, makes real analysis more mind blowing, but also harder to reason about. My brain uses non-constructive methods subconsciously, so it’s hard for me to notice when I’ve transgressed the rules of constructivism.

As a general reflection on undergraduate mathematics imho there is way too much emphasis on real analysis. Yes, knowing how to be rigorous is important, being aware of pathological counterexample is importanting, and real analysis is used all over the place. But there is so much more to learn in mathematics than real analysis and the focus on minor technical issues here is often a distraction to developing a broad & deep mathematical background.

For most mathematicians (and scientists using serious math) real analysis is a only a small part of the toolkit. Understanding well the different kinds of limits can ofc be crucial in functional analysis, stochastic processes and various parts of physics. But there are so many topics that are important to know and learn here!

The reason it is so prominent in the undergraduate curriculum seems to be more tied to institutional inertia, its prominence on centralized exams, relation with calculus, etc

Really, what’s going on is that in the general case, as mathematics is asked to be more and more general, you will start encountering pathological examples more, and paying attention to detail more is a valuable skill in both math and real life.

And while being technical about the pathological cases is kind of annoying, it’s also one that actually matters in real life, as you aren’t guaranteed to have an elegant solution to your problems.

Update: huh, nonstandard analysis is really cool. Not only are things much more intuitive (by using infinitesimals from hyperreals instead of using epsilon-delta formulation for everything), by the transfer principle all first order statements are equivalent between standard and nonstandard analysis!

Selection vs Control as distinguishing different types of “space of possibilities”

Selection as having that space explicitly given & selectable numerous times by the agent

Control as having that space only given in terms of counterfactuals, and the agent can access it only once.

These distinctions correlate with the type of algorithm being used & its internal structure, where Selection uses more search-like process using maps, while Control may just use explicit formula … although it may very well use internal maps to Select on counterfactual outcomes!

In other words, the Selection vs Control may very well be viewed as a different cluster of Analysis. Example:

If we decide to focus our Analysis of “space of possibilities” on eg “Real life outcome,” then a guided missile is always Control.

But if we decide to focus on “space of internal representation of possibilities,” then a guided missle that uses internal map to search on becomes Selection.

Greater emphasis on internal structure—specifically, “maps.”

Maps are capital investment, allowing you to be able to optimize despite not knowing what to exactly optimize for (by compressing info)

I have several thoughts on these framings, but one trouble is the excessive usage of words to represent “clusters” i.e. terms to group a bunch of correlated variables. Selection vs Control, for example, doesn’t have a clear definition/criteria but rather points at a number of correlated things, like internal structure, search, maps, control-like things, etc.

Sure, deconfusing and pointing out clusters is useful because clusters imply correlations and correlations perhaps imply hidden structure + relationships—but I think the costs from cluster-representing-words doing hidden inference is much greater than the benefits, and it would be better to explicitly lay out the features-of-clusters that the one is referring to instead of just using the name of the cluster.

This is similar to the trouble I had with “wrapper-minds,” which is yet another example of a cluster pointing at a bunch of correlated variables, and people using the same term to mean different things.

Anyways, I still feel totally confused about optimization—and while these clusters/frames are useful, I think thinking in terms of them would ensue even more confusion within myself. It’s probably better to take the useful individual parts within the cluster and start deconfusing from the ground-up using those as the building blocks.

By boundaries, I mean a sustaining/propagating system that informationally/causally insulates its ‘viscera’ from the ‘environment,’ and only allows relatively small amounts of deliberate information flow through certain channels in both directions. Living systems are an example of it (from bacteria to humans). It doesn’t even have to be a physically distinct chunk of spacetime, they can be over more abstract variables like societal norms. Agents are an example of it.

I find them very relevant to alignment especially from the direction of detecting such boundary-possessing/agent-like structures embedded in a large AI system and backing out a sparse relationship between these subsystems, which can then be used to e.g., control the overall dynamic. Check out theseposts for more.

A prototypical deliverable would be an algorithm that can detect such ‘boundaries’ embedded in a dynamical system when given access to some representation of the system, performs observations & experiments and returns a summary data structure of all the ‘boundaries’ embedded in a system and their desires/wants, how they game-theoretically relate to one another (sparse causal relevance graph?), the consequences of interventions performed on them, etc—that’s versatile enough to detect e.g., gliders embedded in Game of Life / Particle Lenia, agents playing Minecraft while only given coarse grained access to the physical state of the world, boundary-like things inside LLMs, etc. (I’m inspired by this)

Why do I find the aforementioned directions relevant to this goal?

Critch’s Boundaries operationalizes boundaries/viscera/environment as functions of the underlying variable that executes policies that continuously prevents information ‘flow’ ^{[1]} between disallowed channels, quantified via conditional transfer entropy.

Relatedly, Fernando Rosas’s paper on Causal Blankets operationalize boundaries using a similar but subtly different^{[2]} form of mutual information constraint on the boundaries/viscera/environment variables than that of Critch’s. Importantly, they show that such blankets always exist between two coupled stochastic processes (using a similar style of future morph equivalence relation characterization from compmech, and also a metric they call “synergistic coefficient” that quantifies how boundary-like this thing is.^{[3]}

More on compmech, epsilon transducers generalize epsilon machines to input-output processes. PALO (Perception Action Loops) and Boundaries as two epsilon transducers coupled together?

These directions are interesting, but I find them still unsatisfactory because all of them are purely behavioral accounts of boundaries/agency. One of the hallmarks of agentic behavior (or some boundary behaviors) is adapting ones policy if an intervention changes the environment in a way that the system can observe and adapt to.^{[4]}^{[5]}

(is there an interventionist extension of compmech?)

Discovering agents provide a genuine causal, interventionist account of agency and an algorithm to detect them, motivated by the intentional stance. I think the paper is very enlightening from a conceptual perspective, but there are many problems yet to be solved before we can actually implement this. Here’s my take on it.

More fundamentally, (this is more vibes, I’m really out of my depth here) I feel there is something intrinsically limiting with the use of Bayes Nets, especially with the fact that choosing which variables to use in your Bayes Net already encodes a lot of information about the specific factorization structure of the world. I heard good things about finite factored sets and I’m eager to learn more about them.

Not exactly a ‘flow’, because transfer entropy conflates between intrinsic information flow and synergistic information—a ‘flow’ connotes only the intrinsic component, while transfer entropy just measures the overall amount of information that a system couldn’t have obtained on its own. But anyways, transfer entropy seems like a conceptually correct metric to use.

Specifically, Fernando’s paper criticizes blankets of the following form (V for viscera, A and P for active/passive boundaries, E for environment):

Vt→At,Pt→Et

DIP implies I(Vt;At,Pt)≥I(Vt;Et)

This clearly forbids dependencies formed in the past that stays in ‘memory’.

but Critch instead defines boundaries as satisfying the following two criteria:

Vt+1,At+1→Vt,At,Pt→Et (infiltration)

DIP implies I(Vt;At,Pt)≥I(Vt;Et)

Et+1,Pt+1→At,Pt,Et→Vt (exfiltration)

DIP implies I(Vt+1,At+1;At,Pt)≥I(Vt+1,At+1;Et)

and now that the independencies are entangled across different t, there is no longer a clear upper bound on I(Vt;Et), so I don’t think the criticisms apply directly.

My immediate curiosities are on how these two formalisms relate to one another. e.g., Which independency requirements are more conceptually ‘correct’? Can we extend the future-morph construction to construct Boundaries for Critch’s formalism? etc etc

For example, a rock is very goal-directed relative to ‘blocking-a-pipe-that-happens-to-exactly-match-its-size,’ until one performs an intervention on the pipe size to discover that it can’t adapt at all.

Also, interventions are really cheap to run on digital systems (e.g., LLMs, cellular automata, simulated environments)! Limiting oneself to behavioral accounts of agency would miss out on a rich source of cheap information.

Does anyone know if Shannon arrive at entropy from the axiomatic definition first, or the operational definition first?

I’ve been thinking about these two distinct ways in which we seem to arrive at new mathematical concepts, and looking at the countless partial information decomposition measures in the literature all derived/motivated based on an axiomatic basis, and not knowing which intuition to prioritize over which, I’ve been assigning less premium on axiomatic conceptual definitions than i used to:

decision theoretic justification of probability > Cox’s theorem

shannon entropy as min description length > three information axioms

The basis of comparison would be its usefulness and ease-of-generalization to better concepts:

at least in the case of fernando’s synergistic information, it seems far more useful because i at least know what i’m exactly getting out of it, unlike having to compare between the axiomatic definitions based on handwavy judgements.

for ease of generalization, the problem with axiomatic definitions is that there are many logically equivalent ways to state the initial axiom (from which they can then be relaxed), and operational motivations seem to ground these equivalent characterizations better, like logical inductors from the decision theoretic view of probability theory

I’m not sure what you mean by operational vs axiomatic definitions.

But Shannon was unaware of the usage of S=−Σipilnpi in statistical mechanics. Instead, he was inspired by Nyquist and Hartley’s work, which introduced ad-hoc definitions of information in the case of constant probability distributions.

And in his seminal paper, “A mathematical theory of communication”, he argued in the introduction for the logarithm as a measure of information because of practicality, intuition and mathematical convenience. Moreover, he explicitly derived the entropy of a distribution from three axioms: 1) that it be continuous wrt. the probabilities, 2) that it increase monotonically for larger systems w/ constant probability distributions, 3) and that it be a weighted sum the entropy of sub-systems. See section 6 for more details.

‘Symmetry’ implies ‘redundant coordinate’ implies ‘cyclic coordinates in your Lagrangian / Hamiltonian’ implies ‘conservation of conjugate momentum’

And because the action principle (where the true system trajectory extremizes your action, i.e. integral of Lagrangian) works in various dynamical systems, the above argument works in non-physical dynamical systems.

Thus conserved quantities usually exist in a given dynamical system.

mmm, but why does the action principle hold in such a wide variety of systems though? (like how you get entropy by postulating something to be maximized in an equilibrium setting)

Mildly surprised how some verbs/connectives barely play any role in conversations, even in technical ones. I just tried directed babbling with someone, and (I think?) I learned quite a lot about Israel-Pakistan relations with almost no stress coming from eg needing to make my sentences grammatically correct.

Example of (a small part of) my attempt to summarize my understanding of how Jews migrated in/out of Jerusalem over the course of history:

They here *hand gesture on air*, enslaved out, they back, kicked out, and boom, they everywhere.

(audience nods, given common knowledge re: gestures, meaning of “they,” etc)

My (completely amateur) understanding is that the “extra” semantic and syntactic structure of written and spoken language does two things.

One, it adds redundancy and reduces error. Simple example, gendered pronouns mean that when you hear “Have you seen Laurence? She didn’t get much sleep last night.” you have a chance to ask the speaker for clarification and catch if they had actually said “Laura” and you misheard.

Two, it can be used as a signal. The correct use of jargon is used by listeners or readers as a proxy for competence. Or many typos in your text will indicate to readers that you haven’t put much effort into what you’re saying.

Why haven’t mosquitos evolved to be less itchy? Is there just not enough selection pressure posed by humans yet? (yes probably) Or are they evolving towards that direction? (they of course already evolved towards being less itchy while biting, but not enough to make that lack-of-itch permanent)

this is a request for help i’ve been trying and failing to catch this one for god knows how long plz halp

tbh would be somewhat content coexisting with them (at the level of houseflies) as long as they evolved the itch and high-pitch noise away, modulo disease risk considerations.

The reason mosquito bites itch is because they are injecting saliva into your skin. Saliva contains mosquito antigens, foreign particles that your body has evolved to attack with an inflammatory immune response that causes itching. The compound histamine is a key signaling molecule used by your body to drive this reaction.

In order for the mosquito to avoid provoking this reaction, they would either have to avoid leaving compounds inside of your body, or mutate those compounds so that they do not provoke an immune response. The human immune system is an adversarial opponent designed with an ability to recognize foreign particles generally. If it was tractable for organisms to reliably evolve to avoid provoking this response, that would represent a fundamental vulnerability in the human immune system.

Mosquitoe saliva does in fact contain anti-inflammatory, antihemostatic, and immunomodulatory compounds. So they’re trying! But also this means that mosquitos are evolved to put saliva inside of you when they feed, which means they’re inevitably going to expose the foreign particles they produce to your immune system.

There’s also a facet of selection bias making mosquitos appear unsuccessful at making their bites less itchy. If a mosquito did evolve to not provoke (as much of) an immune response and therefore less itching, redness and swelling, you probably wouldn’t notice they’d bitten you. People often perceive that some are prone to getting bitten, others aren’t. It may be that some of this is that some people don’t have as serious an immune response to mosquito bites, so they think they get bitten less often.

I’m sure there are several PhDs worth of research questions to investigate here—I’m a biomedical engineer with a good basic understanding of the immune system, but I don’t study mosquitos.

Because they have no reproductive advantage to being less itchy. You can kill them while they’re feeding, which is why they put lots of evolutionary effort into not being noticed. (They have an anesthetic in their saliva so you are unlikely to notice the bite.) By the time you develop the itchy bump, they’ve flown away and you can’t kill them.

There’s still some pressure, though. If the bites were permanently not itchy, then I may have not noticed that the mosquitos were in my room in the first place, and consequently would less likely pursue them directly. I guess that’s just not enough.

There’s also positive selection for itchiness. Mosquito spit contains dozens of carefully evolved proteins. We don’t know what they all are, but some of them are anticoagulants and anesthetics. Presumably they wouldn’t be there if they didn’t have a purpose. And your body, when it detects these foreign proteins, mounts a protective reaction, causing redness, swelling, and itching. IIRC, that reaction does a good job of killing any viruses that came in with the mosquito saliva. We’ve evolved to have that reaction. The itchiness is probably good for killing any bloodsuckers that don’t flee quickly. It certainly works against ticks.

Evolution is not our friend. It doesn’t give us what we want, just what we need.

I believe mosquitos do inject something to suppress your reaction to them, which is why you don’t notice bug bites until long after the bug is gone. There’s no reproductive advantage to the mosquito to extending that indefinitely.

