Reply to Eliezer on Biological Anchors

The "biological anchors" method for forecasting transformative AI is the biggest non-trust-based input into my thinking about likely timelines for transformative AI. While I'm sympathetic to parts of Eliezer Yudkowsky's recent post on it, I overall disagree with the post, and think it's easy to get a misimpression of the "biological anchors" report (which I'll abbreviate as "Bio Anchors") - and Open Philanthropy's take on it - by reading it. 

This post has three sections:

  • Most of Eliezer's critique seems directed at assumptions the report explicitly does not make about how transformative AI will be developed, and more broadly, about the connection between its (the report's) compute estimates and all-things-considered AI timelines. One way of putting this is that most of Eliezer's critique doesn't apply to the "bounding-based" interpretation of the report discussed in this post (which is my best explanation for skeptics of why I find the framework valuable; I will also give quotes below from the original report showing that its intended interpretation is along the same lines as mine).
  • Much of Eliezer's critique is some form of "Look at the reference class you're in," invoking "Platt's Law" and comparing the report to past attempts at biological anchoring. Based on my understanding of the forecasts he's comparing it to and the salient alternatives, I don't think this does much to undermine the report.
  • I also make a few minor points.

A few notes before I continue:

  • I think the comments on the post are generally excellent and interesting, and I recommend them. (I will mostly not be repeating things from the comments here.)
  • I generally view Bio Anchors as a tool for informing AI timelines rather than as a comprehensive generator of all-things-considered AI timelines, and will be discussing it as such. Bio Anchors also presents itself this way - see section Translating into views on TAI timelines.
  • Something like half of this post is blockquotes. I've often been surprised by the degree to which people (including people I respect a lot, such as Eliezer in this case) seem to mischaracterize specific pieces they critique , and I try to avoid this for myself by quoting extensively from a piece when critiquing it. (This still leaves the possibility that I'm quoting out of context; readers may want to spot-check that.)
  • This post doesn't address what some have referred to as the "meta-level core thing", though I might write some thoughts related to that in a future post.

Bounding vs. pinpointing

Here are a number of quotes from Eliezer in which I think he gives the impression that Biological Anchors assumes transformative AI will be arrived at via modern machine learning methods:

OpenPhil: Because AGI isn't like biology, and in particular, will be trained using gradient descent instead of evolutionary search, which is cheaper. We do note inside our report that this is a key assumption, and that, if it fails, the estimate might be correspondingly wrong - ...

OpenPhil: Well, search by evolutionary biology is more costly than training by gradient descent, so in hindsight, it was an overestimate. Are you claiming this was predictable in foresight instead of hindsight?

Eliezer: I'm claiming that, at the time, I snorted and tossed Somebody's figure out the window while thinking it was ridiculously huge and absurd, yes.

OpenPhil: Because you'd already foreseen in 2006 that gradient descent would be the method of choice for training future AIs, rather than genetic algorithms?

Eliezer: Ha! No. Because it was an insanely costly hypothetical approach whose main point of appeal, to the sort of person who believed in it, was that it didn't require having any idea whatsoever of what you were doing or how to design a mind.

OpenPhil: Suppose one were to reply: "Somebody" didn't know better-than-evolutionary methods for designing a mind, just as we currently don't know better methods than gradient descent for designing a mind; and hence Somebody's estimate was the best estimate at the time, just as ours is the best estimate now? ...

OpenPhil:  It seems to us that Moravec's estimate, and the guess of your nineteen-year-old past self, are both predictably vast underestimates.  Estimating the computation consumed by one brain, and calling that your AGI target date, is obviously predictably a vast underestimate because it neglects the computation required for training a brainlike system.  It may be a bit uncharitable, but we suggest that Moravec and your nineteen-year-old self may both have been motivatedly credulous, to not notice a gap so very obvious.

Eliezer:  I could imagine it seeming that way if you'd grown up never learning about any AI techniques except deep learning, which had, in your wordless mental world, always been the way things were, and would always be that way forever.

I mean, it could be that deep learning will still be the bleeding-edge method of Artificial Intelligence right up until the end of the world.  But if so, it'll be because Vinge was right and the world ended before 2030, not because the deep learning paradigm was as good as any AI paradigm can ever get.  That is simply not a kind of thing that I expect Reality to say "Gotcha" to me about, any more than I expect to be told that the human brain, whose neurons and synapses are 500,000 times further away from the thermodynamic efficiency wall than ATP synthase, is the most efficient possible consumer of computations ...

OpenPhil:  How could anybody possibly miss anything so obvious?  There's so many basic technical ideas and even philosophical ideas about how you do AI which make it supremely obvious that the best and only way to turn computation into intelligence is to have deep nets, lots of parameters, and enormous separate training phases on TPU pods ...

OpenPhil:  How quaint and archaic!  But that was 13 years ago, before time actually got started and history actually started happening in real life.  Now we've got the paradigm which will actually be used to create AGI, in all probability; so estimation methods centered on that paradigm should be valid.

However, the argument given in Bio Anchors does not hinge on an assumption that modern deep learning is what will be used, nor does it set aside the possibility of paradigm changes.

From the section What if TAI is developed through a different path?

I believe that this analysis can provide a useful median estimate even if TAI is produced through a very different path: essentially, by the time it is affordable to develop TAI through a particular highlighted route, it is plausible that somebody develops it through that route or any cheaper route. I consider the example of a distributed economic transition facilitated by a broad range of different technologies below, but the same reasoning applies to the possibility that a unified transformative program may be developed using a qualitatively different “AI paradigm” that can’t be usefully considered a descendant of modern machine learning ... 

Because this model estimates when one particular path toward transformative AI (let’s call it the “big model path”) out of many will be attainable, that means if this analysis is correct (i.e., if I am correct to assume the big model path is possible at all due to the theoretical feasibility of local search, and if we correctly estimated the probability that it would be attainable in year Y for all Y), then the probability estimates generated should be underestimates ...

However, once sources of distortion (many of which tend to push our estimates upward) are properly taken into account, I think it is fairly unclear whether these estimates should actually be considered underestimates [one such source given is similar to my comments here following "When it comes to translating my 'sense of mild surprise' into a probability] ... 

