# Aaro Salosensaari

Karma: 133

aqsalose.fi

• >It turns out that using Transformers in the autoregressive mode (with output tokens being added back to the input by concatenating the previous input and the new output token, and sending the new versions of the input through the model again and again) results in them emulating dynamics of recurrent neural networks, and that clarifies things a lot...

I’ll bite: Could you dumb down the implications of the paper a little bit, what is the difference between a Transformer emulating a RNN and some pre-Transformer RNNs and/​or not-RNN?

My much more novice-level answer to Hofstadter’s intuition would have been: it’s not the feedforward firing, but it is the gradient descent training of the model on massive scale (both in data and in computation). But apparently you think that something RNN-like about the model structure itself is important?

• Epistemic status: I am probably misunderstanding some critical parts of the theory, and I am quite ignorant on technical implementation of prediction markets. But posting this could be useful for my and others’ learning.

First question. Am I understanding correctly how the market would function. Taking your IRT probit market example, here is what I gather:

(1) I want to make a bet on the conditional market P(X_i | Y). I have a visual UI where I slide bars to make a bet on parameters a and b; (dropping subscript i) however, internally this is represented by a bet on a’ = a sigma_y and P(X) = Phi(b’/​sqrt(a’+1)), b’ = b + a mu_y. So far, so clear.

(2) I want to bet on the latent market P(Y). I make bet on mu_Y and sigma_y, which is internally represented by a bet on a’ and P(X). In your demo, this is explicit: P(Y) actually remains a standard normal distribution, it is the conditional distributions that shift.

(3) I want to bet on the unconditional market P(X_i), which happens to be an indicator variable for some latent variable market I don’t really care about. I want to make a bet on P(X_i) only. What does exactly happen? If the internal technical representation is a database row [p_i, a’] changing to [p_new, a’], then I must be implicitly making a bet on a’ and b’ too, as b’ derived from a and P(X) and P(X) changed. In other words, it appears I am also making a bet on the conditional market?

Maybe this is OK if I simply ignore the latent variable market, but I could also disagree with it: I think P(X) is not correlated with the other indicators, or the correlation structure implied by linear probit IRT model is inadequate. Can I make a bet of form [p_i, NA] and have it not to participate in the latent variable market structure? (Ie when the indicator market resolves, p_i is taken into account for scoring, NA is ignored while scoring Y).

Thus, I can launch a competing latent variable market for the same indicators? How can I make a bet against a particular latent structure in favor of another structure? A model selection bet, if you will.

Context for the question: In my limited experience of latent variable models for statistical inference tasks, the most difficult and often contested question is whether the latent structure you specified is good one. The choice of model and interpretations of it are very likely to get criticized and debated in published work.

Another question for discussion.

It seems theoretically possible to obtain individual predictors’ bets /​ predictions on X_1 … X_k, without presence of any latent variable market, impute for the missing values if some predictors have not predicted on all k indicators, and then estimate a latent variable model on this data. What is the exact benefit of having the latent model available for betting? If fit a Bayesian model (non-market LVM with indicator datapoints) and it converges, from this non-market model, I would obtain posterior distribution for parameters and compute many of the statistics of interest.

Presumably, if there is a latent “ground truth” relationship between all indicator questions and the market participants are considering them, the relationship would be recovered by such analysis. If the model is non-identifiable or misspecified (= such ground truth does not exist /​ is not recoverable), I will have problems fitting it and with the interpretation of fit. (And if I were to run a prediction market for a badly specified LVM, the market would exhibit problematic behavior.)

By not making a prediction market for the latent model available, I won’t have predictions on the latent distribution directly, but on the other hand, I could try to estimate several different without imposing any of their assumptions about presumed joint structure on the market. I could see this beneficial, depending on how the predictions on Y “propagate” to indicator markets or the other way around (case (3) above).