I had something like locality in mind when writing this shortform, the context being: [I’m in my room → I notice itch → I realize there’s a mosquito somewhere in my room → I deliberately pursue and kill the mosquito that I wouldn’t have known existed without the itch]

But, again, this probably wouldn’t amount to much selection pressure, partially due to the fact that the vast majority of mosquito population exists in places where such locality doesn’t hold i.e. in an open environment.

But the evolutionary timescale at which mosquitos can adapt to avoid detection must be faster than that of humans adapting to find mosquitos itchy! Or so I thought—my current boring guess is that (1) mechanisms for the human body to detect foreign particles are fairly “broad”, (2) the required adaptation from the mosquitos to evade them are not-way-too-simple, and (3) we just haven’t put enough selection pressure to make such change happen.

Just noticing that the negation of a statement exists is enough to make meaningful updates.

e.g. I used to (implicitly) think “Chatbot Romance is weird” without having evaluated anything in-depth about the subject (and consequently didn’t have any strong opinions about it)—probably as a result of some underlying cached belief.

But after seeing this post, just reading the title was enough to make me go (1) “Oh! I just realized it is perfectly possible to argue in favor of Chatbot Romance … my belief on this subject must be a cached belief!” (2) hence is probably by-default biased towards something like the consensus opinion, and (3) so I should update away from my current direction, even without reading the post.

People mean different things when they say “values” (object vs meta values)

I noticed that people often mean different things when they say “values,” and they end up talking past each other (or convergence only happens after a long discussion). One of the difference is in whether they contain meta-level values.

Some people refer to the “object-level” preferences that we hold.

Often people bring up the “beauty” of the human mind’s capacity for its values to change, evolve, adopt, and grow—changing mind as it learns more about the world, being open to persuasion via rational argumentation, changing moral theories, etc.

Some people include the meta-values (that are defined on top of other values, and the evolution of such values).

e.g., My “values” include my meta-values, like wanting to be persuaded by good arguments, wanting to change my moral theories when I get to know better, even “not wanting my values to be fixed”

example of this view: carado’s post on you want what you want, and one of Vanessa Cosoy’s shortform/comment (can’t remember the link)

Unidimensional Continuity of Preference ≈ Assumption of “Resources”?

tl;dr, the unidimensional continuity of preference assumption in the money pumping argument used to justify the VNM axioms correspond to the assumption that there exists some unidimensional “resource” that the agent cares about, and this language is provided by the notion of “souring / sweetening” a lottery.

Various coherence theorems—or more specifically, various money pumping arguments generally have the following form:

If you violate this principle, then [you are rationally required] / [it is rationally permissible for you] to follow this trade that results in you throwing away resources. Thus, for you to avoid behaving pareto-suboptimally by throwing away resources, it is justifiable to call this principle a ‘principle of rationality,’ which you must follow.

… where “resources” (the usual example is money) are something that, apparently, these theorems assume exist. They do, but this fact is often stated in a very implicit way. Let me explain.

In the process of justifying the VNM axioms using money pumping arguments, one of the three main mathematical primitives are: (1) lotteries (probability distribution over outcomes), (2) preference relation (general binary relation), and (3) a notion of Souring/Sweetening of a lottery. Let me explain what (3) means.

Souring of A is denoted A−, and a sweetening of A is denoted A+.

A− is to be interpreted as “basically identical with A but strictly inferior in a single dimension that the agent cares about.” Based on this interpretation, we assume A>A−. Sweetening is the opposite, defined in the obvious way.

Formally, souring could be thought of as introducing a new preference relation A>uniB, which is to be interpreted as “lottery B is basically identical to lottery A, but strictly inferior in a single dimension that the agent cares about”.

On the syntactic level, such B is denoted as A−.

On the semantic level, based on the above interpretation, >uni is related to > via the following: A>uniB⟹A>B

This is where the language to talk about resources come from. “Something you can independently vary alongside a lottery A such that more of it makes you prefer that option compared to A alone” sounds like what we’d intuitively call a resource^{[1]}.

Now that we have the language, notice that so far we haven’t assumed sourings or sweetenings exist. The following assumption does it:

Unidimensional Continuity of Preference: If X>Y, then there exists a prospect X− such that 1) X− is a souring of X and 2) X>X−>Y.

Which gives a more operational characterization of souring as something that lets us interpolate between the preference margins of two lotteries—intuitively satisfied by e.g., money due to its infinite divisibility.

So the above assumption is where the assumption of resources come into play. I’m not aware of any money pump arguments for this assumption, or more generally, for the existence of a “resource.” Plausibly instrumental convergence.

I don’t actually think this + the assumption below fully capture what we intuitively mean by “resources”, enough to justify this terminology. I stuck with “resources” anyways because others around here used that term to (I think?) refer to what I’m describing here.

Thinking about for some time my feeling has been that resources are about fungibility implicitly embedded in a context of trade, multiple agents (very broadly construed. E.g. an agent in time can be thought of as multiple agents cooperating intertemporally perhaps).

A resource over time has the property that I can spend it now or I can spend it later. Glibly, one could say the operational meaning of the resource arises from the intertemporal bargaining of the agent.

Perhaps it’s useful to distinguish several levels of resources and resource-like quantities.

Discrete vs continuous, tradeable / meaningful to different agents, ?? Fungibility, ?? Temporal and spatial locatedness, ?? Additivity?, submodularity ?

Addendum: another thing to consider is that the input of the vNM theorem is in some sense more complicated than the output.
The output is just a utility function u: X → R, while your input is a preference order on the very infinite set of lotteries (= probability distributions ) L(X).

Thinking operationally about a preference ordering on a space of distribution is a little wacky.
It means you are willing to trade off uncertain options against one another. For this to be a meaningful choice would seem to necessitate some sort of (probabilistic) world model.

Damn, why did Pearl recommend readers (in the preface of his causality book) to read all the chapters other than chapter 2 (and the last review chapter)? Chapter 2 is literally the coolest part—inferring causal structure from purely observational data! Almost skipped that chapter because of it …

it’s true it’s cool, but I suspect he’s been a bit disheartened by how complicated it’s been to get this to work in real-world settings.

in the book of why, he basically now says it’s impossible to learn causality from data, which is a bit of a confusing message if you come from his previous books.

but now with language models, I think his hopes are up again, since models can basically piggy-back on causal relationships inferred by humans

Bayes Net inference algorithms maintain its efficiency by using dynamic programming over multiple layers.

Level 0: Naive Marginalization

No dynamic programming whatsoever. Just multiply all the conditional probability distribution (CPD) tables, and sum over the variables of non-interest.

Level 1: Variable Elimination

Cache the repeated computations within a query.

For example, given a chain-structured Bayes Net A⟶B⟶C⟶D, instead of doing P(D)=∑A∑B∑CP(A,B,C,D), we can do P(D)=∑CP(D|C)∑BP(C|B)∑AP(A)P(B|A). Check my post for more.

Suppose you have a fixed Bayes Net, and you want to compute the marginalization not only P(D), but also P(A). Clearly running two instances of Variable Elimination as above is going to contain some overlapping computation.

Clique-tree is a data structure where, given the initial factors (in this case the CPD tables), you “calibrate” a tree whose nodes correspond to a subset of the variables. Cost can be amortized by running many queries over the same Bayes Net.

Calibration can be done by just two passes across the tree, after which you have the joint marginals for all the nodes of the clique tree.

Incorporating evidence is equally simple. Just zero-out the entries of variables that you are conditioning on for some node, then “propagate” that information downwards via a single pass across the tree.

Level 3: Specialized query-set answering algorithms over a calibrated clique tree.

Cache the repeated computations across a certain query-class

e.g., computing P(X,Y) for every pair of variables can be done by using yet another layer of dynamic programming by maintaining a table of P(Ci,Cj) for each pair of clique-tree nodes ordered according to their distance in-between.

what are macrostates? Variables which are required to make your thermodynamics theory work! If they don’t, add more macrostates!

nonequilibrium? Define it as systems that don’t admit a thermodynamic description!

inductive biases? Define it as the amount of correction needed for a system to obey Bayesian updating, i.e. correction terms in the exponent of the Gibbs measure!

coarse graining? Define the coarse-grained variables to keep the dynamics as close as possible to that of the micro-dynamics!

or in a similar spirit—does your biological system deviate from expected utility theory? Well, there’s discovery (and money) to be made!

It’s easy to get confused and think the circularity is a problem (“how can you define thermodynamics in terms of equilibriums, when equilibriums are defined using thermodynamics?”), but it’s all about carving nature at the right joints—and a sign that you made the right carving is that the amount of corrections needed to be applied aren’t too numerous, and they all seem “natural” (and of course, all of this while letting you make nontrivial predictions. that’s what matters at the end of the day).

Then, it’s often the case that those corrections also turn out to be meaningful and natural quantities of interest.

One of the rare insightful lessons from high school: Don’t set your AC to the minimum temperature even if it’s really hot, just set it to where you want it to be.

It’s not like the air released gets colder with lower target temperature, because most ACs (according to my teacher, I haven’t checked lol) are just a simple control system that turns itself on/off around the target temperature, meaning the time it takes to reach a certain temperature X is independent of the target temperature (as long it’s lower than X)

Well is he is right about some ACs being simple on/off units.

But there also exists units than can change cycle speed, its basically the same thing except the motor driving the compression cycle can vary in speed.

In case you where wondering, they are called inverters. And when buying new today, you really should get an inverter (efficiency).

(Note: This was a post, but in retrospect was probably better to be posted as a shortform)

(Epistemic Status: 20-minute worth of thinking, haven’t done any builder/breaker on this yet although I plan to, and would welcome any attempts in the comment)

Have an algorithmic task whose input/output pair could (in reasonable algorithmic complexity) be generated using highly specific combination of modular components (e.g., basic arithmetic, combination of random NN module outputs, etc).

Train a small transformer (or anything, really) on the input/output pairs.

Take a large transformer that takes the activation/weights, and outputs a computational graph.

Train that large transformer over the small transformer, across a diverse set of such algorithmic tasks (probably automatically generated) with varying complexity. Now you have a general tool that takes in a set of high-dimensional matrices and backs-out a simple computational graph, great! Let’s call it Inspector.

Apply the Inspector in real models and see if it recovers anything we might expect (like induction heads).

To go a step further, apply the Inspector to itself. Maybe we might back-out a human implementable general solution for mechanistic interpretability! (Or, at least let us build a better intuition towards the solution.)

(This probably won’t work, or at least isn’t as simple as described above. Again, welcome any builder/breaker attempts!)

I want to focus on having a better mechanistic picture of agent value formation & distinguishing between hypotheses (e.g., shard theory, Thane Ruthenis’s value-compilation hypothesis, etc) and forming my own.

I think I have a specific but very high uncertainty baseline model of what-to-expect from agent value-formation using greedy search optimization. It’s probably time to allocate more resources on reducing that uncertainty by touching reality i.e. running experiments.

(and also think about related theoretical arguments like Selection Theorem)

So I’ll probably allocate my research time:

Studying math (more linear algebra / dynamical systems / causal inference / statistical mechanics)

Sketching a better picture of agent development, assigning confidence, proposing high-bit experiments (that might have the side-effect of distinguishing between different conflicting pictures), formalization, etc.

and read relevant literature (eg ones on theoretic DL and inductive biases)

Upskilling mechanistic interpretability to actually start running quick experiments

Unguided research brainstorming (e.g., going through various alignment exercises, having a writeup of random related ideas, etc)

Possibly participate in programs like MATS? Probably the biggest benefit to me would be (1) commitment mechanism / additional motivation and (2) high-value conversations with other researchers.

Useful perspective when thinking of mechanistic pictures of agent/value development is to take the “perspective” of different optimizers, consider their relative “power,” and how they interact with each other.

E.g., early on SGD is the dominant optimizer, which has the property of (having direct access to feedback from U / greedy). Later on early proto-GPS (general-purpose search) forms, which is less greedy, but still can largely be swayed by SGD (such as having its problem-specification-input tweaked, having the overall GPS-implementation modified, etc). Much later, GPS becomes the dominant optimizing force “at run-time” which shortens the relevant time-scale and we can ignore the SGD’s effect. This effect becomes much more pronounced after reflectivity + gradient hacking when the GPS’s optimization target becomes fixed.

(very much inspired by reading Thane Ruthenis’s value formation post)

This is a very useful approximation at the late-stage when the GPS self-modifies the agent in pursuit of its objective! Rather than having to meticulously think about local SGD gradient incentives and such, since GPS is non-greedy, we can directly model it as doing what’s obviously rational from a birds-eye-perspective.

(kinda similar to e.g., separation of timescale when analyzing dynamical systems)

It seems like retrieval-based transformers like RETRO is “obviously” the way to go—(1) there’s just no need to store all the factual information as fixed weights, (2) and it uses much less parameter/memory. Maybe mechanistic interpretability should start paying more attention to these type of architectures, especially since they’re probably going to be a more relevant form of architecture.

They might also be easier to interpret thanks to specialization!

I’ve noticed during my alignment study that just the sheer amount of relevant posts out there is giving me a pretty bad habit of (1) passively engaging with the material and (2) not doing much independent thinking. Just keeping up to date & distilling the stuff in my todo read list takes up most of my time.

I guess the reason I do it is because (at least for me) it takes a ton of mental effort to switch modes between “passive consumption” and “active thinking”:

I noticed then when self-studying math; like, my subjective experience is that I enjoy both “passively listening lectures+taking notes” and “solving practice problems,” the problem is that it takes a ton of mental energy to switch between the two equilibriums.

(This is actually still a problem—too much wide & passive consumption rather than actively practicing them and solving problems.)

Also relevant is wanting to just progress/upskill as fast and wide of a subject as I can, sacrificing mastery for diversity. This probably makes sense to some degree (especially in the sense that having more frames is good), but I think I’m taking this wayyyy too far.

My r for opening new links far exceeds 1. This definitely helped me when I was trying to get a rapid overview of the entire field, but now it’s just a bad adaptation + akrasia.