For each biological anchor hypothesis, I am acting on the assumption that there is a relatively broad space of “unknown unknown” paths to solving a transformative task within that range of technical difficulty, not just the particular concrete path I have written down for illustration in association with each hypothesis (which is often fairly conjunctive) ...

some of our technical advisors are still relatively confident these probability estimates are low-end estimates. This is partly because they would assign a higher probability to some of the low-end biological anchor hypotheses than I do, partly because they are overall more confident in the argument given above that these numbers ought to be considered underestimates ...

For now, I feel that the most reasonable way to interpret the probability estimates generated by the biological anchors framework is as a rough central estimate for when TAI will be developed rather than as particularly conservative or particularly aggressive. In making this judgment, I am admittedly mentally running together a large cloud of heterogeneous considerations which in a maximally-principled and transparent analysis should be handled separately.

That is, Ajeya (the author) sees the "median" estimate as structurally likely to be overly conservative (a soft upper bound) for reasons including those Eliezer gives, but is also adjusting in the opposite direction to account for factors including the generic burden of proof. (More discussion of "soft bounds" provided by Bio Anchors in this section and this section of the report.)

I made similar arguments in a recent piece, “Biological anchors” is about bounding, not pinpointing, AI timelines. This is my best explanation for skeptics of why I find the framework valuable.

As far as I can tell, the only part of Eliezer's piece that addresses an argument along the lines of the "soft bounding" idea is:

OpenPhil:  Doesn't our calculation at least provide a soft upper bound on how much computation is required to produce human-level intelligence?  If a calculation is able to produce an upper bound on a variable, how can it be uninformative about that variable?

Eliezer:  You assume that the architecture you're describing can, in fact, work at all to produce human intelligence.  This itself strikes me as not only tentative but probably false.  I mostly suspect that if you take the exact GPT architecture, scale it up to what you calculate as human-sized, and start training it using current gradient descent techniques... what mostly happens is that it saturates and asymptotes its loss function at not very far beyond the GPT-3 level - say, it behaves like GPT-4 would, but not much better.

This is what should have been told to Moravec:  "Sorry, even if your biology is correct, the assumption that future people can put in X amount of compute and get out Y result is not something you really know."  And that point did in fact just completely trash his ability to predict and time the future.

The same must be said to you.  Your model contains supposedly known parameters, "how much computation an AGI must eat per second, and how many parameters must be in the trainable model for that, and how many examples are needed to train those parameters".  Relative to whatever method is actually first used to produce AGI, I expect your estimates to be wildly inapplicable, as wrong as Moravec was about thinking in terms of just using one supercomputer powerful enough to be a brain.  Your parameter estimates may not be about properties that the first successful AGI design even has.  Why, what if it contains a significant component that isn't a neural network?  I realize this may be scarcely conceivable to somebody from the present generation, but the world was not always as it was now, and it will change if it does not end.

I don't literally think that the "exact GPT architecture" would work to produce transformative AI, but I think something not too far off would be a strong contender - such that having enough compute to afford this extremely brute-force method, combined with decades more time to produce new innovations and environments, does provide something of a "soft upper bound" on transformative AI timelines.

Another way of putting this is that a slightly modified version of what Eliezer calls "tentative [and] probably false]" seems to me to be "tentative and probably true." There's room for disagreement about this, but this is not where most of Eliezer's piece focused.

While I can't be confident, I also suspect that the person in the 2006 or thereabouts part of Eliezer's piece may have intended to argue for something more like a "(soft) upper bound" than a median estimate.

Finally, I want to point out this quote from Bio Anchors, which reinforces that it is intended as a tool for informing AI timelines rather than as a comprehensive generator of all-things-considered AI timelines:

This model is not directly estimating the probability of transformative AI, but rather the probability that the amount of computation that would be required to train a transformative model using contemporary ML methods would be attainable for some AI project, assuming that algorithmic progress, spending, and compute prices progress along a “business-as-usual” trajectory ...

How does the probability distribution output by this model relate to TAI timelines? In the very short-term (e.g. 2025), I’d expect this model to overestimate the probability of TAI because it feels especially likely that other elements such as datasets or robustness testing or regulatory compliance will be a bottleneck even if the raw compute is technically affordable, given that a few years is not a lot of time to build up key infrastructure. In the long-term (e.g. 2075), I’d expect it to underestimate the probability of TAI, because it feels especially likely that we would have found an entirely different path to TAI by then.

It seems that Eliezer places higher probability on an "entirely different path" sooner than Bio Anchors, but he does not seem to argue for this (and see below for why I don't think it would be a great bet). Instead, he largely argues that the possibility is ignored by Bio Anchors, which is not the case.

Platt's Law and past forecasts

Eliezer writes:

Eliezer:  So does the report by any chance say - with however many caveats and however elaborate the probabilistic methods and alternative analyses - that AGI is probably due in about 30 years from now?

OpenPhil:  Yes, in fact, our 2020 report's median estimate is 2050; though, again, with very wide credible intervals around both sides.  Is that number significant?

Eliezer:  It's a law generalized by Charles Platt, that any AI forecast will put strong AI thirty years out from when the forecast is made.  Vernor Vinge referenced it in the body of his famous 1993 NASA speech, whose abstract begins, "Within thirty years, we will have the technological means to create superhuman intelligence.  Shortly after, the human era will be ended." ... 

OpenPhil:  That part about Charles Platt's generalization is interesting, but just because we unwittingly chose literally exactly the median that Platt predicted people would always choose in consistent error, that doesn't justify dismissing our work, right? ...

Eliezer:  Oh, nice.  I was wondering what sort of tunable underdetermined parameters enabled your model to nail the psychologically overdetermined final figure of '30 years' so exactly.

I have a couple issues here.

First, I think Eliezer exaggerates the precision of Platt's Law and its match to the Bio Anchors projection:

  • Some aggregated data for assessing Platt's Law is in this comment by Matthew Barnett as well as here.
  • While Matthew says "Overall I find the law to be pretty much empirically validated, at least by the standards I'd expect from a half in jest Law of Prediction," I don't agree: I don't think an actual trendline on the chart would be particularly close to the Platt's Law line. I think it would, instead, predict that Bio Anchors should point to longer timelines than 30 years out.
  • Note that my own median projection for transformative AI is 40 years, not 30, and I know several people who have much shorter medians (15 years and under) based on their own interpretations of the analysis in the report. So I don't think it's the case that Bio Anchors "automatically" lands one on a particular view, nor that it obviously pushes against timelines as short as Eliezer's. It is a tool for informing AI timelines, and after taking it and other data points into account, Ajeya and I both are estimating longer timelines than Eliezer.