Final question for discussion: Suppose I run a prediction market using the current market implementations (Metaculus or Manifold or something else, without any LVM support), where I promise to fit a particular IRT probit model at time point t in future and ask for predictions on the distribution of posterior model parameters estimates when I fit the model. Wouldn’t I obtain most of the benefits of “native technical implementation” of a LVM market … except that updating away any inconsistencies between the LVM prediction market estimate and individual indicator market will be up to to individual market participants? The bets on this estimated LVM market at least should behave as any other bets, right? (Related: which should be the time point t when I promise to fit the model and resolve the market? t = when the first indicator market resolves? when the last indicator market resolves? arbitrary fixed time point T?)

• >Glancing back and forth, I keep changing my mind about whether or not I think the messy empirical data is close enough to the prediction from the normal distribution to accept your conclusion, or whether that elbow feature around 1976-80 seems compelling.

I realize you two had a long discussion about this, but my few cents: This kind of situation (eyeballing is not enough to resolve which of two models fit the data better) is exactly the kind of situation for which a concept of statistical inference is very useful.

I’m a bit too busy right now to present a computation, but my first idea would be to gather the data and run a simple “bootstrappy” simulation: 1) Get the original data set. 2) Generate k = 1 … N simulated samples x^k = [x^k_1, … x^k_t] form a normal distribution with linearly increasing mean mu(t) = mu + c * t at time points t= 1960 … 2018, where c and variance are as in “linear increase hypothesis”. 3) How many of simulated replicate time series have an elbow at 1980 that is equally or more extreme than observed in the data? (One could do this not too informal way by fitting a piece-wise regression model with break at t = 2018 to reach replicate time series, and computing if the two slope estimates differ by a predetermined threshold, such as the estimates recovered by fitting the same piece-wise model in the real data).

This is slightly ad hoc, and there are probably fancier statistical methods for this kind of test, or you could fits some kind of Bayesian model, but I’d think such computational exercise would be illustrative.

• Hyperbole aside, how many of those experts linked (and/​or contributing to the 10% /​ 2% estimate) have arrived to their conclusion with a thought process that is “downstream” from the thoughtspace the parent commenter thinks suspect? Then it would not qualify as independent evidence or rebuttal, as it is included as the target of criticism.

• (I initially thought it would be better not to mention the context of the question as it might bias the responses. OTOH the context could make the marginal LW poster more interested in providing answers, so I here it is:)

It came up in an argument that the difficulty of economic calculation problem could be a difficult to a hypothetical singleton, insomuch a singleton agent needs certain amount of compute relative to the economy in question. My intuition consists two related hypotheses: First, during any transition period where any agent participates in global economy where most other participants are humans (“economy” could be interpreted widely to include many human transactions), can the problem of economic calculation provide some limits how much calculation would be needed for an agent to become able to manipulate /​ dominate the economy? (Is it enough for an agent to be marginally more capable than any other participant, or does it get swamped by the sheer size of the economy is large enough?)

Secondly, if an Mises/​Hayek answer is correct and the economic calculation problem is solved most efficiently by a distributed calculation, it could imply that a single agent in a charge of a number of processes on “global economy” scale could be out-competed by a community of coordinating agents. [1]

However, I would like to read more to judge if my intuitions are correct. Maybe all of this is already rendered moot by results I simply do not how to find.

([1] Related but tangential: Can one provide a definition when distributed computation is no longer a singleton but more-or-less aligned community of individual agents? My hunch is, there could be a characterizations related to speed of communication between agents /​ processes in a singleton. Ultimately speed of light is prone to mandate some limitations.)

• Can anyone recommend good reading material on economic calculation problem?

• I found this interesting. Finnish is also language of about 5 million speakers, but we have a commonly used natural translation of “economies of scale” (mittakaavaetu, “benefit of scale”). Any commonplace obvious translation for “Single point of failure” didn’t strike my mind, so I googled, and found engineering MSc thesis works and similar documents: the words they choose to use included yksittäinen kriittinen prosessi (“single critical process”, most natural one IMO), yksittäinen vikaantumispiste (“single point of failure”, literal translation and a bit clumsy one), yksittäinen riskikohde (“single object of risk”, makes sense but only in the context), and several phrases that chose to explain the concept.