Okay, then, don’t do that! Some directions to move towards:

Independent brainstorming/investigation sessions to form concrete inside views

There are lots of posts but the actual content is very thing. I would say there is plausibly more content in your real analysis book than there is in the entire alignment field.

(Epistemic Status: I don’t endorse this yet, just thinking aloud. Please let me know if you want to act/research based on this idea)

It seems like it should be possible to materialize certain forms of AI alignment failure modes with today’s deep learning algorithms, if we directly optimize for their discovery. For example, training a Gradient Hacker Enzyme.

A possible benefit of this would be that it gives us bits of evidence wrt how such hypothesized risks would actually manifest in real training environments. While the similarities would be limited because the training setups would be optimizing for their discovery, it should at least serve as a good lower bound for the scenarios in which these risks could manifest.

Perhaps having a concrete bound for when dangerous capabilities appear (eg a X parameter model trained in Y modality has Z chance of forming a gradient hacker) would make it easier for policy folks to push for regulations.

Is AI gain-of-function equally dangerous as biotech gain-of-function? Some arguments in favor (of the former being dangerous):

The malicious actor argument is probably stronger for AI gain-of-function.

if someone publicly releases a Gradient Hacker Enzyme, this lowers the resource that would be needed for a malicious actor to develop a misaligned AI (eg plug in the misaligned Enzyme at an otherwise benign low-capability training run).

Risky researcher incentive is equally strong.

e.g., a research lab carelessly pursuing gain-of-function research, deliberately starting risky training runs for financial/academic incentives, etc.

Some arguments against:

Accident risks from financial incentives are probably weaker for AI gain-of-function.

The standard gain-of-function risk scenario is: research lab engineers a dangerous pathogen, it accidentally leaks, and a pandemic happens.

I don’t see how these events would happen “accidentally” when dealing with AI programs; e.g., the researcher would have to deliberately cut parts of the network weights and replace it with the enzyme, which is certainly intentional.

Random alignment-related idea: train and investigate a “Gradient Hacker Enzyme”

TL;DR, Use meta-learning methods like MAML to train a network submodule i.e. circuit that would resist gradient updates in a wide variety of contexts (various architectures, hyperparameters, modality, etc), and use mechanistic interpretability to see how it works.

It should be possible to have a training setup for goals other than “resist gradient updates,” such as restricting the meta-objective to a specific sub-sub-circuit. In that case, the outer circuit might (1) instrumentally resist updates, or (2) somehow get modified while keeping its original behavioral objective intact.

This setup doesn’t have to be restricted to circuits of course; there was a previous work which did this on the level of activations, although iiuc the model found a trivial solution by exploiting relu—it would be interesting to extend this to more diverse setup.

Anyways, varying this “sub-sub-circuit/activation-to-be-preserved” over different meta-learning episodes would incentivize the training process to find “general” Gradient Hacker designs that aren’t specific to a particular circuit/activation—a potential precursor for various forms of advanced Gradient Hackers (and some loose analogies to how enzymes accelerate reactions).

What is the Theory of Impact for training a “Gradient Hacker Enzyme”?

(note: while I think these are valid, they’re generated post-hoc and don’t reflect the actual process for me coming up with this idea)

Estimating the lower-bound for the emergence Gradient Hackers.

By varying the meta-learning setups we can get an empirical estimate for the conditions in which Gradient Hackers are possible.

Perhaps gradient hackers are actually trivial to construct using tricks we haven’t thought of before (like the relu example before). Maybe not! Perhaps they require [high-model-complexity/certain-modality/reflective-agent/etc].

Why lower-bound? In a real training environment, gradient hackers appear because of (presumably) convergent training incentives. Instead in the meta-learning setup, we’re directly optimizing for gradient hackers.

Mechanistically understanding how Gradient Hackers work.

Applying mechanistic interpretability here might not be too difficult, because the circuit is cleanly separated from the rest of the model.

There has been severalspeculationson how such circuits might emerge. Testing them empirically sounds like a good idea!

This is just a random idea and I’m probably not going to work on it; but if you’re interested, let me know. While I don’t think this is capabilities-relevant, this probably falls under AI gain-of-function research and should be done with caution.

Update: I’m trying to upskill mechanistic interpretability, and training a Gradient Hacker Enzyme seems like a fairly good project just to get myself started.

I don’t think this project would be highly valuable in and of itself (although I would definitely learn a lot!), so one failure mode I need to avoid is ending up investing too much of my time in this idea. I’ll probably spend a total of ~1 week working on it.

Thoughtdump on why I’m interested in computational mechanics:

one concrete application to natural abstractions from here: tl;dr, belief structures generally seem to be fractal shaped. one major part of natural abstractions is trying to find the correspondence between structures in the environment and concepts used by the mind. so if we can do the inverse of what adam and paul did, i.e. ‘discover’ fractal structures from activations and figure out what stochastic process they might correspond to in the environment, that would be cool

… but i was initially interested in reading compmech stuff not with a particular alignment relevant thread in mind but rather because it seemed broadly similar in directions to natural abstractions.

re: how my focus would differ from my impression of current compmech work done in academia: academia seems faaaaaar less focused on actually trying out epsilon reconstruction in real world noisy data. CSSR is an example of a reconstruction algorithm. apparently people did compmech stuff on real-world data, don’t know how good, but effort-wise far too less invested compared to theory work

would be interested in these reconstruction algorithms, eg what are the bottlenecks to scaling them up, etc.

tangent: epsilon transducers seem cool. if the reconstruction algorithm is good, a prototypical example i’m thinking of is something like: pick some input-output region within a model, and literally try to discover the hmm model reconstructing it? of course it’s gonna be unwieldly large. but, to shift the thread in the direction of bright-eyed theorizing …

the foundational Calculi of Emergence paper talked about the possibility of hierarchical epsilon machines, where you do epsilon machines on top of epsilon machines and for simple examples where you can analytically do this, you get wild things like coming up with more and more compact representations of stochastic processes (eg data stream → tree → markov model → stack automata → … ?)

this … sounds like natural abstractions in its wildest dreams? literally point at some raw datastream and automatically build hierarchical abstractions that get more compact as you go up

haha but alas, (almost) no development afaik since the original paper. seems cool

and also more tangentially, compmech seemed to have a lot to talk about providing interesting semantics to various information measures aka True Names, so another angle i was interested in was to learn about them.

eg crutchfield talks a lot about developing a right notion of information flow—obvious usefulness in eg formalizing boundaries?

many other information measures from compmech with suggestive semantics—cryptic order? gauge information? synchronization order? check ruro1 and ruro2 for more.

I agree with you.

Epsilon machine (and MSP) construction is most likely computationally intractable [I don’t know an exact statement of such a result in the literature but I suspect it is true] for realistic scenarios.

Scaling an approximate version of epsilon reconstruction seems therefore of prime importance. Real world architectures and data has highly specific structure & symmetry that makes it different from completely generic HMMs. This must most likely be exploited.

The calculi of emergence paper has inspired many people but has not been developed much. Many of the details are somewhat obscure, vague. I also believe that most likely completely different methods are needed to push the program further. Computational Mechanics’ is primarily a theory of hidden markov models—it doesn’t have the tools to easily describe behaviour higher up the Chomsky hierarchy. I suspect more powerful and sophisticated algebraic, logical and categorical thinking will be needed here. I caveat this by saying that Paul Riechers has pointed out that actually one can understand all these gadgets up the Chomsky hierarchy as infinite HMMs which may be analyzed usefully just as finite HMMs.

The still-underdeveloped theory of epsilon transducers I regard as the most promising lens on agent foundations. This is uncharcted territory; I suspect the largest impact of computational mechanics will come from this direction.

Your point on True Names is well-taken. More basic examples than gauge information, synchronization order are the triple of quantites entropy rate h, excess entropy E and Crutchfield’s statistical/forecasting complexity C. These are the most important quantities to understand for any stochastic process (such as the structure of language and LLMs!)

Just read through Robust agents learn causal world models and man it is really cool! It proves a couple of bona fide selection theorems, talking about the internal structure of agents selected against a certain criteria.

Tl;dr, agents selected to perform robustly in various local interventional distributions must internally represent something isomorphic to a causal model of the variables upstream of utility, for it is capable of answering all causal queries for those variables.

Thm 1: agents achieving optimal policy (util max) across various local interventions must be able to answer causal queries for all variables upstream of the utility nodeThm 2: relaxation of above to nonoptimal policies, relating regret bounds to the accuracy of the reconstructed causal modelthe proof is constructive—an algorithm that, when given access to regret-bounded-policy-oracle wrt an environment with some local intervention, queries them appropriately to construct a causal model

one implication is an algorithm for causal inference that converts black box agents to explicit causal models (because, y’know, agents like you and i are literally that aforementioned ‘regret-bounded-policy-oracle‘)

These selection theorems could be considered the converse of the well-known statement that given access to a causal model, one can find an optimal policy. (this and its relaxation to approximate causal models is stated in

Thm 3)Thm 1 / 2is like a ‘causal good regulator‘ theorem.gooder regulator theorem is not structural—as in, it gives conditions under which a model of the regulator must be isomorphic to the posterior of the system—a black box statement about the input-output behavior.

theorem is limited. only applies to cases where the decision node is not upstream of the environment nodes (eg classification. a negative example would be an mdp). but authors claim this is mostly for simpler proofs and they think this can be relaxed.

yes !! discovered this last week—seems very important the quantitative regret bounds for approximatiions is especially promising

I think you can drop this premise and modify the conclusion to “you can find a causal model for all variables upstream of the utility

and not downstream of the decision.”Quick paper review of Measuring Goal-Directedness from the causal incentives group.

tl;dr, goal directedness of a policy wrt a utility function is measured by its min distance to one of the policies implied by the utility function, as per the intentional stance—that one should model a system as an agent insofar as doing so is useful.

## Details

how is “policies implied by the utility function” operationalized? given a value u, we define a set containing policies of maximum entropy (of the decision variable, given its parents in the causal bayes net) among those policies that attain the utility u.

then union them over all the achievable values of u to get this “wide set of maxent policies,” and define goal directedness of a policy π wrt a utility function U as the maximum (negative) cross entropy between π and an element of the above set. (actually we get the same result if we quantify the min operation over just the set of maxent policies achieving the same utility as π.)

## Intuition

intuitively, this is measuring:

“how close is my policyπto being ‘deterministic,’ while ‘optimizingUat the competence levelu(π)’ and not doing anything else ‘deliberately’?”“close” / “deterministic” ~ large negative CE means small CE(π,πmaxent)=H(π)+KL(π||πmaxent)

“not doing anything else deliberately’” ~ because we’re quantifying over

maxentpolicies. the policy is maximally uninformative/uncertain, the policy doesn’t take any ‘deliberate’ i.e. low entropy action, etc.“at the competence level u(π)” ~ … under the constraint that it is identically competent to π

and you get the nice property of the measure being invariant to translation / scaling of U.

obviously so, because a policy is maxent among all policies achieving u on U iff that same policy is maxent among all policies achieving au+b on aU+b, so these two utilities have the same “wide set of maxent policies.”

## Critiques

I find this measure problematic in many places, and am confused whether this is conceptually correct.

one property claimed is that the measure is maximum for uniquely optimal / anti-optimal policy.

it’s interesting that this measure of goal-directedness isn’t exactly an ~increasing function of u(π), and i think it makes sense. i want my measure of goal-directedness to, when evaluated relative to human values, return a large number for both aligned ASI and signflip ASI.

… except, going through the proof one finds that the latter property heavily relies on the “uniqueness” of the policy.

My policy can get the maximum goal-directedness measure if it is the only policy of its competence level while being very deterministic. It isn’t clear that this always holds for the optimal/anti-optimal policies or always relaxes smoothly to epsilon-optimal/anti-optimal policies.

Relatedly, the fact that the quantification is only happening over policies of the same competence level, which feels problematic.

~~minimum for uniformly random policy~~(this would’ve been a good property, but unless I’m mistaken I think the proof for the lower bound is incorrect, because negative cross entropy is not bounded below.)honestly the maxent motivation isn’t super clear to me.

not causal. the reason you need causal interventions is because you want to rule out accidental agency/goal-directedness, like a rock that happens to be the perfect size to seal a water bottle—does your rock adapt when I intervene to change the size of the hole? discovering agents is excellent in this regards.

Reminds me a little bit of this idea from Vanessa Kosoy.

Thanks for the feedback!

Yeah, uniqueness definitely doesn’t always hold for the optimal/anti-optimal policy. I think the way MEG works here makes sense: if you’re following the unique optimal policy for some utility function, that’s a lot of evidence for goal-directedness. If you’re following one of many optimal policies, that’s a bit less evidence—there’s a greater chance that it’s an accident. In the most extreme case (for the constant utility function) every policy is optimal—and we definitely don’t want to ascribe maximum goal-directedness to optimal policies there.

With regard to relaxing smoothly to epsilon-optimal/anti-optimal policies, from memory I think we do have the property that MEG is increasing in the utility of the policy for policies with greater than the utility of the uniform policy, and decreasing for policies with less than the utility of the uniform policy. I think you can prove this via the property that the set of maxent policies is (very nearly) just Boltzman policies with varying temperature. But I would have to sit down and think about it properly. I should probably add that to the paper if it’s the case.

Thanks for this. The proof is indeed nonsense, but I think the proposition is still true. I’ve corrected it to this.

This link doesn’t work for me:

Thanks, it seems like the link got updated. Fixed!

Thanks for writing this up! Having not read the paper, I am wondering if in your opinion there’s a potential connection between this type of work and comp mech type of analysis/point of view? Even if it doesn’t fit in a concrete way right now, maybe there’s room to extend/modify things to combine things in a fruitful way? Any thoughts?

Here’s my current take, I wrote it as a separate shortform because it got too long. Thanks for prompting me to think about this :)

EDIT: I no longer think this setup is viable, for reasons that connect to why I think Critch’s operationalization is incomplete and why boundaries should ultimately be grounded in Pearlian Causality and interventions. Checkupdate.I believe there’s nothing much in the way of actually implementing an approximation of Critch’sboundaries^{[1]}using deep learning.Recall, Critch’s boundaries are:

Given a world (markovian stochastic process) Wt, map its values W (vector) bijectively using f into ‘features’ that can be split into four vectors each representing a boundary-possessing system’s Viscera, Active Boundary, Passive Boundary, and Environment.