I think a softer "It's suspicious that Bio Anchors is in the same 'reasonable-sounding' general range ('a few decades') that AI forecasts have been in for a long time" comment would've been more reasonable than what Eliezer wrote, so from here I'll address that. First, I want to comment on Moravec specifically.

Eliezer characterizes Open Philanthropy as though we think that Hans Moravec's projection was foreseeably silly and overaggressive (see quote above), but now think we have the right approach. This isn't the case. 

  • On one hand, I do think that if Ajeya or I had been talking with Moravec in 1990, we would've had a further-out median timeline estimate by some amount. This isn't because I think we would've been doing similar estimates to today (we didn't have enough information at the time for this to make much sense), or because I think we would've rejected the framework as irrelevant without today's information. It's simply because we each (myself more than her) have an inclination to apply a fair amount of adjustment in a conservative direction, for generic "burden of proof" reasons, rather than go with the timelines that seem most reasonable based on the report in a vacuum.
  • But more importantly, even if we set the above point aside, I simply don't think it's a mark against Bio Anchors to be in the same reference class as Moravec, and I think his prediction was (according to my views, and more so according to Eliezer's apparent views) impressively good when judged by a reasonable standard and compared to reasonable alternatives.

To expand on what I mean by a reasonable standard and reasonable alternatives:

  • Bio Anchors is, first and foremost, meant as a tool for updating one's timelines from the place they would naively be after considering broader conventional wisdom and perhaps semi-informative priors. Re: the former, I'm referring not to surveys of experts or conventional wisdom in futurist circles (both of which are often dismissed outside of these circles), but to what I perceive as most people's "This is nowhere close to happening, ignore it" intuition.
  • According to my current views (median expectation of transformative AI around 2060), Moravec's 1988 prediction of 2010-2020 looks much better than these alternatives, and even looks impressive. Specifically, it looks impressive by the standards of: "multi-decade forecasting of technologies for which no roadmap exists, with capabilities far exceeding those of anything that exists today." (The more strongly one expects forecasts in this class to be difficult, the more one should be impressed here, in my view.)
  • Eliezer pretty clearly expects shorter timelines than I do, so according to his views, I think Moravec's prediction looks more impressive still (by the standards and alternatives I'm using here). It is implied in the dialogue that Eliezer's median would be somewhere between 2025-2040; if you assume this will turn out to be right, that would make a 1988 prediction of "2010-2020" look extremely good, in my view. (Good enough that, to the extent there's doubt about whether the underlying reasoning is valid or noise, this should be a noticeable update toward the former.)
  • I suspect Eliezer has a different picture of the salient context and alternatives here. I suspect that he's mostly operating in a context where it's near-universal to expect transformative AI at least as early as I do; that he has non-biological-anchor-inspired views that point to much shorter timelines; and that a lot of his piece is a reaction to "Humbali" types (whom he notes are distinct from Open Philanthropy) asking him to update away from his detailed short-timelines views.
  • I'm sympathetic to that, in the sense that I think Bio Anchors is not very useful for the latter purpose. In particular, perhaps it's helpful for me to say here that if you think timelines are short for reasons unrelated to biological anchors, I don't think Bio Anchors provides an affirmative argument that you should change your mind. (I do think it is a useful report for deconstructing - or at least clarifying - several specific, biologically inspired short-timelines arguments that have been floating around, none of which I would guess Eliezer has any interest in.) Most of the case I'd make against shorter timelines would come down to a lack of strong affirmative arguments plus a nontrivial burden of proof.

Returning to the softened version of Platt's Law: according to my current views on timelines (and more so according to Eliezer's), "a few decades" has been a good range for a prediction to be in for the last few decades (again, keeping in mind what context and alternatives I am using). I think this considerably softens the force of an objection like: "You're forecasting a few decades, as many others have over the last few decades; this in itself undermines your case."

None of the above points constitute arguments for the correctness of Bio Anchors. My point is that "Your prediction is like these other predictions" (the thrust of much of Eliezer's piece) doesn't seem to undermine the argument, partly because the other predictions look broadly good according to both my and Eliezer's current views. 

A few other reactions to specific parts

Eliezer: ... The software for a human brain is not going to be 100% efficient compared to the theoretical maximum, nor 10% efficient, nor 1% efficient, even before taking into account the whole thing with parallelism vs. serialism, precision vs. imprecision, or similarly clear low-level differences ...

Eliezer: The makers of AGI aren't going to be doing 10,000,000,000,000 rounds of gradient descent, on entire brain-sized 300,000,000,000,000-parameter models, algorithmically faster than today.  They're going to get to AGI via some route that you don't know how to take, at least if it happens in 2040.  If it happens in 2025, it may be via a route that some modern researchers do know how to take, but in this case, of course, your model was also wrong.

On one hand, I think it's a distinct possibility that we're going to see dramatically new approaches to AI development by the time transformative AI is developed.

On the other, I think quotes like this overstate the likelihood in the short-to-medium term.

  • Deep learning has been the dominant source of AI breakthroughs for nearly the last decade, and the broader "neural networks" paradigm - while it has come in and out of fashion - has broadly been one of the most-attended-to "contenders" throughout the history of AI research.
  • AI research prior to 2012 may have had more frequent "paradigm shifts," but this is probably related to the fact that it was seeing less progress.
  • With these two points in mind, it seems off to me to confidently expect a new paradigm to be dominant by 2040 (even conditional on AGI being developed), as the second quote above implies. As for the first quote, I think the implication there is less clear, but I read it as expecting AGI to involve software well over 100x as efficient as the human brain, and I wouldn't bet on that either (in real life, if AGI is developed in the coming decades - not based on what's possible in principle.)

Eliezer:  The problem is that the resource gets consumed differently, so base-rate arguments from resource consumption end up utterly unhelpful in real life.  The human brain consumes around 20 watts of power.  Can we thereby conclude that an AGI should consume around 20 watts of power, and that, when technology advances to the point of being able to supply around 20 watts of power to computers, we'll get AGI?