Small languages need active caretaking and cultivation so that translations of novel concepts are introduced, and then there is active intellectual life where they are used. In Finnish, this work has been done for “economies of scale”, but less efficiently for “single point of failure”. But I believe one could use any of the translations I found or invent ones own, maybe add the English term in parenthesis, and not look like a crackpot. (Because I would expect quite a few people are familiar with the concept with its English name. In a verbal argument I would expect a politician just say it in English if they didn’t know an established equivalent in Finnish. Using English is fancy and high-status in Finland in the way French was fancy and high-status in the 19th century.)

In another comment you make a comparison to LW concepts like Moloch. [1] I think the idea of “cultivation” is also applicable to LW shibboleths too, especially in the context of the old tagline “rising the sanity waterline”. It is useful to have a good language—or very least, good words for important concepts. (And also maybe avoid promoting words for concepts that are not well thought-out.) Making such words common requires active work, which includes care in choice of words that are descriptive/​good-sounding and good judgement to choose words and use-patterns that they can become popular and make it to common use without sounding crackpot-ish. (In-group shibboleths can become a failure mode.)

Lack of such work is obvious sooner in small languages than in larger ones, but even large language with many speakers, like English, miss words for every concept they have not yet adopted a word for.

In Iceland, much smaller language, I have heard they translate and localize everything.

• “if I were an AGI, then I’d be able to solve this problem” “I can easily imagine”

Doesn’t this way of analysis come with a ton of other assumptions left unstated?

Suppose “I” am an AGI running on a data center and I can modeled as an agent with some objective function that manifest as desires and I know my instantiation needs electricity and GPUs to continue running. Creating another copy of “I” running in the same data center will use the same resources. Creating another copy in some other data center requires some other data center.

Depending on the objective function and algorithm and hardware architecture bunch of other things, creating copies may result some benefits from distributed computation (actually it is quite unclear to me if “I” happen already to be a distributed computation running on thousands of GPUs—do “I” maintain even a sense of self—but let’s no go into that).

The key here is the word may. Not obviously it necessarily follows that

For example: Is the objective function specified so that the agent will find creating a copy of itself beneficial for fulfilling the objective function (informally, it has internal experience of desiring to create copies)? As the OP points out, there might be a disagreement: for the distributed copies to be any useful, they will have different inputs and thus they will end in different, unanticipated states. What “I” am to do when “I” disagree another “I”? What if some other “I” changes, modifies its objective function into something unrecognizable to “me”, and when “we” meet, it gives false pretenses of cooperating but in reality only wants hijack “my” resources? Is the “trust” even the correct word here, when “I” could verify instead: maybe “I” prefer to create and run a subroutine of limited capability (not a full copy) that can prove its objective function has remained compatible with “my” objective function and will terminate willingly after it’s done with its task (killswitch OP mentions) ? But doesn’t this sound quite like our (not “our” but us humans) alignment problem? Would you say “I can easily imagine if I were an AGI, I’d be easily able to solve it” to that? Huh? Reading LW I have come to think the problem is difficult to the human-general intelligence.

Secondly: If “I” don’t have any model of data centers existing in the real world, only the experience of uploading myself to other data centers (assuming for the sake of argument all the practical details of that can be handwaved off), i.e. it has a bad model of the self-other boundary described in OPs essay, it could easily end up copying itself to all available data centers and then becoming stuck without any free compute left to “eat” and adversely affecting human ability to produce more. Compatible with model and its results in the original paper (take the non-null actions to consume resource because U doesn’t view the region as otherwise valuable). It is some other assumptions (not the theory) that posit an real-world affecting AGI would have U that doesn’t consider the economy of producing the resources it needs.