Then, we characterize boundary-ness (i.e. minimal information flow across features unmediated by a boundary) using two mutual information criterion each representing infiltration and exfiltration of information.

And a policy of the boundary-posessing system (under the ‘stance’ of viewing the world implied by f) can be viewed as a stochastic map (that has no infiltration/exfiltration by definition) that best approximates the true Wt dynamics.

The interpretation here (under low exfiltration and infiltration) is that f can be viewed as a policy taken by the system in order to perpetuate its boundary-ness into the future and continue being well-described as a boundary-posessing system.

All of this seems easily implementable using very basic techniques from deep learning!

Bijective feature map are implemented using two NN maps each way, with an autoencoder loss.

Mutual information is approximated with standard variational approximations. Optimize f to minimize it.

(the interpretation here being—we’re optimizing our ‘stance’ towards the world in a way that best views the world as a boundary-possessing system)

After you train your ‘stance’ using the above setup, learn the policy using an NN with standard SGD, with fixed f.

A very basic experiment would look something like:

Test the above setup on two cellular automata (e.g., GoL, Lenia, etc) systems, one containing just random ash, and the other some boundary-like structure like noise-resistant glider structures found via optimization (there are a lot of such examples in the Lenia literature).

^{[2]}Then (1) check if the infiltration/exfiltration values are lower for the latter system, and (2) do some interp to see if the V/A/P/E features or the learned policy NN have any interesting structures.

I’m not sure if I’d be working on this any time soon, but posting the idea here just in case people have feedback.

I think research on boundaries—both conceptual work and developing practical algorithms for approximating them & schemes involving them—are quite important for alignment for reasons discussed earlier in my shortform.

Ultimately we want our setup to detect boundaries that aren’t just physically contiguous chunks of matter, like informational boundaries, so we want to make sure our algorithm isn’t just always exploiting basic locality heuristics.

I can’t think of a good toy testbed (ideas appreciated!), but one easy thing to try is to just destroy all locality by mapping the automata lattice (which we were feeding as input) with the output of a complicated fixed bijective map over it, so that our system will have to

learnlocality if it turns out to be ausefulnotion in its attempt at viewing the system as a boundary.I don’t see much hope in capturing a technical definition that doesn’t fall out of some sort of game theory, and even the latter won’t directly work for boundaries as representation of respect for autonomy helpful for alignment (as it needs to apply to radically weaker parties).

Boundaries seem more like a landmark feature of human-like preferences that serves as a test case for whether toy models of preference are reasonable. If a moral theory insists on tiling the universe with something, it fails the test. Imperative to merge all agents fails the test unless the agents end up essentially reconstructed. And with computronium, we’d need to look at the shape of things it’s computing rather than at the computing substrate.

I think it’s plausible that the general concept of boundaries can possibly be characterized somewhat independently of preferences, but at the same time have boundary-preservation be a quality that agents mostly satisfy (discussion here. very unsure about this). I see Critch’s definition as a first iteration of an operationalization for boundaries in the general, somewhat-preference-independent sense.

But I do agree that ultimately all of this should tie back to game theory. I find Discovering Agents most promising in this regards, though there are still a lot of problems—some of which I suspect might be easier to solve if we treat systems-with-high-boundaryness as a sort of primitive for the kind-of-thing that we can associate agency and preferences with in the first place.

There are two different points here, boundaries as a formulation of agency, and boundaries as a major component of human values (which might be somewhat sufficient by itself for some alignment purposes). In the first role, boundaries are an acausal norm that many agents end up adopting, so that it’s natural to consider a notion of agency that implies boundaries (after the agent had an opportunity for sufficient reflection). But this use of boundaries is probably open to arbitrary ruthlessness, it’s not respect for autonomy of someone the powers that be wouldn’t sufficiently care about. Instead, boundaries would be a convenient primitive for describing interactions with other live players, a Schelling concept shared by agents in this sense.

The second role as an aspect of values expresses that the agent does care about autonomy of others outside game theoretic considerations, so it only ties back to game theory by similarity, or through the story of formation of such values that involved game theory. A general definition might be useful here, if pointing AIs at it could instill it into their values. But technical definitions don’t seem to work when you consider what happens if you try to protect humanity’s autonomy using a boundary according to such definitions. It’s like machine translation, the problem could well be well-defined, but impossible to formally specify, other than by gesturing at a learning process.

I no longer think the setup above is viable, for reasons that connect to why I think Critch’s operationalization is incomplete and why boundaries should ultimately be grounded in Pearlian Causality and interventions.

(Note: I am thinking as I’m writing, so this might be a bit rambly.)## The world-trajectory distribution is ambiguous.

Intuition: Why does a robust glider in Lenia intuitively feel like a system possessing boundary? Well, I imagine various situations that happen in the world (like bullets) and this pattern mostly stays stable in face of them.

Now, notice that the measure of infiltration/exfiltration depends on ϕ∈Δ(Wω), a distribution over world history. Infil(ϕ):=Aggt≥0MutWω∼ϕ((Vt+1,At+1);Et∣(Vt,At,Pt))

So, for the above measure to capture my intuition, the approximate Markov condition (operationalized by low infil & exfil) must consider the world state Wω that contains the Lenia pattern with it avoiding bullets.

Remember, W is the

rawworld state, no coarse graining. So ϕ is the distribution over the raw world trajectory. It already captures all the “potentially occurring trajectories under which the system may take boundary-preserving-action.” Since everything is observed, our distribution already encodes all of “Nature’s Intervention.” So in some sense Critch’s definition is already causal (in a very trivial sense), by the virtue of requiring a distribution over the raw world trajectory, despite mentioning no Pearlian Causality.## Issue: Choice of ϕ

Maybe there is some canonical true ϕ for our physical world that minds can intersubjectively arrive at, so there’s no ambiguity.

But when I imagine trying to implement this scheme on Lenia, there’s immediately an ambiguity as to which distribution (representing my epistemic state on which raw world trajectories that will “actually happen”) we should choose:

Perhaps a very simple distribution: assigning uniform probability over world trajectories where the world contains nothing but the glider moving in a random direction with some initial point offset.

I suspect many stances other the one factorizing the world into gliders would have low infil/exfil, because the world is so simple. This is the case of “accidental boundary-ness.”

Perhaps something more complicated: various trajectories where e.g., the Lenia patterns encounters bullets, evolves with various other patterns, etc.

This I think rules out “accidental boundary-ness.”

I think the latter works. But now there’s a subjective choice of the distribution, and what are the set of possible/realistic “Nature’s Intervention”—all the situations that can ever be encountered by the system under which it has boundary-like behaviors—that we want to implicitly encode into our observational distribution. I don’t think it’s natural for ϕ assign much probability to a trajectory whose initial conditions are set in a very precise way such that everything decays into noise. But this feels quite subjective.

## Hints toward a solution: Causality

I think the discussion above hints at a very crucial insight:

ϕ

must arise as a consequence of the stable mechanisms in the world.Suppose the world of Lenia contains various stable mechanisms like a gun that shoots bullets at random directions, scarce food sources, etc.

We want ϕ to describe distributions that the boundary system will “actually” experience in some sense. I want the “Lenia pattern dodges bullet” world trajectory to be considered, because there

isa plausible mechanism in the world that can cause such trajectories to exist. For similar reasons, I think the empty world distributions are impoverished, and a distribution containing trajectories where the entire world decays into noise is bad because no mechanism can implement it.Thus, unless you have a canonical choice ofϕ, a better starting point would be to consider theabstract causal modelthat encodes the stable mechanisms in the world, and using Discovering Agents-style interventional algorithms that operationalize the notion “boundaries causally separate environment and viscera.”Well, because of everything mentioned above on how the causal model informs us on which trajectories are realistic, especially in the absence of a canonical ϕ. It’s also far more efficient, because the knowledge of the mechanism informs the algorithm of the precise interventions to query the world for, instead of having to implicitly bake them in ϕ.

There are still a lot more questions, but I think this is a pretty clarifying answer as to how Critch’s boundaries are limiting and why DA-style causal methods will be important.

I think the update makes sense in general, isn’t there however some way mutual information and causality is linked? Maybe it isn’t strong enough for there to be an easy extrapolation from one to the other.

Also I just wanted to drop this to see if you find it interesting, kind of on this topic? Im not sure its fully defined in a causality based way but it is about structure preservation.

https://youtu.be/1tT0pFAE36c?si=yv6mbswVpMiywQx9

Active Inference people also have the boundary problem as core in their work so they have some interesting stuff on it.

Yeah I’d like to know if there’s a unified way of thinking about information theoretic quantities and causal quantities, though a quick literature search doesn’t show up anything interesting. My guess is that we’d want separate boundary metrics for informational separation and causal separation.

Perhaps I should

~~one day in the far far future~~write a sequence on bayes nets.Some low-effort TOC (this is basically mostly koller & friedman):

why bayesnets and markovnets? factorized cognition, how to do efficient bayesian updates in practice, it’s how our brain is probably organized, etc. why would anyone want to study this subject if they’re doing alignment research? explain philosophy behind them.

simple examples of bayes nets. basic factorization theorems (the I-map stuff and separation criterion)

tangent on why bayes nets aren’t causal nets, though Zack M Davis had a good post on this exact topic, comment threads there are high insight

how inference is basically marginalization (basic theorems of: a reduced markov net represents conditioning, thus inference upon conditioning is the same as marginalization on a reduced net)

why is marginalization hard? i.e. NP-completeness of exact and approximate inference worst-case

what is a workaround? solve by hand simple cases in which inference can be greatly simplified by just shuffling in the order of sums and products, and realize that the exponential blowup of complexity is dependent on a graphical property of your bayesnet called the treewidth

exact inference algorithms (bounded by treewidth) that can exploit the graph structure and do inference efficiently: sum-product / belief-propagation

approximate inference algorithms (works in even high treewidth! no guarantee of convergence) - loopy belief propagation, variational methods, etc

connections to neuroscience: “the human brain is just doing belief propagation over a bayes net whose variables are the cortical column” or smth, i just know that there is some connection

Any thoughts on how to customize LessWrong to make it LessAddictive? I just really, really like the editor for various reasons, so I usually write a bunch (drafts, research notes, study notes, etc) using it but it’s quite easy to get distracted.

You could use the ad & content blocker uBlock Origin to zap any addictive elements of the site, like the main page feed or the Quick Takes or Popular Comments. Then if you

dowant to access these, you can temporarily turn off uBlock Origin.Incidentally, uBlock Origin can also be installed on mobile Firefox, and you can manually sync its settings across devices.

Maybe make a habit of blocking

`https://www.lesswrong.com/posts/*`

while writing?moments of microscopic fun encountered while studying/researching:

Quantum mechanics call vector space & its dual bra/ket because … bra-c-ket. What can I say? I like it—But where did the letter ‘c’ go, Dirac?

Defining cauchy sequences and limits in real analysis: it’s really cool how you “bootstrap” the definition of Cauchy sequences / limit on real using the definition of Cauchy sequences / limit on rationals. basically:

(1) define Cauchy sequence on rationals

(2) use it to define limit (on rationals) using rational-Cauchy

(3) use it to define reals

(4) use it to define Cauchy sequence on reals

(5) show it’s consistent with Cauchy sequence on rationals in both directions

a. rationals are embedded in reals hence the real-Cauchy definition subsumes rational-Cauchy definition

b. you can always find a rational number smaller than a given real number hence a sequence being rational-Cauchy means it is also real-Cauchy)

(6) define limit (on reals)

(7) show it’s consistent with limit on rationals

(8) … and that they’re equivalent to real-Cauchy

(9) proceed to ignore the distinction b/w real-Cauchy/limit and their rational counterpart. Slick!

(will probably keep updating this in the replies)

Maybe he dropped the “c” because it changes the “a” phoneme from æ to ɑː and gives a cleaner division in sounds: “brac-ket” pronounced together collides with “bracket” where “braa-ket” does not.

Any advice on reducing neck and shoulder pain while studying? For me that’s my biggest blocker to being able to focus longer (especially for math, where I have to look down at my notes/book for a long period of time). I’m considering stuff like getting a standing desk or doing regular back/shoulder exercises. Would like to hear what everyone else’s setups are.

Train skill of noticing tension and focus on it. Tends to dissolve. No that’s not so satisfying but it works. Standing desk can help but it’s just not that comfortable for most.

weight training?

I still have lots of neck and shoulder tension, but the only thing I’ve found that can reliably lessen it is doing some hard work on a punching bag for about 20 minutes every day, especially hard straights and jabs with full extension.

I’ve used Pain Science in the past as a resource and highly, highly endorse it. Here is an article they have on neck pain.

(Quality: Low, only read when you have nothing better to do—also not much citing)30-minute high-LLM-temp stream-of-consciousness on “How do we make mechanistic interpretability work for non-transformers, or just any architectures?”

We want a general way to reverse engineer circuits

e.g., Should be able to rediscover properties we discovered from transformers

Concrete Example: we spent a bunch of effort reverse engineering transformer-type architectures—then boom, suddenly some parallel-GPU-friendly-LSTM architecutre turns out to have better scaling properties, and everyone starts using it. LSTMs have different inductive biases, like things in the same layer being able to communicate multiple times with each other (unlike transformers), which incentivizes e.g., reusing components (more search-y?).

Formalize:

You have task X. You train a model A with inductive bias I_A. You also train a model B with inductive bias I_B. Your mechanistic interpretability techniques work well on deciphering A, but not B. You want your mechanistic interpretability techniques to work well for B, too.

Proposal: Communication channel

Train a Transformer on task X

Existing Mechanistic interpretability work does well on interpreting this architecture

Somehow stitch the LSTM to the transformer (?)