If the world were such that:

  • We had some reasonable framework for "power usage" that didn't include gratuitously wasted power, and measured the "power used meaningfully to do computations" in some important sense;
  • AI performance seemed to systematically improve as this sort of power usage increased;
  • Power usage was just now coming within a few orders of magnitude of the human brain;
  • We were just now starting to see AIs have success with tasks like vision and speech recognition (tasks that seem likely to have been evolutionarily important, and that we haven't found ways to precisely describe GOFAI-style);
  • It also looked like AI was starting to have insect-like capabilities somewhere around the time it was consuming insect-level amounts of power;
  • And we didn't have some clear candidate for a better metric with similar properties (as I think we do in the case of computations, since the main thing I'd expect increased power usage to be useful for is increased computation);

...Then I would be interested in a Bio Anchors-style analysis of projected power usage. As noted above, I would be interested in this as a tool for analysis rather than as "the way to get my probability distribution." That's also how I'm interested in Bio Anchors (and how it presents itself).

I also think we have some a priori reason to believe that human scientists can "use computations" somewhere near as efficiently as the brain does (software), more than we have reason to believe that human scientists can "use power" somewhere nearly as efficiently as the brain does (hardware).

(As a side note, there is some analysis of how nature vs. humans use power in this section of Bio Anchors.)

Somebody: All of that seems irrelevant to my novel and different argument. I am not foolishly estimating the resources consumed by a single brain; I'm estimating the resources consumed by evolutionary biology to invent brains!

Eliezer: And the humans wracking their own brains and inventing new AI program architectures and deploying those AI program architectures to themselves learn, will consume computations so utterly differently from evolution that there is no point comparing those consumptions of resources. That is the flaw that you share exactly with Moravec, and that is why I say the same of both of you, "This is a kind of thinking that fails to bind upon reality, it doesn't work in real life." I don't care how much painstaking work you put into your estimate of 10^43 computations performed by biology. It's just not a relevant fact.

It's hard for me to understand how it is not a relevant fact: I think we have good reason to believe that humans can use computations at least as intelligently as evolution did.

I think it's perfectly reasonable to push back on 10^43 as a median estimate, but not as a number that has some sort of relevance.

OpenPhil:  We have commissioned a Very Serious report on a biologically inspired estimate of how much computation will be required to achieve Artificial General Intelligence, for purposes of forecasting an AGI timeline.  (Summary of report.)  (Full draft of report.)  Our leadership takes this report Very Seriously.

I thought this was a pretty misleading presentation of how Open Philanthropy has communicated about this work. It's true that Open Philanthropy's public communication tends toward a cautious, serious tone (and I think there are good reasons for this); but beyond that, I don't think we do much to convey the sort of attitude implied above. The report's publication announcement was on LessWrong as a draft report for comment, and the report is still in the form of several Google docs. We never did any sort of push to have it treated as a fancy report.



 

New Comment
22 comments, sorted by Click to highlight new comments since: Today at 10:49 AM

Thanks for this post!

That being said, my model of Yudkowsky, which I built by spending time interpreting and reverse engineering the post you're responding to, feels like you're not addressing his points (obviously, I might have missed the real Yudkowsky's point)

My interpretation is that he is saying that Evolution (as the generator of most biological anchors) explores the solution space in a fundamentally different path than human research.  So what you have is two paths through a space. The burden of proof for biological anchors thus lies in arguing that there are enough connections/correlations between the two paths to use one in order to predict the other.

Here it sounds like you're taking as an assumption that human research follows the same or a faster path towards the same point in search space. But that's actually the assumption that IMO Yudkowsky is criticizing!

In his piece, Yudkowsky is giving arguments that the human research path should lead to more efficient AGIs than evolution, in part due to the ability of humans to have and leverage insights, which the naive optimization process of evolution can't do. He also points to the inefficiency of biology in implementing new (in geological-time) complex solutions. On the other hand, he doesn't seem to see a way of linking the amount of resources needed by evolution to the amount of resources needed by human research, because they are so different.

If the two paths are very different and don't even aim at the same parts of the search space, there's nothing telling you that computing the optimization power of the first path helps in understanding the second one.

I think Yudkowsky would agree that if you could estimate the amount of resources needed to simulate all evolution  until humans at the level of details that you know is enough to capture all relevant aspects, that amount of resources would be an upper bound on the time taken by human research because that's a way to get AGI if you have the resources. But the number is so vastly large (and actually unknown due to the "level of details" problem) that it's not really relevant for timelines calculations

(Also, I already had this discussion with Daniel Kokotajlo in this thread, but I really think that Platt's law is one of the least cruxy aspects of the original post. So I don't think discussing it further or pointing attention to it is a good idea)

It displeases me that this is currently the most upvoted response: I believe you are focusing on EY's weakest rather than strongest points.

My interpretation is that he is saying that Evolution (as the generator of most biological anchors) explores the solution space in a fundamentally different path than human research. So what you have is two paths through a space. The burden of proof for biological anchors thus lies in arguing that there are enough connections/correlations between the two paths to use one in order to predict the other.

It's hardly surprising there are 'two paths through a space' - if you reran either (biological or cultural/technological) evolution with slightly different initial conditions you'd get a different path. However technological evolution is aware of biological evolution and thus strongly correlated to and influenced by it. IE deep learning is in part brain reverse engineering (explicitly in the case of DeepMind, but there are many other examples). The burden proof is thus arguably more opposite of what you claim (EY claims).

In his piece, Yudkowsky is giving arguments that the human research path should lead to more efficient AGIs than evolution, in part due to the ability of humans to have and leverage insights, which the naive optimization process of evolution can't do. He also points to the inefficiency of biology in implementing new (in geological-time) complex solutions.

To the extent EY makes specific testable claims about the inefficiency of biology, those claims are in err - or at least easily contestable.

EY' strongest point is that the Bio Anchors framework puts far too much weight on scaling of existing models (ie transformers) to AGI, rather than modeling improvement in asymptotic scaling itself. GPT-3 and similar model scaling is so obviously inferior to what is probably possible today - let alone what is possible in the near future - that it should be given very little consideration/weight, just as it would be unwise to model AGI based on scaling up 2005 DL tech.

Thanks for pushing back on my interpretation.

I feel like you're using "strongest" and "weakest" to design "more concrete" and "more abstract", with maybe the value judgement (implicit in your focus on specific testable claims) that concreteness is better. My interpretation doesn't disagree with your point about Bio Anchors, it simply says that this is a concrete instantiation of a general pattern, and that the whole point of the original post as I understand it is to share this pattern. Hence the title who talks about all biology-inspired timelines, the three examples in the post, and the seven times that Yudkowsky repeats his abstract arguments in differents ways.

