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.
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.)
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.)