I’m trying to get at to the idea of “interface conversion,” that by the virtue of SGD being greedy, it will try to convert the outputs of transformer-friendly types

Now you can better understand the intermediate outputs of the LSTM by just running mechanistic interpretability on the transformer layers whose input are from the LSTM

(I don’t know if I’m making any sense here, my LLM temp is > 1)

Proposal: approximation via large models?

Train a larger transformer architecture to approximate the smaller LSTM model (either just input output pairs, or intermediate features, or intermediate features across multiple time-steps, etc):

the basic idea is that a smaller model would be more subject to following its natural gradient shaped by the inductive bias, while larger model (with direct access to the intermediate outputs of the smaller model) would be able to approximate it despite not having as much inductive bias incentive towards it.

probably false but illustrative example: Train small LSTM on chess. By the virtue of being able to run serial computation on same layers, it focuses on algorithms that have repeating modular parts. In contrast, a small Transformer would learn algorithms that don’t have such repeating modular parts. But instead, train a large transformer to “approximate” the small LSTM—it should be able to do so by, e.g., inefficiently having identical modules across multiple layers. Now use mechanistic interpretability on that.

Proposal: redirect GPS?

Thane’s value formation picture says GPS should be incentivized to reverse-engineer the heuristics because it has access to inter-heuristic communication channel. Maybe, in the middle of training, gradually swap different parts of the model with those that have different inductive biases, see GPS gradually learn to reverse-engineer those, and mechanistically-interpret how GPS exactly does that, and reimplement in human code?

Proposal: Interpretability techniques based on behavioral constraints

e.g., Discovering Latent Knowledge without Supervision, putting constraints?

How to do we “back out” inductive biases, just given e.g., architecture, training setup? What is the type signature?

(I need to read more literature)

I am curious as to how often the asymptotic results proven using features of the problem that seem basically practically-irrelevant become relevant in practice.

Like, I understand that there are many asymptotic results (e.g., free energy principle in SLT) that are useful in practice, but i feel like there’s something sus about similar results from information theory or complexity theory where the way in which they prove certain bounds (or inclusion relationship, for complexity theory) seem totally detached from practicality?

joint source coding theoremis often stated as why we can consider the problem of compression and redundancy separately, but when you actually look at the proof it only talks about possibility (which is proven in terms of insanely long codes) and thus not-at-all trivial that this equivalence is something that holds in the context of practical code-engineeringcomplexity theorytalks about stuff like quantifying some property over all possible boolean circuits of a given size which seems to me considering a feature of the problem just soutterlyirrelevant to real programs that I’m suspicious it can say meaningful things about stuff we see in practiceas an aside, does the P vs NP distinction even matter in practice? we just … seem to have very good approximation to NP problems by algorithms that take into account the structures specific to the problem and domains where we want things to be fast; and as long as complexity methods doesn’t take into account those fine structures that are specific to a problem, i don’t see how it would characterize such well-approximated problems using complexity classes.

Wigderson’s book had a short section on average complexity which I hoped would be this kind of a result, and I’m unimpressed (the problem doesn’t sound easier—now how do you specify the natural distribution??)

P v NP: https://en.wikipedia.org/wiki/Generic-case_complexity

One result to mention in computational complexity is the PCP theorem which not only gives probabilistically checkable proofs but also gives approximation case hardness. Seems deep but I haven’t understood the proof yet.

Great question. I don’t have a satisfying answer. Perhaps a cynical answer is survival bias—we remember the asymptotic results that eventually become relevant (because people develop practical algorithms or a deeper theory is discovered) but don’t remember the irrelevant ones.

Existence results are categorically easier to prove than explicit algorithms. Indeed, classical existence may hold (the former) while intuitioinistically (the latter) might not. We would expect non-explicit existence results to appear before explicit algorithms.

One minor remark on ‘quantifying over all boolean algorithms’. Unease with quantification over large domains may be a vestige of set-theoretic thinking that imagines types as (platonic) boxes. But a term of a for-all quantifier is better thought of as an algorithm/ method to check the property for any given term (in this case a Boolean circuit). This doesn’t sound divorced from practice to my ears.

Yes, it does, for several reasons:

It basically is necessary to prove P != NP to get a lot of other results to work, and for some of those results, proving P != NP is sufficient.

If P != NP (As most people suspect), it fundamentally rules out solving lots of problems generally and quickly without exploiting structure, and in particular lets me flip the burden of proof to the algorithm maker to explain why their solution to a problem like SAT is efficient, rather than me having to disprove the existence of an efficient algorithm.

It’s either by exploiting structure, somehow having a proof that P=NP, or relying on new physics models that enable computing NP-complete problems efficiently, and the latter 2 need very, very strong evidence behind them.

This in particular applies to basically all learning problems in AI today.

It explains why certain problems cannot be reasonably solved optimally, without huge discoveries, and the best examples are travelling salesman problems for inability to optimally solve, as well as a whole lot of other NP-complete problems. There are also other NP problems where there isn’t a way to solve them efficiently at all, especially if FPT != W[1] holds.

Also a note that we also expect a lot of NP-complete problems to also not be solvable by fast algorithms even in the average case, which basically means it’s likely to be very relevant quite a lot of the time, so we don’t have to limit ourselves to the worst case either.

I recently learned about metauni, and it looks amazing. TL;DR, a bunch of researchers give out lectures or seminars on Roblox—Topics include AI alignment/policy, Natural Abstractions, Topos Theory, Singular Learning Theory, etc.

I haven’t actually participated in any of their live events yet and only watched their videos, but they all look really interesting. I’m somewhat surprised that there hasn’t been much discussion about this on LW!

Discovering agents provide a genuine causal, interventionist account of agency and an algorithm to detect them, motivated by the intentional stance. I find this paper

veryenlightening from a conceptual perspective!I’ve tried to think of problems that needed to be solved before we can actually implement this on real systems—both conceptual and practical—on approximate order of importance.

There are no ‘dynamics,’ no learning.As soon as a mechanism node is edited, it is assumed that agents immediately change their ‘object decision variable’ (a conditional probability distribution given its object parent nodes) to play the subgame equilibria.Assumption of factorization of variables into ‘object’ / ‘mechanisms,’ and the resulting subjectivity.The paper models the process by which an agent adapts its policy given changes in the mechanism of the environment via a ‘mechanism decision variable’ (that depends on its mechanism parent nodes), which modulates the conditional probability distribution of its child ‘object decision variable’, the actual policy.For example, the paper says a learned RL policy isn’t an agent, because interventions in the environment won’t make it change its already-learned policy—but that a human or a RL policy

togetherwith its training processisan agent, because itcanadapt. Is this reasonable?Say I have a gridworld RL policy that’s learned to get cheese (3 cell world, cheese always on left) by always going to the left. Clearly it can’t change its policy when I change the cheese distribution to favor right, so it seems right to call this not an agent.

Now, say the policy now has sensory access to the grid state, and correctly generalized (despite only being trained on left-cheese) to move in the direction where it sees the cheese, so when I change the cheese distribution, it adapts accordingly. I think it is right to call this an agent?

Now, say the policy is an LLM agent (static weight) on an open world simulation which reasons in-context. I just changed the mechanism of the simulation by lowering the gravity constant, and the agent observes this, reasons in-context, and adapts its sensorimotor policy accordingly. This is clearly an agent?

I think this is because the paper considers, in the case of the RL policy alone, the ‘object policy’ to be the policy of the trained neural network (whose induced policy distribution is definitionally fixed), and the ‘mechanism policy’ to be a trivial delta function assigning the already-trained object policy. And in the case of the RL policy together with its training process, the ‘mechanism policy’ is now defined as the training process that assigns the fully-trained conditional probability distribution to the object policy.

But what if the ‘mechanism policy’ was the in-context learning process by which it induces an ‘object policy’? Then changes in the environment’s mechanism can be related to the ‘mechanism policy’ and thus the ‘object policy’ via in-context learning as in the second and third example, making them count as agents.

Ultimately, the setup in the paper forces us to factorize the means-by-which-policies-adapt into mechanism vs object variables, and the results (like whether a system is to be considered an agent) depends on this factorization. It’s not always clear what the right factorization is, how to discover them from data, or if this is the right frame to think about the problem at all.

Implicit choice of variables that are convenient for agent discovery.The paper does mention that the algorithm is dependent in the choice of the variable, as in: if the node corresponding to the ‘actual agent decision’ is missing but its children is there, then the algorithm will label its children to be the decision nodes. But this is already a very convenient representation!Prototypical example: Minecraft world with RL agents interacting represented as a coarse-grained lattice (dynamical Bayes Net?) with each node corresponding to a physical location and its property, like color. Clearly no single node here is an agent, because agents move! My naive guess is that in principle, everything will be labeled an agent.

So the variables of choice must be abstract variables of the underlying substrate, like functions over them. But then, how do you discover the right representation automatically, in a way that interventions in the abstract variable level can faithfully translate to actually performable interventions in the underlying substrate?

Given the causal graph, even the slightest satisfaction of the agency-criterion labels the nodes as decision / utility.

No “degree-of-agency”—maybe by summing over the extent to which the independencies fail to satisfy?Then different agents are defined as causally separated chunks (~connected component) of [set-of-decision-nodes / set-of-utility-nodes].

How do we accommodate hierarchical agency (like subagents), systems with different degrees of agency, etc?The interventional distribution on the object/mechanism variables are converted into a causal graph using the obvious [perform-do()-while-fixing-everything-else] algorithm. My impression is that

causal discovery doesn’t really work in practice, especially in noisy reality with a large number of variables via gazillion conditional independence tests.The correctness proof requires lots of unrealistic assumptions, e.g.,

agents always play subgame equilibria, though I think some of this can be relaxed.Complaint with Pugh’s real analysis textbook: He doesn’t even define the limit of a function properly?!

It’s implicitly defined together with the definition of continuity where ∀ϵ>0∃δ>0|x−x0|<δ⟹|f(x)−f(x0)|<ϵ, but in Chapter 3 when defining differentiability he implicitly switches the condition to 0<|x−x0|<δ without even mentioning it (nor the requirement that x0 now needs to be an accumulation point!) While Pugh has its own benefits, coming from Terry Tao’s analysis textbook background, this is

absurd!(though to be fair Terry Tao has the exact same issue in Book 2, where his definition of function continuity via limit in metric space precedes that of defining limit in general … the only redeeming factor is that it’s defined rigorously in Book 1, in the limited context of R)

*sigh* I guess we’re still pretty far from reaching the Pareto Frontier of textbook quality, at least in real analysis.

… Speaking of Pareto Frontiers, would anyone say there

issuch a textbook that is close to that frontier, at least in a different subject? Would love to read one of those.Maybe you should email Pugh with the feedback? (I audited his honors analysis course in fall 2017; he seemed nice.)

As far as the frontier of analysis textbooks goes, I really like how Schröder

Mathematical Analysismanages to be both rigorous and friendly: the early chapters patiently explain standard proof techniques (like the add-and-subtract triangle inequality gambit) to the novice who hasn’t seen them before, but the punishing details of the subject are in no way simplified. (One wonders if the subtitle “A Concise Introduction” was intended ironically.)I used to try out near-random search on ideaspace, where I made a quick app that spat out 3~5 random words from a dictionary of interesting words/concepts that I curated, and I spent 5 minutes every day thinking

very hardon whether anything interesting came out of those combinations.Of course I knew random search on exponential space was futile, but I got a couple cool invention ideas (most of which turned out to already exist), like:

infinite indoor rockclimbing: attach rocks to a vertical treadmill, and now you have an infinite indoor rock climbing wall (which is also safe from falling)! maybe add some fancy mechanism to add variations to the rocks + a VR headgear, I guess.

clever crypto mechanism design (in the spirit of CO2 Coin) to incentivize crowdsourcing of age-reduction molecule design animal trials from the public. (I know what you’re thinking)

You can probably do this smarter now if you wanted, with eg better GPT models.

Having lived ~19 years, I can distinctly remember around 5~6 times when I explicitly noticed myself experiencing totally new qualia with my inner monologue going “oh wow! I didn’t know this dimension of qualia was a thing.” examples:

hard-to-explain sense that my mind is expanding horizontally with fractal cube-like structures (think bismuth) forming around it and my subjective experience gliding along its surface which lasted for ~5 minutes after taking zolpidem for the first time to sleep (2 days ago)

getting drunk for the first time (half a year ago)

feeling absolutely euphoric after having a cool math insight (a year ago)

...

Reminds me of myself around a decade ago, completely incapable of understanding why my uncle smoked, being “huh? The smoke isn’t even sweet, why would you want to do that?” Now that I have [addiction-to-X] as a clear dimension of qualia/experience solidified in myself, I can better model their subjective experiences although I’ve never smoked myself. Reminds me of the SSC classic.

Also one observation is that it feels like the rate at which I acquire these is getting faster, probably because of increase in self-awareness + increased option space as I reach adulthood (like being able to drink).

Anyways, I think it’s really cool, and can’t wait for more.

I observed new visual qualia of colors while using some light machine.

Also, when I first came to Italy, I have a feeling as if the whole rainbow of color qualia changed

Sunlight scattered by the atmosphere on cloudless mornings during the hour before sunrise inspires a subtle feeling (“this is cool, maybe even exciting”) that I never noticed till I started intentionally exposing myself to it for health reasons (specifically, making it easier to fall asleep 18 hours later).

More precisely, I might or might not have noticed the feeling, but if I did notice it, I quickly forgot about it because I had no idea how to reproduce it.

I have to get away from artificial light (streetlamps) (and from

direct(yellow) sunlight) for the (blue) indirect sunlight to have this effect. Also, it is no good looking at a small patch of sky, e.g., through a window in a building: most or all of the upper half of my field of vision must be receiving this indirect sunlight. (The intrinsically-photosensitive retinal ganglion cells are all over the bottom half of the retina, but absent from the top half.)To me, the fact that the human brain basically implements SSL+RL is

very verystrong evidence that the current DL paradigm (with a bit of “engineering” effort, but nothing like fundamental breakthroughs) will kinda just keep scaling until we reach point-of-no-return. Does this broadly look correct to people here? Would really appreciate other perspectives.I mostly think “algorithms that involve both SSL and RL” is a much broader space of possible algorithms than you seem to think it is, and thus that there are parts of this broad space that require “fundamental breakthroughs” to access. For example, both AlexNet and differentiable rendering can be used to analyze images via supervised learning with gradient descent. But those two algorithms are very very different from each other! So there’s more to an algorithm than its update rule.