It's hardly surprising there are 'two paths through a space' - if you reran either (biological or cultural/technological) evolution with slightly different initial conditions you'd get a different path. However technological evolution is aware of biological evolution and thus strongly correlated to and influenced by it. IE deep learning is in part brain reverse engineering (explicitly in the case of DeepMind, but there are many other examples). The burden proof is thus arguably more opposite of what you claim (EY claims).

Maybe a better way of framing my point here is that the optimization processes are fundamentally different (something about which Yudkowsky has written a lot, see for example this post from 13 years ago), and that the burden of proof is on showing that they have enough similarity to extract a lot of info from the evolutionary optimization to the human research optimization.

I also don't think your point about DeepMind works, because DM is working in a way extremely different from evolution. They are in part reverse engineering the brain, but that's a very different (and very human and insight heavy) paths towards AGI than the one evolution took.

Lastly for this point, I don't think the interpretation that "Yudkowsky says that the burden of proof is on showing that the optimization of evolution and human research are non correlated" survives the contact with a text where Yudkowsky constantly berates his interlocutors for assuming such correlation, and keeps drawing again and again the differences.

To the extent EY makes specific testable claims about the inefficiency of biology, those claims are in err - or at least easily contestable.

Hum, I find myself feeling like this comment: Yudkowsky's main point about biology IMO is that brains are not at all the most efficient computational way of implementing AGI. Another way of phrasing it is that Yudkowsky says (according to me) that you could use significantly less hardware and ops/sec to make an AGI.

I don't think I am following the argument here. You seem focused on the comparison with evolution, which is only a minor part of Bio Anchors, and used primarily as an upper bound. (You say "the number is so vastly large (and actually unknown due to the 'level of details' problem) that it's not really relevant for timelines calculations," but actually Bio Anchors still estimates that the evolution anchor implies a ~50% chance of transformative AI this century.)

Generally, I don't see "A and B are very different" as a knockdown counterargument to "If A required ___ amount of compute, my guess is that B will require no more." I'm not sure I have more to say on this point that hasn't already been said - I acknowledge that the comparisons being made are not "tight" and that there's a lot of guesswork, and the Bio Anchors argument doesn't go through without some shared premises and intuitions, but I think the needed intuitions are reasonable and an update from Bio-Anchors-ignorant starting positions is warranted.


 

Thanks for the answer!

Unfortunately, I don't have the time at the moment to answer in detail and have more of a conversation, as I'm fully focused on writing a long sequence about pushing for pluralism in alignment and extracting the core problem out of all the implementation details and additional assumption. I plan on going back to analyzing timeline research in the future, and will probably give better answers then.

That being said, here are quick fire thoughts:

  • I used the evolution case because I consider it the most obvious/straightforward case, in that it sounds so large that everyone instantly assumes that it gives you an upper bound.
  • My general impression about this report (and one I expect Yudkowsky to share) is that it didn't made me update at all. I already updated from GPT and GPT3, and I didn't find new bits of evidence in the report and the discussions around it, despite the length of it. My current impression (please bear in mind that I haven't taken the time to study the report from that angle, so I might change my stance) is that this report, much like a lot of timeline work, seems like it takes as input a lot of assumption, and gives as output far less than was assumed. It's the opposite of compression — a lot of assumptions are needed to conclude things that aren't that strong and constraining.

It's simply because we each (myself more than her) have an inclination to apply a fair amount of adjustment in a conservative direction, for generic "burden of proof" reasons, rather than go with the timelines that seem most reasonable based on the report in a vacuum.

While one can sympathize with the view that the burden of proof ought to lie with advocates of shorter timelines when it comes to the pure inference problem ("When will AGI occur?"), it's worth observing that in the decision problem ("What should we do about it?") this situation is reversed. The burden of proof in the decision problem probably ought instead to lie with advocates of non-action: when one's timelines are >1 generation, it is a bit too easy to kick the can down the road in various ways — leaving one unprepared if the future turns out to move faster than we expected. Conversely someone whose timelines are relatively short may take actions today that will leave us in a better position in the future, even if that future arrives more slowly than they believed originally.

(I don't think OpenPhil is confusing these two, just that in a conversation like this it is particularly worth emphasizing the difference.)

I agree with this. I often default to acting as though we have ~10-15 years, partly because I think leverage is especially high conditional on timelines in that rough range.

Continuing our experiments with the voting system, I've enabled two-axis voting for this thread too.

The two dimensions are:

  • Overall (left and right arrows): what is your overall feeling about the comment? Does it contribute positively to the conversation? Do you want to see more comments like this?
  • Agreement (check and cross): do you agree with the position of this comment?

While Matthew says "Overall I find the law to be pretty much empirically validated, at least by the standards I'd expect from a half in jest Law of Prediction," I don't agree: I don't think an actual trendline on the chart would be particularly close to the Platt's Law line. I think it would, instead, predict that Bio Anchors should point to longer timelines than 30 years out.

I ran OLS regression on the data, and this was the result. Platt's law is in blue.

I agree this trendline doesn't look great for Platt's law, and backs up your observation by predicting that Bio Anchors should be more than 30 years out.

However, OLS is notoriously sensitive to outliers. If instead of using some more robust regression algorithm, we instead super arbitrarily eliminated all predictions after 2100, then we get this, which doesn't look absolutely horrible for the law. Note that the median forecast is 25 years out.

Caveating that I did a lot of skimming on both Bio Anchors and Eliezer's response, the part of Bio Anchors that seemed weakest to me was this:

To be maximally precise, we would need to adjust this probability downward by some amount to account for the possibility that other key resources such as datasets and environments are not available by the time the computation is available

I think the existence of proper datasets/environments is a huge issue for current ML approaches, and you have to assign some nontrivial weight to it being a much bigger bottleneck than computational resources. Like, we're lucky that GPT-3 is trained with the LM objective (predict the next word) for which there is a lot of naturally-occurring training data (written text). Lucky, because that puts us in a position where there's something obvious to do with additional compute. But if we hit a limit following that approach (and I think it's plausible that the signal is too weak in otherwise-unlabeled text) then we're rather stuck. Thus, to get timelines, we'd also need to estimate what dataset/environments are necessary for training AGI. But I'm not sure we know what these datasets/environments look like. An upper bound is "the complete history of earth since life emerged", or something... not sure we know any better.