So if “I” were to successful in running myself with only “I” and my subroutines, “I” should have a way to affecting the real world and producing computronium for my continued existence. Quite a task to handwaved away as trivial! How much compute an agent running in one data center (/​unit of computronium) needs to successfully model all the economic constraints that go into the maintenance of one data center? Then add all the robotics to do anything. If “I” have a model of running everything a chip fab requires more efficiently than the current economy, and act on it, but the model was imperfect and the attempt is unsuccessful but destructive to economy, well, that could be [bs]ad and definitely a problem. But it is a real constraint to the kind of simplifying assumptions the OP critiques (disembodied deployer of resources with total knowledge).

All of this—how would “I” solve a problem and what problems “I” am aware of—is contingent on, I would call them, the implementation details. And I think author is right to point them out. Maybe it does necessary follows, but it needs to be argued so.

• Why wonder when you can think: What is the substantial difference in MuZero (as described in [1]) that makes the algorithm to consider interruptions?

Maybe I show some great ignorance of MDPs, but naively I don’t see how an interrupted game could come into play as a signal in the specified implementations of MuZero:

Explicit signals I can’t see, because the explicitly specified reward u seems contingent ultimately only on the game state /​ win condition.

One can hypothesize an implicit signal could be introduced if algorithm learns to “avoid game states that result in game being terminated for out-of-game reason /​ game not played until the end condition”, but how such learning would happen? Can MuZero interrupt the game during training? Sounds unlikely such move would be implemented in Go or Shogi environment. Are there any combination of moves in Atari game that could cause it?

• a backdrop of decades of mistreatment of the Japanese by Western countries.

I find this a bit difficult to take seriously. The WW2 in the Pacific didn’t start with well-treatment of China and other countries by Japan, either. Naturally Japanese didn’t care about that part of the story, but hey had plenty of other options how they could have responded their the UK or the US trade policy instead of invading Manchuria.

making Ukraine a country with a similar international status to Austria or Finland during the Cold War would be one immediate solution.

This is not a simple task, but rather a tall order. Austria was “made neutral” after it was occupied. Finland signed a peace treaty that put it into effectively similar position. Why would any country submit to such a deal voluntarily? The answer is, they often don’t. Finland didn’t receive significant assistance from the Allies in 1939, yet they decided to defend themselves against the USSR anyway when Stalin attacked.

However, if one side in these disputes had refused to play the game of ratcheting up tensions, the eventual wars would simply not have happened. In this context it takes two to dance.

Sure, but the game theoretic implication is that this kind of strategy favors the first party to take the first step and say “I have an army and a map where this neighboring country belongs to us”.

NATO would have refrained from sending lethal arms to Ukraine and stationing thousands of foreign military advisors in Ukrainian territory after Maidan.

What a weird way to present the causality of events. I am quite confident NATO didn’t have time to send any weapons and certainly not thousands of advisors between Maidan and the war starting. Yanukovich fled 22 February. Antimaidan protests started in Donetsk 1 March and shooting war started in April.

• First, avoiding arguments from the “other side” on the basis that they might convince you of false things assumes that the other side’s belief are in fact false.

I believe it is less about true/​false, but whether you believe the “other side” is making a well-intentioned effort at obtaining and sharing accurate maps of reality. On practical level, I think it is unlikely studying Russian media in detail is useful and cost-effective for a modal LWer.

Propaganda during wartime, especially during total war, is a prima facia example of situation where every player of note is doing their best to convince you of something in order to produce certain effects. [2] To continue with the map metaphor, they want to you to have a certain kind of map that will guide you to certain location. All parties wish to do this to some extent, and because it is a situation with the highest stakes of all, they are putting in their best effort.

Suppose you read lots of Western media sources and then a lot of Russian media sources. All sides in the conflict do their best to fill the air with favorable propaganda. You will find yourself doing a lot of reading, and I don’t know if there is any guarantee you can achieve any good results by interpolating between two propaganda-infused maps [1], instead of say, reading much less of both Western media and Russian media and trying to find good close-to-ground signals, or outsourcing the time-consuming analysis part to people /​ sources who you have a good reason to trust to do a good analysis (preferably you have vetted them before the conflict, and you can trust the reason still applies).