See also 2nd section of this comment, although I was emphasizing alignment-relevant differences there whereas you’re talking about capabilities. Other things include the fact that if I ask you to solve a hard math problem, your brain will be different (different

weights, not just different activations / context) when you’re halfway through compared to when you started working on it (a.k.a. online learning, see also here), and the fact that brain neural networks are not really “deep” in the DL sense. Among other things.Makes sense. I think we’re using the terms differently in scope. By “DL paradigm” I meant to encompass the kind of stuff you mentioned (RL-directing-SS-target (active learning), online learning, different architecture, etc) because they

reallyseemed like “engineering challenges” to me (despite them covering a broad space of algorithms) in the sense that capabilities researchers already seem to be working on & scaling them without facing any apparent blockers to further progress, i.e. in need of any “fundamental breakthroughs”—by which I was pointing more at paradigm shifts away from DL like, idk, symbolic learning.I have a slightly different takeaway. Yes techniques similar to current techniques will most likely lead to AGI but it’s not literally ‘just scaling LLMs’. The actual architecture of the brain is meaningfully different from what’s being deployed right now. So different in one sense. On the other hand it’s not like the brain does something completely different and proposals that are much closer to the brain architecture are in the literature (I won’t name them here...). It’s plausible that some variant on that will lead to true AGI. Pure hardware scaling obviously increases capabilities in a straightforward way but a transformer is not a generally intelligent agent and won’t be even if scaled many more OOMs.

(I think Steven Byrnes has a similar view but I wouldn’t want to misrepresent his views)

So far as I can tell, a transformer has three possible blockers (that would need to stand undefeated

together): (1) in-context learning plateauing at a level where it’s not able to do even a little bit of useful work without changing model weights, (2) terrible sample efficiency that asks for more data than is available on new or rare/situational topics, and (3) absence of a synthetic data generation process that’s both sufficiently prolific and known not to be useless at that scale.A need for online learning and terrible sample efficiency are defeated by OOMs if enough useful synthetic data can be generated, which the anemic in-context learning without changing weights might turn out to be sufficient for. This is the case of defeating (3), with others falling as a result.

Another possibility is that much larger multimodal transformers (there is a lot of video) might suffice without synthetic data if a model learns superintelligent in-context learning. SSL is not just about imitating humans, the problems it potentially becomes adept at solving are arbitrarily intricate. So even if it can’t grow further and learn substantially new things within its current architecture/model, it might happen to already be far enough along at inference time to do the necessary redesign on its own. This is the case of defeating (1), leaving it to the model to defeat the others. And it should help with (3) even at non-superintelligent levels.

Failing that, RL demonstrates human level sample efficiency in increasingly non-toy settings, promising that saner amounts of useful synthetic data might suffice, defeating (2), though at this point it’s substantially not-a-transformer.

generating useful synthetic data and solving novel tasks with little correlation with training data is the exact issue here. Seems straightforwardly true that a transformer arcthiecture doesn’t do that?

I don’t know what superintelligent in-context learning is—I’d be skeptical that scaling a transformer a further 3 OOMS will suddenly make it do tasks that are very far from the text distribution it is trained on, indeed solutions to tasks that are not even remotely in the internet text data like building a recursively self-improving agent (if such a thing is possible...)? Maybe I’m misunderstanding what you’re claiming here.

Not saying it’s impossible, just seems deeply implausible. ofc LLMs being so impressive was also a prior implausible but this seems another OOM of implausibility bits if that makes sense?

I’m imagining some prompts to generate reasoning, inferred claims about the world. You can’t generate new observations about the world, but you can reason about the observations available so far, and having those inferred claims in the dataset likely helps, that’s how humans build intuition about theory. If an average a 1000 inferred claims are generated for every naturally observed statement (or just those on rare/new/situational topics), that could close the gap of sample efficiency with humans. Might take the form of exercises or essays or something.

If this is all done with prompts, using a sufficiently smart order-following chatbot, then it’s straightforwardly just a transformer, with some superficial scaffolding. If this can work, it’ll eventually appear in distillation literature, though I’m not sure if serious effort to check was actually made with current SOTA LLMs, to pre-train exclusively on synthetic data that’s not too simplistically prompted. Possibly you get nothing for a GPT-3 level generator, and then something for GPT-4+, because reasoning needs to be good enough to preserve contact with ground truth. From Altman’s comments I get the impression that it’s plausibly the exact thing OpenAI is hoping for.

In-context learning is capability to make use of novel data that’s only seen in a context, not in pre-training, to do tasks that make use of this novel data, in ways that normally would’ve been expected to require it being seen in pre-training. In-context learning is a model capability, it’s learned. So its properties are not capped by those of the hardcoded model training algorithm, notably in principle in-context learning could have higher sample efficiency (which might be crucial for generating a lot of synthetic data out of a few rare observations). Right now it’s worse in most respects, but that could change with scale without substantially modifying the transformer architecture, which is the premise of this thread.

By superintelligent in-context learning I mean the capabilities of in-context learning significantly exceeding those of humans. Things like fully comprehending a new paper without changing any model weights, becoming able to immediately write the next one in the same context window. I agree that it’s not very plausible, and probably can’t happen without sufficiently deep circuits, which even deep networks don’t seem to normally develop. But it’s not really ruled out by anything that’s been tried so far. Recent stuff on essentially pre-training with some frozen weights without losing resulting performance suggests a trend of increasing feasible model size for given compute. So I’m not sure this can’t be done in a few years. Then there’s things like memory transformers, handing a lot more data than a context to a learned learning capability.

I wonder if the following is possible to study textbooks more efficiently using LLMs:

Feed the entire textbook to the LLM and produce a list of summaries that increases in granularity and length, covering all the material in the textbook just at a different depth (eg proofs omitted, further elaboration on high-level perspectives, etc)

The student starts from the highest-level summary, and gradually moves to the more granular materials.

When I study textbooks, I spend a significant amount of time improving my mental autocompletion, like being able to familiarize myself with the terminologies, which words or proof-style usually come in which context, etc. Doing this seems to significantly improve my ability to read eg long proofs, since I can ignore all the pesky details (which I can trust my mental autocompletion to later fill in the details if needed) and allocate my effort in getting a high-level view of the proof.

Textbooks don’t really admit this style of learning, because the students don’t have prior knowledge of all the concept-dependencies of a new subject they’re learning, and thus are forced to start at the lowest-level and make their way up to the high-level perspective.

Perhaps LLMs will let us reverse this direction, instead going from the highest to the lowest.

What’s a good technical introduction to Decision Theory and Game Theory for alignment researchers? I’m guessing standard undergrad textbooks don’t include, say, content about logical decision theory. I’ve mostly been reading posts on LW but as with most stuff here they feel more like self-contained blog posts (rather than textbooks that build on top of a common context) so I was wondering if there was anything like a canonical resource providing a unified technical / math-y perspective on the whole subject.

The MIRI Research Guide recommends An Introduction to Decision Theory and Game Theory: An Introduction. I have read neither and am simply relaying the recommendation.

i absolutely hate bureaucracy, dumb forms, stupid websites etc. like, I almost had a literal breakdown trying to install Minecraft recently (and eventually failed). God.

I think what’s so crushing about it, is that it reminds me that the wrong people are designing things, and that they wont allow them to be fixed, and I can only find solace in thinking that the inefficiency of their designs is also a sign that they can be defeated.

God, I wish real analysis was at least half as elegant as any other math subject — way too much pathological examples that I can’t care less about. I’ve heard some good things about constructivism though, hopefully analysis is done better there.

Yeah, real analysis sucks. But you have to go through it to get to delightful stuff— I particularly love harmonic and functional analysis. Real analysis is just a bunch of pathological cases and technical persnicketiness that you need to have to keep you from steering over a cliff when you get to the more advanced stuff. I’ve encountered some other subjects that have the same feeling to them. For example, measure-theoretic probability is a dry technical subject that you need to get through before you get the fun of stochastic differential equations. Same with commutative algebra and algebraic geometry, or point-set topology and differential geometry.

Constructivism, in my experience, makes real analysis more mind blowing, but also harder to reason about. My brain uses non-constructive methods subconsciously, so it’s hard for me to notice when I’ve transgressed the rules of constructivism.

As a general reflection on undergraduate mathematics imho there is way too much emphasis on real analysis. Yes, knowing how to be rigorous is important, being aware of pathological counterexample is importanting, and real analysis is used all over the place. But there is so much more to learn in mathematics than real analysis and the focus on minor technical issues here is often a distraction to developing a broad & deep mathematical background.

For most mathematicians (and scientists using serious math) real analysis is a only a small part of the toolkit. Understanding well the different kinds of limits can ofc be crucial in functional analysis, stochastic processes and various parts of physics. But there are so many topics that are important to know and learn here!

The reason it is so prominent in the undergraduate curriculum seems to be more tied to institutional inertia, its prominence on centralized exams, relation with calculus, etc

Really, what’s going on is that in the general case, as mathematics is asked to be more and more general, you will start encountering pathological examples more, and paying attention to detail more is a valuable skill in both math and real life.

And while being technical about the pathological cases is kind of annoying, it’s also one that actually matters in real life, as you aren’t guaranteed to have an elegant solution to your problems.

Update: huh, nonstandard analysis is really cool. Not only are things much more intuitive (by using infinitesimals from hyperreals instead of using epsilon-delta formulation for everything), by the transfer principle all first order statements are equivalent between standard and nonstandard analysis!

There were various notions/frames of optimization floating around, and I tried my best to distill them:

Eliezer’s Measuring Optimization Power on unlikelihood of outcome + agent preference ordering

Alex Flint’s The ground of optimization on robustness of system-as-a-whole evolution

Selection vs Control as distinguishing different types of “space of possibilities”

Selection as having that space explicitly given & selectable numerous times by the agent

Control as having that space only given in terms of counterfactuals, and the agent can access it only once.

These distinctions correlate with the type of algorithm being used & its internal structure, where Selection uses more search-like process using maps, while Control may just use explicit formula … although it may very well use internal maps to Select on counterfactual outcomes!

In other words, the Selection vs Control may very well be viewed as a

different cluster of Analysis. Example:If we decide to focus our Analysis of “space of possibilities” on eg “Real life outcome,” then a guided missile is always Control.

But if we decide to focus on “space of internal representation of possibilities,” then a guided missle that uses internal map to search on becomes Selection.

“Internal Optimization” vs “External Optimization”

Similar to Selection vs Control, but the analysis focuses more on internal structure:

Why? Motivated by the fact that, as with the guided missile example, Control systems can be viewed as Selection systems depending on perspective

… hence, better to focus on internal structures where it’s much less ambiguous.

IO: Internal search + selection

EO: Flint’s definition of “optimizing system”

IO is included in EO, if we assume accurate map-to-environment correspondence.

To me, this doesn’t really get at what the internals of actually-control-like systems look like, which presumably a subset of EO—IO.

Search-in-Territory vs Search-in-Map

Greater emphasis on internal structure—specifically, “maps.”

Maps are capital investment, allowing you to be able to optimize despite not knowing what to exactly optimize for (by compressing info)

I have several thoughts on these framings, but one trouble is the excessive usage of words to represent “clusters” i.e. terms to group a bunch of correlated variables. Selection vs Control, for example, doesn’t have a clear definition/criteria but rather points at a number of correlated things, like internal structure, search, maps, control-like things, etc.

Sure, deconfusing and pointing out clusters is useful because clusters imply correlations and correlations perhaps imply hidden structure + relationships—but I think the costs from cluster-representing-words doing hidden inference is much greater than the benefits, and it would be better to explicitly lay out the features-of-clusters that the one is referring to instead of just using the name of the cluster.

This is similar to the trouble I had with “wrapper-minds,” which is yet another example of a cluster pointing at a bunch of correlated variables, and people using the same term to mean different things.

Anyways, I still feel totally confused about optimization—and while these clusters/frames are useful, I think thinking in terms of them would ensue even more confusion within myself. It’s probably better to take the useful individual parts

withinthe cluster and start deconfusing from the ground-up using those as the building blocks.I find the intersection of computational mechanics, boundaries/frames/factored-sets, and some works from the causal incentives group—especially discovering agents and robust agents learn causal world model (review) - to be a very interesting theoretical direction.

By boundaries, I mean a sustaining/propagating system that informationally/causally insulates its ‘viscera’ from the ‘environment,’ and only allows relatively small amounts of deliberate information flow through certain channels in both directions. Living systems are an example of it (from bacteria to humans). It doesn’t even have to be a physically distinct chunk of spacetime, they can be over more abstract variables like societal norms. Agents are an example of it.

I find them very relevant to alignment especially from the direction of detecting such boundary-possessing/agent-like structures embedded in a large AI system and backing out a sparse relationship between these subsystems, which can then be used to e.g., control the overall dynamic. Check out these posts for more.

A prototypical deliverable would be an algorithm that can detect such ‘boundaries’ embedded in a dynamical system when given access to some representation of the system, performs observations & experiments and returns a summary data structure of all the ‘boundaries’ embedded in a system and their desires/wants, how they game-theoretically relate to one another (sparse causal relevance graph?), the consequences of interventions performed on them, etc—that’s versatile enough to detect e.g., gliders embedded in Game of Life / Particle Lenia, agents playing Minecraft while only given coarse grained access to the physical state of the world, boundary-like things inside LLMs, etc. (I’m inspired by this)

Why do I find the aforementioned directions relevant to this goal?