I think parts of Eliezer's response intersects with this concern, e.g. the energy use analogy. It is the same sort of question, how well do we know what the missing ingredients are? Do we know that compute doesn't occupy enough of the surface area of possible bottlenecks for a compute-based analysis to be worth much? And I'm specifically suggesting that environments/datasets occupy enough of that surface area to seriously undermine the analysis.

Does Bio Anchors deal with this concern beyond the brief mention above (and I missed it, very possible)? Or are there other arguments out there that suggest compute really is all that's missing?

I occasionally hear people make this point but it really seems wrong to me, so I'd like to hear more! Here are the reasons it seems wrong to me:

1. Data generally seems to be the sort of thing you can get more of by throwing money at. It's not universally true but it's true in most cases, and it only needs to be true for at least one transformative or dangerous task. Moreover, investment in AI is increasing; when a tech company is spending $10,000,000,000 on compute for a single AI training run, they can spend 10% as much money to hire 2,000 trained professionals and pay them $100k/yr salaries for 5 years to build training environments and collect/curate data. Not to mention, you are probably doing multiple big AI training runs and you are probably a big tech company that has already set up a huge workforce and data-gathering pipeline with loads of products and stuff.

2. Bitter lesson etc. etc. Most people who I trust on this topic seem to think compute is the main bottleneck, not data.

3. Humans don't need much data; one way we could get transformative or dangerous AI is if we get "human-level AGI," there are plausible versions of this scenario that involve creating something that needs as little or even less data than humans, and while it's possible that the only realistic way to create such a thing is via some giant evolution-like training process that itself requires lots of data... I wouldn't bet on it.

Q. Can you give an example of what progress in data/environments looks like? What progress has been made in this metric over the past 10 years? It'll help me get a sense of what you count as data/environment and what you count as "algorithms." The Bio Anchors report already tracks progress in algorithms.

Final point: I totally agree we should assign nontrivial weight to data being a huge bottleneck, one that won't be overcome by decades and billions of dollars in spending. But everyone agrees on this already; for example, Ajeya assigns 10% of her credence to "basically no amount of compute would be enough with 2020's ideas, not even 10^70 flops." How could recapitulating evolution a billion times over not be enough??? Well, two reasons: Maybe there's some special sauce architecture that can't be found by brute evolutionary search, and maybe data/training environments is an incredibly strong bottleneck. I think the line between these two is blurry. Anyhow, the point is, on my interpretation of Ajeya her framework already assigns something like 5% credence to data/environments being a strong bottleneck. If you think it should be more than that, no problem! Just go into her spreadsheet and make it 25% or whatever you think it is.

Yes, good questions, but I think there are convincing answers. Here's a shot:

1. Some kinds of data can be created this way, like parallel corpora for translation or video annotated with text. But I think it's selection bias that it seems like most cases are like this. Most of the cases we're familiar with seem like this because this is what's easy to do! But transformative tasks are hard, and creating data that really contains latent in it the general structures necessary for task performance, that is also hard. I'm not saying research can't solve it, but that if you want to estimate a timeline, you can't consign this part of the puzzle to a footnote of the form "lots of research resources will solve it". Or, if you do, you might as well relax the whole project and bring only that level of precision across the board.

2. At least in NLP (the AI subfield with which I'm most familiar), my sense of the field's zeitgeist is quite contrary to "compute is the issue". I think there's a large, maybe majority current of thought that our current benchmarks are crap, that performance on them doesn't relate to any interesting real-world task, that optimizing on them is of really unclear value, and that the field as a whole is unfortunately rudderless right now. I think this current holds true among many young DL researchers, not just the Gary Marcuses of the world. That's not a formal survey or anything, just my sense from reading NLP papers and twitter. But similarly, I think the notion that compute is the bottleneck is overrepresented in the LessWrong sphere, vs. what others think.

3. Humans not needing much data is misleading IMO because the human brain comes highly optimized out of the box at birth, and indeed that's the result of a big evolutionary process. To be clear, I agree achieving human-level AI is enough to count as transformative and may well be a second-long epoch on the way to much more powerful AI. But anyway, you have basically the same question to answer there. Namely, I'd still object that Bio Anchors doesn't address the datasets/environments issue regarding making even just human-level AI. Changing the scope to "merely" human doesn't answer the objection.

Q/A. As for recent progress: no, I think there has been very little! I'm only really familiar with NLP, so there might be more in the RL environments. (My very vague sense of RL is that it's still just "video games you can put an agent is" and basically always has been, but don't take it from me.) As for NLP, there is basically nothing new in the last 10 years. We have lots of unlabeled text for language models, we have parallel corpora for translation, and we have labeled datasets for things like question-answering (see here for a larger list of supervised tasks). I think it's really unclear whether any of these have latent in them the structures necessary for general language understanding. GPT is the biggest glimmer of hope recently, but part of the problem even there is we can't even really quantify how close it is to general language understanding. We don't have a good way of measuring this! Without it, we certainly can't train, as we can't compute a loss function. I think there are maybe some arguments that, in the limit, unlabeled text with the LM objective is enough: but that limit might really be more text than can fit on earth, and we'd need to get a handle on that for any estimates.

Final point: I'm more looking for a qualitative acknowledgement that this problem of datasets/environments is hard and unsolved (or arguments for why it isn't), is as important as compute, and, building on that, serious attention paid to an analysis of what it would take to make the right datasets/environments. Rather than consign it to an "everything else" parameter, analyze what it might take to make better datasets/environments, including trying to get a handle on whether we even know how. I think this would make for a much better analysis, and would address some of Eliezer's concerns because it would cover more of the specific, mechanistic story about the path to creating transformative AI.

(Full disclosure: I've personally done work on making better NLP benchmarks, which I guess has given me an appreciation for how hard and unsolved this problem feels. So, discount appropriately.)

Thanks, nice answers!

I agree it would be good to extend the bio anchors framework to include more explicit modelling of data requirements and the like instead of just having it be implicit in the existing variables. I'm generally a fan of making more detailed, realistic models and this seems reasonably high on the priority list of extensions to make. I'd also want to extend the model to include multiple different kinds of transformative task and dangerous task, and perhaps to include interaction effects between them (e.g. once we get R&D acceleration then everything else may come soon...) and perhaps to make willingness-to-spend not increase monotonically but rather follow a boom and bust cycle with a big boom probably preceding the final takeoff.

I still don't think the problem of datasets/environments is hard and unsolved, but I would like to learn more.