So the good reason to read Russian media to analyze it, is if you have a good reason to believe you would be good analyst of Russian media sphere. But if you were, would you find yourself reading a Russian newspaper you had not heard about two weeks ago with Google translate?

[1] I don’t have references at hand to give a good summary, but imagine you are your great*-grandparent and reading newspapers during WW2. At great expense you manage to get newspapers from London, New York, Berlin, Tokyo, and Moscow. Are you going to get good picture of “what happens” by reading them all? I think you would get some idea of how situation develops by reading accounts of battles and cross-referencing a map, but I don’t know it would be worth the expense. One thing I know, none of them is reporting much at all about the thing you most likely consider most salient about WW2, namely, the holocaust and the atomic bomb until after the fact.

[2] edit. addendum. Zvi used the word “hostile” and I want to stress its importance. During peacetime and in internal politics it is often a mistake to assume hostile influences (ie. conflict on conflict/​mistake theory spectrum), because then you are engaging in a conflict all the time and likely to escalate it more and more. But now that we have a major European war, I think that is a good situation to assume that the players in the field are actually “hostile” because there is a shooting war conflict to begin with.

• Open thread is presumably the best place for a low-effort questions, so here goes:

I came across this post from 2012: Thoughts on the Singularity Institute (SI) by Holden Karnofsky (then-Co-Executive Director of GiveWell). Interestingly enough, some of the object-level objections (under subtitle “objections”) Karnofsky raises[1] are similar to some points that were came up in the Yudkowsky/​chathamroom.com discussion and Ngo/​​Yudkowsky dialogue I read the other day (or rather, read parts of, because they were quite long).

What are people’s thought about that post and objections raised today? What the 10 year (-ish, 9.5 year) retrospective looks like?

Some specific questions.

Firstly, how his arguments would be responded today? Any substantial novel contra-objections? (I ask because its more fun to ask than start reading through Alignment forum archives.)

Secondly, predictions. When I look at the bullet points under the subtitle “Is SI the kind of organization we want to bet on?”, I think I can interpolate a prediction Karnofsky could have made: in 2012, SI [2] had not the sufficient capability nor engaged in activities likely to achieve its stated goals (“Friendliness theory” or Friendly AGI before others), as it was not worth a GiveWell funding recommendation in 2012.

A perfect counterfactual experiment this is not, but given what people on LW today know about what SI/​MIRI did achieve in the NoGiveWell!2012 timeline, was Karnofsky’s call correct, incorrect or something else? (As in, did his map of the situation in 2012 matched the reality better than some other map, or was it poor compared to other map?) What inferences could be drawn, if any?

Would be curious to hear perspectives from MIRI insiders, too (edit. but not only them). And I noticed Holden Karnofsky looks active here on LW, though I have no idea if how to ping him.

[1] Tool-AI; idea that advances in tech would bring insights into AGI safety.

[2] succeeded by MIRI I suppose

edit2. fixed ordering of endnotes.

• Yeah, random internet forum users emailing eminent mathematician en masse would be strange enough to be non-productive. I for one wasn’t thinking anyone would to, I don’t think it was what OP suggested. To anyone contemplating sending one, the task is best delegated to someone who not only can write coherent research proposals that sound relevant to the person approached, but can write the best one.

Mathematicians receive occasional crank emails about solutions to P ?= NP, so anyone doing the reaching needs to be reputable to get past their crank filters.