Critch’s Boundaries operationalizes boundaries/viscera/environment as functions of the underlying variable that executes policies that continuously prevents information ‘flow’

^{[1]}between disallowed channels, quantified via conditional transfer entropy.Relatedly, Fernando Rosas’s paper on Causal Blankets operationalize boundaries using a similar but subtly different

^{[2]}form of mutual information constraint on the boundaries/viscera/environment variables than that of Critch’s. Importantly, they show that such blankets always exist between two coupled stochastic processes (using a similar style of future morph equivalence relation characterization from compmech, and also a metric they call “synergistic coefficient” that quantifies how boundary-like this thing is.^{[3]}More on compmech, epsilon transducers generalize epsilon machines to input-output processes. PALO (Perception Action Loops) and Boundaries as two epsilon transducers coupled together?

These directions are interesting, but I find them still unsatisfactory because all of them are purely behavioral accounts of boundaries/agency. One of the hallmarks of agentic behavior (or some boundary behaviors) is adapting ones policy if an intervention changes the environment in a way that the system can observe and adapt to.

^{[4]}^{[5]}(is there an interventionist extension of compmech?)

Discovering agents provide a genuine causal, interventionist account of agency and an algorithm to detect them, motivated by the intentional stance. I think the paper is very enlightening from a conceptual perspective, but there are many problems yet to be solved before we can actually implement this. Here’s my take on it.

More fundamentally, (this is more vibes, I’m really out of my depth here) I feel there is something intrinsically limiting with the use of Bayes Nets, especially with the fact that choosing which variables to use in your Bayes Net already encodes a lot of information about the specific factorization structure of the world. I heard good things about finite factored sets and I’m eager to learn more about them.

Not exactly a ‘flow’, because transfer entropy conflates between intrinsic information flow and synergistic information—a ‘flow’ connotes only the intrinsic component, while transfer entropy just measures the overall amount of information that a system couldn’t have obtained on its own. But anyways, transfer entropy seems like a conceptually correct metric to use.

Specifically, Fernando’s paper criticizes blankets of the following form (V for viscera, A and P for active/passive boundaries, E for environment):

Vt→At,Pt→Et

DIP implies I(Vt;At,Pt)≥I(Vt;Et)

This clearly forbids dependencies formed in the past that stays in ‘memory’.

but Critch instead defines boundaries as satisfying the following two criteria:

Vt+1,At+1→Vt,At,Pt→Et (infiltration)

DIP implies I(Vt;At,Pt)≥I(Vt;Et)

Et+1,Pt+1→At,Pt,Et→Vt (exfiltration)

DIP implies I(Vt+1,At+1;At,Pt)≥I(Vt+1,At+1;Et)

and now that the independencies are entangled across different t, there is no longer a clear upper bound on I(Vt;Et), so I don’t think the criticisms apply directly.

My immediate curiosities are on how these two formalisms relate to one another. e.g., Which independency requirements are more conceptually ‘correct’? Can we extend the future-morph construction to construct Boundaries for Critch’s formalism? etc etc

For example, a rock is very goal-directed relative to ‘blocking-a-pipe-that-happens-to-exactly-match-its-size,’ until one performs an intervention on the pipe size to discover that it can’t adapt at all.

Also, interventions are really cheap to run on digital systems (e.g., LLMs, cellular automata, simulated environments)! Limiting oneself to behavioral accounts of agency would miss out on a rich source of cheap information.

Does anyone know if Shannon arrive at entropy from the axiomatic definition first, or the operational definition first?

I’ve been thinking about these two distinct ways in which we seem to arrive at new mathematical concepts, and looking at the countless partial information decomposition measures in the literature all derived/motivated based on an axiomatic basis, and not knowing which intuition to prioritize over which, I’ve been assigning less premium on axiomatic conceptual definitions than i used to:

decision theoretic justification of probability > Cox’s theorem

shannon entropy as min description length > three information axioms

fernando’s operational definition of synergistic information > rest of the literature with its countless non-operational PID measures

The basis of comparison would be its usefulness and ease-of-generalization to better concepts:

at least in the case of fernando’s synergistic information, it seems far more useful because i at least know what i’m exactly getting out of it, unlike having to compare between the axiomatic definitions based on handwavy judgements.

for ease of generalization, the problem with axiomatic definitions is that there are many logically equivalent ways to state the initial axiom (from which they can then be relaxed), and operational motivations seem to ground these equivalent characterizations better, like logical inductors from the decision theoretic view of probability theory

(obviously these two feed into each other)

I’m not sure what you mean by operational vs axiomatic definitions.

But Shannon was unaware of the usage of S=−Σi pi ln pi in statistical mechanics. Instead, he was inspired by Nyquist and Hartley’s work, which introduced ad-hoc definitions of information in the case of constant probability distributions.

And in his seminal paper, “A mathematical theory of communication”, he argued in the introduction for the logarithm as a measure of information because of practicality, intuition and mathematical convenience. Moreover, he explicitly derived the entropy of a distribution from three axioms:

1) that it be continuous wrt. the probabilities,

2) that it increase monotonically for larger systems w/ constant probability distributions,

3) and that it be a weighted sum the entropy of sub-systems.

See section 6 for more details.

I hope that answers your question.

‘Symmetry’ implies ‘redundant coordinate’ implies ‘cyclic coordinates in your Lagrangian / Hamiltonian’ implies ‘conservation of conjugate momentum’And because the action principle (where the true system trajectory extremizes your action, i.e. integral of Lagrangian) works in various dynamical systems, the above argument works in non-physical dynamical systems.

Thus conserved quantities usually exist in a given dynamical system.

mmm, but why does the action principle hold in such a wide variety of systems though? (like how you get entropy by postulating something to be maximized in an equilibrium setting)

Mildly surprised how some verbs/connectives barely play any role in conversations, even in technical ones. I just tried directed babbling with someone, and (I think?) I learned quite a lot about Israel-Pakistan relations with almost no stress coming from eg needing to make my sentences grammatically correct.

Example of (a small part of) my attempt to summarize my understanding of how Jews migrated in/out of Jerusalem over the course of history:

Could you explain more what you mean by this?

My (completely amateur) understanding is that the “extra” semantic and syntactic structure of written and spoken language does two things.

One, it adds redundancy and reduces error. Simple example, gendered pronouns mean that when you hear “Have you seen Laurence? She didn’t get much sleep last night.” you have a chance to ask the speaker for clarification and catch if they had actually said “Laura” and you misheard.

Two, it can be used as a signal. The correct use of jargon is used by listeners or readers as a proxy for competence. Or many typos in your text will indicate to readers that you haven’t put much effort into what you’re saying.

Why haven’t mosquitos evolved to be less itchy? Is there just not enough selection pressure posed by humans yet? (yes probably) Or

arethey evolving towards that direction? (they of course already evolved towards being less itchy while biting, but not enough to make that lack-of-itch permanent)~~this is a request for help i’ve been trying and failing to catch this one for god knows how long plz halp~~tbh would be somewhat content coexisting with them (at the level of houseflies) as long as they evolved the itch and high-pitch noise away, modulo disease risk considerations.

The reason mosquito bites itch is because they are injecting saliva into your skin. Saliva contains mosquito antigens, foreign particles that your body has evolved to attack with an inflammatory immune response that causes itching. The compound histamine is a key signaling molecule used by your body to drive this reaction.

In order for the mosquito to avoid provoking this reaction, they would either have to avoid leaving compounds inside of your body, or mutate those compounds so that they do not provoke an immune response. The human immune system is an adversarial opponent designed with an ability to recognize foreign particles generally. If it was tractable for organisms to reliably evolve to avoid provoking this response, that would represent a fundamental vulnerability in the human immune system.

Mosquitoe saliva

doesin fact contain anti-inflammatory, antihemostatic, and immunomodulatory compounds. So they’re trying! But also this means that mosquitos are evolvedtoput saliva inside of you when they feed, which means they’re inevitably going to expose the foreign particles they produce to your immune system.There’s also a facet of selection bias making mosquitos appear unsuccessful at making their bites less itchy. If a mosquito did evolve to not provoke (as much of) an immune response and therefore less itching, redness and swelling, you probably wouldn’t notice they’d bitten you. People often perceive that some are prone to getting bitten, others aren’t. It may be that some of this is that some people don’t have as serious an immune response to mosquito bites, so they think they get bitten less often.

I’m sure there are several PhDs worth of research questions to investigate here—I’m a biomedical engineer with a good basic understanding of the immune system, but I don’t study mosquitos.

Because they have no reproductive advantage to being less itchy. You can kill them while they’re feeding, which is why they put lots of evolutionary effort into not being noticed. (They have an anesthetic in their saliva so you are unlikely to notice the bite.) By the time you develop the itchy bump, they’ve flown away and you can’t kill them.

There’s still

somepressure, though. If the bites were permanently not itchy, then I may have not noticed that the mosquitos were in my room in the first place, and consequently would less likely pursue them directly. I guess that’s just not enough.There’s also positive selection for itchiness. Mosquito spit contains dozens of carefully evolved proteins. We don’t know what they all are, but some of them are anticoagulants and anesthetics. Presumably they wouldn’t be there if they didn’t have a purpose. And your body, when it detects these foreign proteins, mounts a protective reaction, causing redness, swelling, and itching. IIRC, that reaction does a good job of killing any viruses that came in with the mosquito saliva. We’ve evolved to have that reaction. The itchiness is probably good for killing any bloodsuckers that don’t flee quickly. It certainly works against ticks.

Evolution is not our friend. It doesn’t give us what we want, just what we need.

I believe mosquitos do inject something to suppress your reaction to them, which is why you don’t notice bug bites until long after the bug is gone. There’s no reproductive advantage to the mosquito to extending that indefinitely.

Oh wow, that would make a ton of sense. Thanks Elizabeth!

I had something like locality in mind when writing this shortform, the context being: [I’m in my room → I notice itch → I realize there’s a mosquito somewhere in my room → I deliberately pursue and kill the mosquito that I wouldn’t have known existed without the itch]

But, again, this probably wouldn’t amount to much selection pressure, partially due to the fact that the vast majority of mosquito population exists in places where such locality doesn’t hold i.e. in an open environment.

In NZ we have biting bugs called sandflies which don’t do this—you can often tell the moment they get you.

The reason you find them itchy is because humans are selected to find them itchy most likely?

But the evolutionary timescale at which mosquitos can adapt to avoid detection must be faster than that of humans adapting to find mosquitos itchy! Or so I thought—my current boring guess is that (1) mechanisms for the human body to detect foreign particles are fairly “broad”, (2) the required adaptation from the mosquitos to evade them are not-way-too-simple, and (3) we just haven’t put enough selection pressure to make such change happen.

Yeah that would be my thinking as well.

Just noticing that the negation of a statement exists is enough to make meaningful updates.e.g. I used to (implicitly) think “Chatbot Romance is weird” without having evaluated anything in-depth about the subject (and consequently didn’t have any strong opinions about it)—probably as a result of some underlying cached belief.

But after seeing this post,

just reading the titlewas enough to make me go (1) “Oh! I just realized it isperfectly possibleto argue in favor of Chatbot Romance … my belief on this subject must be a cached belief!” (2) hence is probably by-default biased towards something like the consensus opinion, and (3) so I should update away from my current direction, even without reading the post.People mean different things when they say “values” (object vs meta values)I noticed that people often mean different things when they say “values,” and they end up talking past each other (or convergence only happens after a long discussion). One of the difference is in whether they contain meta-level values.

Some people refer to the “object-level” preferences that we hold.

Often people bring up the “beauty” of the human mind’s capacity for its values to change, evolve, adopt, and grow—changing mind as it learns more about the world, being open to persuasion via rational argumentation, changing moral theories, etc.

Some people include the meta-values (that are defined on top of other values, and the evolution of such values).

e.g., My “values” include my meta-values, like wanting to be persuaded by good arguments, wanting to change my moral theories when I get to know better, even “not wanting my values to be fixed”

example of this view: carado’s post on you want what you want, and one of Vanessa Cosoy’s shortform/comment (can’t remember the link)

## Unidimensional Continuity of Preference ≈ Assumption of “Resources”?

tl;dr, the unidimensional continuity of preference assumption in themoney pumping argumentused to justify the VNM axioms correspond to the assumption that there exists some unidimensional “resource” that the agent cares about, and this language is provided by the notion of “souring / sweetening” a lottery.Various coherence theorems—or more specifically, various money pumping arguments generally have the following form:

… where “resources” (the usual example is money) are something that, apparently, these theorems assume exist. They

do, but this fact is often stated in a very implicit way. Let me explain.In the process of justifying the VNM axioms using money pumping arguments, one of the three main mathematical primitives are: (1) lotteries (probability distribution over outcomes), (2) preference relation (general binary relation), and

(3) a notion of Souring/Sweetening of a lottery.Let me explain what (3) means.Souring of A is denoted A−, and a sweetening of A is denoted A+.

A− is to be interpreted as “basically identical with A but strictly inferior in a single dimension that the agent cares about.” Based on this interpretation, we assume A>A−. Sweetening is the opposite, defined in the obvious way.

Formally, souring could be thought of as introducing a new preference relation A>uniB, which is to be interpreted as “lottery B is basically identical to lottery A, but strictly inferior in a single dimension that the agent cares about”.

On the syntactic level, such B is denoted as A−.

On the semantic level, based on the above interpretation, >uni is related to > via the following: A>uniB⟹A>B

This is where the language to talk about resources come from. “Something you can independently vary alongside a lottery A such that more of it makes you prefer that option compared to A alone” sounds like what we’d intuitively call a resource

^{[1]}.Now that we have the language, notice that so far we haven’t assumed sourings or sweetenings exist. The following assumption does it:

Which gives a more operational characterization of souring as something that lets us interpolate between the preference margins of two lotteries—intuitively satisfied by e.g., money due to its infinite divisibility.

So the above assumption is where the assumption of

resourcescome into play. I’m not aware of any money pump arguments for this assumption, or more generally, for the existence of a “resource.” Plausibly instrumental convergence.I don’t actually think this + the assumption below fully capture what we intuitively mean by “resources”, enough to justify this terminology. I stuck with “resources” anyways because others around here used that term to (I think?) refer to what I’m describing here.

Thinking about for some time my feeling has been that resources are about fungibility implicitly embedded in a context of trade, multiple agents (very broadly construed. E.g. an agent in time can be thought of as multiple agents cooperating intertemporally perhaps).