What's so bad about text prediction / entropy (GPT-3's main metric) as a metric for NLP? I've heard good things about it, e.g. this.

Re: Humans not needing much data: They are still proof of concept that you don't need much data (and/or not much special, hard-to-acquire data, just ordinary life experience, access to books and conversations with people, etc.) to be really generally intelligent. Maybe we'll figure out how to make AIs like this too. MAYBE the only way to do this is to recapitulate a costly evolutionary process that itself is really tricky to do and requires lots of data that's hard to get... but I highly doubt it. For example, we might study how the brain works and copy evolution's homework, so to speak. Or it may turn out that most of the complexity and optimization of the brain just isn't necessary. To your "Humans not needing much data is misleading IMO because the human brain comes highly optimized out of the box at birth, and indeed that's the result of a big evolutionary process" I reply with this post.

I still don't think we have good reason to think ALL transformative tasks or dangerous tasks will require data that's hard to get. You say this stuff is hard, but all we know is that it's too hard for the amount of effort that's been expended so far. For all we know it's not actually that much harder than that.

If there hasn't been much progress in the last ten years on the data problem, then... do you also expect there to be not much progress in the next ten years either? And do you thus have, like, 50-year timelines?

Re: humans/brains, I think what humans are a proof of concept of is that, if you start with an infant brain, and expose it to an ordinary life experience (a la training / fine-tuning), then you can get general intelligence. But I think this just doesn't bear on the topic of Bio Anchors, because Bio Anchors doesn't presume we have a brain, it presumes we have transformers. And transformers don't know what to do with a lifetime of experience, at least nowhere near as well as an infant brain does. I agree we might learn more about AI from examining humans! But that's leaving the Bio Anchors framing of "we just need compute" and getting into the framing of algorithmic improvements etc. I don't disagree broadly that some approaches to AI might not have as big a (pre-)training phase the way current models do, if, for instance, they figure out a way to "start with" infant brains. But I don't see the connection to the Bio Anchors framing.

What's so bad about perplexity? I'm not saying perplexity is bad per se, just that it's unclear how much data you need, with perplexity as your objective, to achieve general-purpose language facility.  It's unclear both because the connection between perplexity and extrinsic linguistic tasks is unclear, and because we don't have great ways of measuring extrinsic linguistic tasks. For instance, the essay you cite itself cites two very small experiments showing correlation between perplexity and extrinsic tasks. One of them is a regression on 8 data points, the other has 24 data points. So I just wouldn't put too much stake in extrapolations there. Furthermore, and this isn't against perplexity, but I'd be skeptical of the other variable i.e. the linguistic task perplexity is regressed against: in both cases, a vague human judgement of whether model output is "human-like". I think there's not much reason to think that is correlated to some general-purpose language facility. Attempts like this to (roughly speaking) operationalize the Turing test have generally been disappointing; getting humans to say "sure, that sounds human" seems to leave a lot to be desired; I think most researchers find them to be disappointingly game-able, vague a critique though that may be. The google Meena paper acknowledges this, and reading between the lines I get the sense they don't think too much of their extrinsic, human-evaluation metric either. E.g., the best they say is, "[inter-annotator] agreement is reasonable considering the questions are subjective and the final results are always aggregated labels".

This is sort of my point in a nutshell: we have put very little effort into telling whether the datasets we have contain adequate signal to learn the functions we want to learn, in part because we aren't even sure how to evaluate those functions. It's not surprising that perplexity correlates with extrinsic tasks to a degree. For instance, it's pretty clear that, to get state-of-the-art low perplexity on existing corpora, transformers can learn the latent rules of grammar, and, naturally doing so correlates with better human judgements of model output. So, grammar is latent in the corpora. But is physics latent in the corpora? It would improve a model's perplexity at least a bit to learn physics: some of these corpora contain physics textbooks with answers to exercises way at the back of the books, so to predict the answers at the back you would have to be able to learn how to do the exercises. But it's unclear whether current corpora contain enough signal to do that. Would we even know how to tell if the model was or wasn't learning physics? I'm personally skeptical that it's happening at all, but I admit that's just based in my subjective assessment of GPT-3 output... again, part of the problem of not having a good way to judge performance outside of perplexity.

As for why all transformative tasks might have hard-to-get-data... well this is certainly speculative, but people sometimes talk about AI-complete tasks, analogizing to the concept of completeness for complexity classes in complexity theory (e.g., NP-complete). I think that's the relevant idea here. The goal being general intelligence, I think it's plausible that most (all? I don't know) transformative tasks are reducible to each other. And I think you also get a hint of this in NLP tasks, where they are weirdly reducible to each other, given the amazing flexibility of language. Like, for a dumb example, the task of question answering entails the task of translation, because you can ask, "How do you say [passage] in French?" So I think the sheer number of tasks, as initially categorized by humans, can be misleading. Tasks aren't as independent as they may appear. Anyway, that's far from a tight argument, but hopefully it provides some intuition.

Honestly I haven't thought about how to incorporate the dataset bottleneck into a timeline. But, I suppose, I could wind up with even longer timelines if I think that we haven't made progress because we don't have the faintest idea how and the lack of progress isn't for lack of trying. Missing fundamental ideas, etc. How do you forecast when a line of zero slope eventually turns up? If I really think we have shown ourselves to be stumped (not sure), I guess I'd have to fall back on general-purpose tools for forecasting big breakthroughs, and that's the sort of vague modeling that Bio Anchors seems to be trying to avoid.

BioAnchors is poorly named, the part you are critiquing should be called GPT-3_Anchors.

A better actual BioAnchor would be based on trying to model/predict how key params like data efficiency and energy efficiency are improving over time, and when they will match/surpass the brain.

GPT-3 could also obviously be improved for example by multi-modal training, active learning, curriculum learning, etc. It's not like it even represents the best of what's possible for a serious AGI attempt today.

Fwiw, I think nostalgebraist's recent post hit on some of the same things I was trying to get at, especially around not having adequate testing to know how smart the systems are getting -- see the section on what he calls (non-)ecological evaluation.