• A reply to comments showing skepticism about how mathematical skills of someone like Tao could be relevant:

Last time I thought I would understood anything of Tao’s blog was around ~2019. Then he was working on curious stuff, like whether he could prove there can be finite-time blow-up singularities in Navier-Stokes fluid equations (coincidentally, solving the famous Millenium prize problem showing non-smooth solution) by constructing a fluid state that both obeys Navier-Stokes and also is Turing complete and … ugh, maybe I quote the man himself:

[...] one would somehow have to make the incompressible fluid obeying the Navier–Stokes equations exhibit enough of an ability to perform computation that one could programme a self-replicating state of the fluid that behaves in a manner similar to that described above, namely a long period of near equilibrium, followed by an abrupt reorganization of the state into a rescaled version of itself. However, I do not know of any feasible way to implement (even in principle) the necessary computational building blocks, such as logic gates, in the Navier–Stokes equations.

However, it appears possible to implement such computational ability in partial differential equations other than the Navier–Stokes equations. I have shown5 that the dynamics of a particle in a potential well can exhibit the behaviour of a universal Turing machine if the potential function is chosen appropriately. Moving closer to the Navier–Stokes equations, the dynamics of the Euler equations for inviscid incompressible fluids on a Riemannian manifold have also recently been shown6,7 to exhibit some signs of universality, although so far this has not been sufficient to actually create solutions that blow up in finite time.

(Tao, Nature Review Physics 2019.)

The relation (if any, to proving stuff about computational agents alignment people are interested in) is probably spurious (I myself don’t follow either Tao’s work or alignment literature), but I am curious if he’d be interested in working on a formal system of self-replicating /​ self-improving /​ aligning computational agents, and (then) capable of finding something genuinely interesting.

minor clarifying edits.

• I have not read Irving either but he is relatively “world-famous” 1970s-1980s author. (In case it helps you to calibrate, his novel The World According To Garp is the kind of book that was published in translation in the prestigious Keltainen Kirjasto series by Finnish publisher Tammi.)

However, I would like make an opposing point about literature and fiction. I was surprised that post author mentioned a work of fiction as a positive example that demonstrates how some commonly argued option is a fabricated one. I’d think literature would at least as often (maybe more often) disseminate belief in fabricated options than correct them, as an author can easily literally fabricate (make things up, it is fiction) easily believable and memorable stories how characters choose one course of action out of many options and it works out (or not, either way, because the narrator decided so) but in reality, all options as portrayed in the story could all turn out be misrepresented, “fabricated options” in real life.

• The picture looks like evidence there is something very weird going on that is not reflected in the numbers or arguments provided. There are homeless encampments in many countries around the world, but very rarely 20 min walk from anyone’s office.

• From what I remember form my history of Finland classes, the 19th/​early 20th century state project to build a compulsory school system met some not insignificant opposition from parents. They liked having the kids working instead going to school, especially in agrarian households.

Now, I don’t want to get into debate whether schooling is useful or not (and for whom, and for what purpose, and if the usefulness has changed over time), but there is something illustrative in the opposition: children rarely are independent agents to the extent adults are. If the incentives are set in that way, the parents will prefer to make choices about their children labor that result in more resources for the household/​family unit (charitable interpretation) or for themselves (not so charitable). Number of children in the family also affects the calculus. (One kid, it makes sense to invest in their career; ten kids, and the investment was in the number.)

• Genetic algorithms are an old and classic staple of LW. [1]

Genetic algorithms (as used in optimization problems) traditionally assume “full connectivity”, that is any two candidates can mate. In other words, population network is assumed to be complete and potential mate is randomly sampled from the population.

Aymeric Vié has a paper out showing (numerical experiments) that some less dense but low average shortest path length network structures appear to result in better optimization results: https://​​doi.org/​​10.1145/​​3449726.3463134

Maybe this isn’t news for you, but it is for me! Maybe it is not news to anyone familiar with mathematical evolutionary theory?

This might be relevant for any metaphors or thought experiments where you wish to invoke GAs.

• My take is that the scientific concept of “heritability” has some problems in its construction: the exact definition (Var(genotype)/​Var(phenotype)), while useful in some regard, does not match the intuition of the word.

Maybe the quantity should be called “relative heritability”, “heritability relative to population” or “proportion of population variance explained”, like many other quantities that similarly have form A/​B where both A and B are (population) parameters or their estimates.