A resource over time has the property that I can spend it now or I can spend it later. Glibly, one could say the operational meaning of the resource arises from the intertemporal bargaining of the agent.

Perhaps it’s useful to distinguish several levels of resources and resource-like quantities.

Discrete vs continuous, tradeable / meaningful to different agents, ?? Fungibility, ?? Temporal and spatial locatedness, ?? Additivity?, submodularity ?

Addendum: another thing to consider is that the input of the vNM theorem is in some sense more complicated than the output. The output is just a utility function u: X → R, while your input is a preference order on the very infinite set of lotteries (= probability distributions ) L(X).

Thinking operationally about a preference ordering on a space of distribution is a little wacky. It means you are willing to trade off uncertain options against one another. For this to be a meaningful choice would seem to necessitate some sort of (probabilistic) world model.

Damn, why did Pearl recommend readers (in the preface of his causality book) to read all the chapters other than chapter 2 (and the last review chapter)? Chapter 2 is literally the coolest part—inferring causal structure from purely observational data! Almost skipped that chapter because of it …

it’s true it’s cool, but I suspect he’s been a bit disheartened by how complicated it’s been to get this to work in real-world settings.

in the book of why, he basically now says it’s impossible to learn causality from data, which is a bit of a confusing message if you come from his previous books.

but now with language models, I think his hopes are up again, since models can basically piggy-back on causal relationships inferred by humans

You should also check out Timeless Causality, if you haven’t done so already.

Bayes Net inference algorithms maintain its efficiency by using dynamic programming over multiple layers.

Level 0: Naive Marginalization

No dynamic programming whatsoever. Just multiply all the conditional probability distribution (CPD) tables, and sum over the variables of non-interest.

Level 1: Variable Elimination

Cache the repeated computations within a query.

For example, given a chain-structured Bayes Net A⟶B⟶C⟶D, instead of doing P(D)=∑A∑B∑CP(A,B,C,D), we can do P(D)=∑CP(D|C)∑BP(C|B)∑AP(A)P(B|A). Check my post for more.

Level 2: Clique-tree based algorithms — e.g., Sum-product (SP) / Belief-update (BU) calibration algorithms

Cache the repeated computations across queries.

Suppose you have a fixed Bayes Net, and you want to compute the marginalization not only P(D), but also P(A). Clearly running two instances of Variable Elimination as above is going to contain some overlapping computation.

Clique-tree is a data structure where, given the initial factors (in this case the CPD tables), you “calibrate” a tree whose nodes correspond to a subset of the variables. Cost can be amortized by running many queries over the same Bayes Net.

Calibration can be done by just two passes across the tree, after which you have the joint marginals for all the nodes of the clique tree.

Incorporating evidence is equally simple. Just zero-out the entries of variables that you are conditioning on for some node, then “propagate” that information downwards via a single pass across the tree.

Level 3: Specialized query-set answering algorithms over a calibrated clique tree.

Cache the repeated computations across a certain query-class

e.g., computing P(X,Y) for every pair of variables can be done by using yet another layer of dynamic programming by maintaining a table of P(Ci,Cj) for each pair of clique-tree nodes ordered according to their distance in-between.

Man, deviation arguments are so cool:

what are macrostates? Variables which are required to make your thermodynamics theory work! If they don’t, add more macrostates!

nonequilibrium? Define it as systems that don’t admit a thermodynamic description!

inductive biases? Define it as the amount of correction needed for a system to obey Bayesian updating, i.e. correction terms in the exponent of the Gibbs measure!

coarse graining? Define the coarse-grained variables to keep the dynamics as close as possible to that of the micro-dynamics!

or in a similar spirit—does your biological system deviate from expected utility theory? Well, there’s discovery (and money) to be made!

It’s easy to get confused and think the circularity is a problem (“how can you define thermodynamics in terms of equilibriums, when equilibriums are defined using thermodynamics?”), but it’s all about carving nature at the right joints—and a sign that you made the right carving is that the amount of corrections needed to be applied aren’t too numerous, and they all seem “natural” (and of course, all of this while letting you make nontrivial predictions. that’s what matters at the end of the day).

Then, it’s often the case that those corrections also turn out to be meaningful and natural quantities of interest.

One of the rare insightful lessons from high school: Don’t set your AC to the minimum temperature even if it’s really hot, just set it to where you want it to be.

It’s not like the air released gets colder with lower target temperature, because most ACs (according to my teacher, I haven’t checked lol) are just a simple control system that turns itself on/off around the target temperature, meaning the time it takes to reach a certain temperature X is independent of the target temperature (as long it’s lower than X)

… which is

embarrassinglyobvious in hindsight.Well is he is right about some ACs being simple on/off units.

But there also exists units than can change cycle speed, its basically the same thing except the motor driving the compression cycle can vary in speed.

In case you where wondering, they are called inverters. And when buying new today, you really should get an inverter (efficiency).

(Note: This was a post, but in retrospect was probably better to be posted as a shortform)

(Epistemic Status: 20-minute worth of thinking, haven’t done anybuilder/breakeron this yet although I plan to, and would welcome any attempts in the comment)Have an algorithmic task whose input/output pair could (in reasonable algorithmic complexity) be generated using highly specific combination of modular components (e.g., basic arithmetic, combination of random NN module outputs, etc).

Train a small transformer (or anything, really) on the input/output pairs.

Take a large transformer that takes the activation/weights, and outputs a computational graph.

Train that large transformer over the small transformer, across a diverse set of such algorithmic tasks (probably automatically generated) with varying complexity. Now you have a general tool that takes in a set of high-dimensional matrices and backs-out a simple computational graph, great! Let’s call it

Inspector.Apply the Inspector in real models and see if it recovers anything we might expect (like induction heads).

To go a step further, apply the Inspector to itself.

Maybe we might back-out a human implementable general solution for mechanistic interpretability!(Or, at least let us build a better intuition towards the solution.)(This probably won’t work, or at least isn’t as simple as described above. Again, welcome anybuilder/breakerattempts!)Quick thoughts on my plans:

I want to focus on having a better mechanistic picture of agent value formation & distinguishing between hypotheses (e.g., shard theory, Thane Ruthenis’s value-compilation hypothesis, etc) and forming my own.

I think I have a specific but very high uncertainty baseline model of what-to-expect from agent value-formation using greedy search optimization. It’s probably time to allocate more resources on reducing that uncertainty by touching reality i.e. running experiments.

(and also think about related theoretical arguments like Selection Theorem)

So I’ll probably allocate my research time:

Studying math (more linear algebra / dynamical systems / causal inference / statistical mechanics)

Sketching a better picture of agent development, assigning confidence, proposing high-bit experiments (that might have the side-effect of distinguishing between different conflicting pictures), formalization, etc.

and read relevant literature (eg ones on theoretic DL and inductive biases)

Upskilling mechanistic interpretability to actually start running quick experiments

Unguided research brainstorming (e.g., going through various alignment exercises, having a writeup of random related ideas, etc)

Possibly participate in programs like MATS? Probably the biggest benefit to me would be (1) commitment mechanism / additional motivation and (2) high-value conversations with other researchers.

Dunno, sounds pretty reasonable!

Useful perspective when thinking of mechanistic pictures of agent/value development is to take the “perspective” of different optimizers, consider their relative “power,” and how they interact with each other.

E.g., early on SGD is the dominant optimizer, which has the property of (having direct access to feedback from U / greedy). Later on early proto-GPS (general-purpose search) forms, which is less greedy, but still can largely be swayed by SGD (such as having its problem-specification-input tweaked, having the overall GPS-implementation modified, etc). Much later, GPS becomes the dominant optimizing force “at

run-time” which shortens the relevant time-scale and we can ignore the SGD’s effect. This effect becomes much more pronounced after reflectivity + gradient hacking when the GPS’s optimization target becomes fixed.(very much inspired by reading Thane Ruthenis’svalue formationpost)This is a

very useful approximationat the late-stage when the GPS self-modifies the agent in pursuit of its objective! Rather than having to meticulously think about local SGD gradient incentives and such, since GPS is non-greedy, we can directly model it as doing what’s obviously rational from a birds-eye-perspective.(kinda similar to e.g., separation of timescale when analyzing dynamical systems)

It seems like retrieval-based transformers like RETRO is “obviously” the way to go—(1) there’s just no need to store all the factual information as fixed weights, (2) and it uses much less parameter/memory. Maybe mechanistic interpretability should start paying more attention to these type of architectures, especially since they’re probably going to be a more relevant form of architecture.

They might also be easier to interpret thanks to specialization!

I’ve noticed during my alignment study that just the sheer amount of relevant posts out there is giving me a pretty bad habit of (1) passively engaging with the material and (2) not doing much independent thinking. Just keeping up to date & distilling the stuff in my todo read list takes up most of my time.

I guess the reason I do it is because (at least for me) it takes a ton of mental effort to switch modes between “passive consumption” and “active thinking”:

I noticed then when self-studying math; like, my subjective experience is that I enjoy

both“passively listening lectures+taking notes” and “solving practice problems,” the problem is that it takes a ton of mental energy to switch between the two equilibriums.(This is actually still a problem—too much wide & passive consumption rather than actively practicing them and solving problems.)

Also relevant is wanting to just progress/upskill as fast and wide of a subject as I can, sacrificing mastery for diversity. This probably makes sense to some degree (especially in the sense that having more frames is good), but I think I’m taking this wayyyy too far.

My r for opening new links far exceeds 1. This definitely helped me when I was trying to get a rapid overview of the entire field, but now it’s just a bad adaptation + akrasia.

Okay, then, don’t do that!Some directions to move towards:Independent brainstorming/investigation sessions to form concrete inside views

like the advice from the field guide post, exercises from the MATS model, et

Commitment mechanisms, like making regular posts or shortforms (eg)

There are lots of posts but the actual content is very thing. I would say there is plausibly more content in your real analysis book than there is in the entire alignment field.

Is there a case for AIgain-of-functionresearch?(Epistemic Status: I don’t endorse this yet, just thinking aloud. Please let me know if you want to act/research based on this idea)

It seems like it should be possible to materialize certain forms of AI alignment failure modes with today’s deep learning algorithms, if we directly optimize for their discovery. For example, training a Gradient Hacker Enzyme.

A possible benefit of this would be that it gives us bits of evidence wrt how such hypothesized risks would actually manifest in real training environments. While the similarities would be limited because the training setups would be optimizing for their discovery, it should at least serve as a good lower bound for the scenarios in which these risks could manifest.

Perhaps having a concrete bound for when dangerous capabilities appear (eg a X parameter model trained in Y modality has Z chance of forming a gradient hacker) would make it easier for policy folks to push for regulations.

Is AI gain-of-function equally dangerous as biotech gain-of-function? Some arguments in favor (of the former being dangerous):

The malicious actor argument is probably stronger for AI gain-of-function.

if someone publicly releases a Gradient Hacker Enzyme, this lowers the resource that would be needed for a malicious actor to develop a misaligned AI (eg plug in the misaligned Enzyme at an otherwise benign low-capability training run).

Risky researcher incentive is equally strong.

e.g., a research lab carelessly pursuing gain-of-function research, deliberately starting risky training runs for financial/academic incentives, etc.

Some arguments against:

Accident risks from financial incentives are probably weaker for AI gain-of-function.

The standard gain-of-function risk scenario is: research lab engineers a dangerous pathogen, it accidentally leaks, and a pandemic happens.

I don’t see how these events would happen “accidentally” when dealing with AI programs; e.g., the researcher would have to deliberately cut parts of the network weights and replace it with the enzyme, which is certainly intentional.

Random alignment-related idea: train and investigate a “Gradient Hacker Enzyme”TL;DR, Use meta-learning methods like MAML to train a network submodule i.e. circuit that would resist gradient updates in a wide variety of contexts (various architectures, hyperparameters, modality, etc), and use mechanistic interpretability to see how it works.

It should be possible to have a training setup for goals other than “resist gradient updates,” such as restricting the meta-objective to a specific sub-sub-circuit. In that case, the outer circuit might (1) instrumentally resist updates, or (2) somehow get modified while keeping its original behavioral objective intact.

This setup doesn’t have to be restricted to circuits of course; there was a previous work which did this on the level of activations, although iiuc the model found a trivial solution by exploiting relu—it would be interesting to extend this to more diverse setup.

Anyways, varying this “sub-sub-circuit/activation-to-be-preserved” over different meta-learning episodes would incentivize the training process to find “general” Gradient Hacker designs that aren’t specific to a particular circuit/activation—a potential precursor for various forms of advanced Gradient Hackers (and some loose analogies to how enzymes accelerate reactions).

What is the Theory of Impact for training a “Gradient Hacker Enzyme”?(note: while I think these are valid, they’re generated post-hoc and don’t reflect the actual process for me coming up with this idea)

Estimating the lower-bound for the emergence Gradient Hackers.

By varying the meta-learning setups we can get an empirical estimate for the conditions in which Gradient Hackers are possible.

Perhaps gradient hackers are actually trivial to construct using tricks we haven’t thought of before (like the relu example before). Maybe not! Perhaps they require [high-model-complexity/certain-modality/reflective-agent/etc].

Why lower-bound?In a real training environment, gradient hackers appear because of (presumably) convergent training incentives. Instead in the meta-learning setup, we’redirectly optimizingfor gradient hackers.Mechanistically understanding how Gradient Hackers work.

Applying mechanistic interpretability here might not be too difficult, because the circuit is cleanly separated from the rest of the model.

There has been several speculations on how such circuits might emerge. Testing them empirically sounds like a good idea!

This is just a random idea and I’m probably not going to work on it; but if you’re interested, let me know. While I don’t think this is capabilities-relevant, this probably falls under AI gain-of-function research and should be done with caution.

Update: I’m trying to upskill mechanistic interpretability, and training a Gradient Hacker Enzyme seems like a fairly good project just to get myself started.

I don’t think this project would be highly valuable in and of itself (although I would definitely learn a lot!), so one failure mode I need to avoid is ending up investing too much of my time in this idea. I’ll probably spend a total of ~1 week working on it.