With apologies for the belated response: I think greghb makes a lot of good points here, and I agree with him on most of the specific disagreements with Daniel. In particular:

  • I agree that "Bio Anchors doesn't presume we have a brain, it presumes we have transformers. And transformers don't know what to do with a lifetime of experience, at least nowhere near as well as an infant brain does." My guess is that we should not expect human-like sample efficiency from a simple randomly initialized network; instead, we should expect to extensively train a network to the point where it can do this human-like learning. (That said, this is far from obvious, and some AI scientists take the opposite point of view.)
  • I'm not super sympathetic to Daniel's implied position that there are lots of possible transformative tasks and we "only need one" of them. I think there's something to this (in particular, we don't need to replicate everything humans can do), but I think once we start claiming that there are 5+ independent tasks such that automating them would be transformative, we have to ask ourselves why transformative events are as historically rare as they are. (More at my discussion of persuasion on another thread.)

Overall, I think that datasets/environments are plausible as a major blocker to transformative AI, and I think Bio Anchors would be a lot stronger if it had more to say about this. 

I am sympathetic to Bio Anchors's bottom-line quantitative estimates despite this, though (and to be clear, I held all of these positions at the time Bio Anchors was drafted). It's not easy for me to explain all of where I'm coming from, but a few intuitions:

  • We're still in a regime where compute is an important bottleneck to AI development, and funding and interest are going up. If we get into a regime where compute is plentiful and data/environments are the big blocker, I expect efforts to become heavily focused there.
  • Several decades is just a very long time. (This relates to the overall burden of proof on arguments like these, particularly the fact that this century is likely to see most of the effort that has gone into transformative AI development to date.)
  • Combining the first two points leaves me guessing that "if there's a not-prohibitively-difficult way to do this, people are going to find it on the time frames indicated." And I think there probably is:
    • The Internet contains a truly massive amount of information at this point about many different dimensions of the human world. I expect this information source to keep growing, especially as AI advances and interacts more productively and richly with humans, and as AI can potentially be used as an increasingly large part of the process of finding data, cleaning it, etc.
    • AI developers will also - especially as funding and interest grow - have the ability to collect data by (a) monitoring researchers, contractors, volunteers, etc.; (b) designing products with data collection in mind (e.g., Assistant and Siri). 
    • The above two points seem especially strong to me when considering that automating science and engineering might be sufficient for transformative AI - these seem particularly conducive to learning from digitally captured information.
    • On a totally separate note, it seems to me that fairly simple ingredients have made the historical human "environment" sufficiently sophisticated to train transformative capabilities. It seems to me that most of what's "interesting and challenging" about our environment comes from competing with each other, and I'd guess it's possible to set up some sort of natural-selection-driven environment in which AIs compete with each other; I wouldn't expect such a thing to be highly sensitive to whether we're able to capture all of the details of e.g. how to get food that occurred in our past. (I would expect it to be compute-intensive, though.)

Hopefully that gives a sense of where I'm coming from. Overall, I think this is one of the most compelling objections to Bio Anchors; I find it stronger than the points Eliezer focuses on above (unless you are pretty determined to steel-man any argument along the lines of "Brains and AIs are different" into a specific argument about the most important difference.)

With these two points in mind, it seems off to me to confidently expect a new paradigm to be dominant by 2040 (even conditional on AGI being developed), as the second quote above implies. As for the first quote, I think the implication there is less clear, but I read it as expecting AGI to involve software well over 100x as efficient as the human brain, and I wouldn’t bet on that either (in real life, if AGI is developed in the coming decades—not based on what’s possible in principle.)

I think this misses the point a bit. The thing to be afraid of is not an all-new approach to replace neural networks, but rather new neural network architectures and training methods that are much more efficient than today's. It's not unreasonable to expect those, and not unreasonable to expect that they'll be much more efficient than humans, given how easy it is to beat humans at arithmetic for example, and given fast recent progress to superhuman performance in many other domains.

Unless I’m mistaken, the Bio Anchors framework explicitly assumes that we will continue to get algorithmic improvements, and even tries to estimate and extrapolate the trend in algorithmic efficiency. It could of course be that progress in reality will turn out a lot faster than the median trendline in the model, but I think that’s reflected by the explicit uncertainty over the parameters in the model.

In other words, Holden’s point about this framework being a testbed for thinking about timelines remains unscathed if there is merely more ordinary algorithmic progress than expected.

Curated. A few weeks ago I curated the post this is a response to. I'm excited to see a response that argues the criticized report was misinterpreted/misrepresented. I'd be even more excited to see a response to the response–by the authors involved so far or anyone else. Someone once said (approx) that successful conversation must permit four levels: saying X, critiquing X, critiquing the critique, and critiquing the critique of the critique. We're at 3 out of 4.

if you think timelines are short for reasons unrelated to biological anchors, I don't think Bio Anchors provides an affirmative argument that you should change your mind.

 

Eliezer:  I wish I could say that it probably beats showing a single estimate, in terms of its impact on the reader.  But in fact, writing a huge careful Very Serious Report like that and snowing the reader under with Alternative Calculations is probably going to cause them to give more authority to the whole thing.  It's all very well to note the Ways I Could Be Wrong and to confess one's Uncertainty, but you did not actually reach the conclusion, "And that's enough uncertainty and potential error that we should throw out this whole deal and start over," and that's the conclusion you needed to reach.

I would be curious to know what the intended consequences of the forecasting piece were.

A lot of Eliezer's argument seems to me to be pushing at something like 'there is a threshold for how much evidence you need before you start putting down numbers, and you haven't reached it', and I take what I've quoted from your piece to be supporting something like 'there is a threshold for how much evidence you might have, and if you're above it (and believe this forecast to be an overestimate) then you may be free to ignore the numbers here', contra the Humbali position. I'm not particularly confident on that, though.

Where this leaves me is feeling like you two have different beliefs about who will (or should) update on reading this kind of thing, and to what end, which is probably tangled up in beliefs about how good people are at holding uncertainty in their mind. But I'm not really sure what these beliefs are.

The Bio Anchors report is intended as a tool for making debates about AI timelines more concrete, for those who find some bio-anchor-related bound helpful (e.g., some think we should lower bound P(AGI) at some reasonably high number for any year in which we expect to hit a particular kind of "biological anchor"). Ajeya's work lengthened my own timelines, because it helped me understand that some bio-anchor-inspired arguments for shorter timelines didn't have as much going for them as I'd thought; but I think it may have shortened some other folks'.

(The presentation of the report in the Most Important Century series had a different aim. That series is aimed at making the case that we could be in the most important century, to a skeptic.)

I don't personally believe I have a high-enough-quality estimate using another framework that I'd be justified in ignoring bio-anchors-based reasoning, but I don't think it's wild to think someone else might have such an estimate.