Someone who is interested in learning and doing good.
My Substack: https://matthewbarnett.substack.com/
I think I understand my confusion, at least a bit better than before. Here's how I'd summarize what happened.
I had three arguments in this essay, which I thought of as roughly having the following form:
You said that (2) was already answered by the bio anchors model. I responded that bio anchors neglected how difficult it will be to develop AI safely. You replied that it will be easy make models to seemingly do what we want, but that the harder part will be making models that actually do what we want.
My reply was trying to say that the inherent difficulty of building TAI safely was inherently baked into (2) already. That might be a dubious reading of the actual textual argument for (2), but I think that interpretation is backed up by my initial reply to your comment.
The reason why I framed my later reply as being about perceptions was because I think the requisite capability level at which people begin to adopt TAI is an important point about how long timelines will be independent of (1) and (3). In other words, I was arguing that people's perceptions of the capability of AI will cause them wait to adopt AI until it's fully developed in the sense I described above; it won't just delay the effects of TAI after it's fully developed, or before then because of regulation.
Furthermore, I assumed that you were arguing something along the lines of "people will adopt AI once it's capable of only seeming to do what we want", which I'm skeptical of. Hence my reply to you.
My understanding was that you are also skeptical about question 2 on short timelines, and that was what you were arguing with your point (2) on overestimating generality.
Since for point 2 you said "I'm assuming that an AI CEO that does the job of CEO well until the point that it executes a treacherous turn", I am not very skeptical of that right now. I think we could probably have AIs do something that looks very similar to what a CEO would do within, idk, maybe five years. [ETA: And I have only slightly longer timelines for stuff that could be really dangerous.]
(Independently of all of this, I've updated towards medium rather than long timelines in the last two years, but mostly because of reflection on other questions, and because I was surprised by the rate of recent progress, rather than because I have fundamental doubts about the arguments I made here, especially (3), which I think is still underrated.
ETA: though also, if I wrote this essay today I would likely fully re-write section (2), since after re-reading it I now don't agree with some of the things I said in it. Sorry if I was being misleading by downplaying how poor some of those points were.)
Sorry for replying to this comment 2 years late, but I wanted to discuss this part of your reasoning,
Fwiw, the problem I think is hard is "how to make models do stuff that is actually what we want, rather than only seeming like what we want, or only initially what we want until the model does something completely different like taking over the world".
I think that's what I meant when I said "I think it will be hard to figure out how to actually make models do stuff we want". But more importantly, I think that's how most people will in fact perceive what it means to get a model to "do what we want".
Put another way, I don't think people will actually start using AI CEOs just because we have a language model that acts like a CEO. Large corporations will likely wait until they're very confident in its reliability, robustness, and alignment. (Although idk, maybe some eccentric investors will find the idea interesting, I just expect that most people will be highly skeptical without strong evidence that it's actually better than a human.)
I think this point can be seen pretty easily in discussion of driverless cars. Regulators are quite skeptical of Tesla's autopilot despite it seeming to do what we want in perhaps over 99% of situations.
If anything, I expect most people to be intuitively skeptical that AI is really "doing what we want" even in cases where it's genuinely doing a better job than humans, and doesn't merely appear that way on the surface. The reason is simple: we have vast amounts of informal data on the reliability of humans, but very little idea how reliable AI will be. That plausibly causes people to start with a skeptical outlook, and only accept AI in safety-critical domains when they've seen it accumulate a long track record of exceptional performance.
For these reasons, I don't fully agree that "one of the major points of the bio anchors framework is to give a reasonable answer to the question of "at what level of scaling might this work"". I mean, I agree that this was what the report was trying to answer, but I disagree that it answered the question of when we will accept and adopt AI for various crucial economic activities, even if such systems were capable of automating everything in principle.
Some people seem to be hoping that nobody will ever make a misaligned human-level AGI thanks to some combination of regulation, monitoring, and enlightened self-interest. That story looks more plausible if we’re talking about an algorithm that can only run on a giant compute cluster containing thousands of high-end GPUs, and less plausible if we’re talking about an algorithm that can run on one 2023 gaming PC.
Isn't the relevant fact whether we could train an AGI with modest computational resources, not whether we could run one? If training runs are curtailed from regulation, then presumably the main effect is that AGI will be delayed until software and hardware progress permits the covert training of an AGI with modest computational resources, which could be a long time depending on how hard it is to evade the regulation.
I don't think this is right -- the main hype effect of chatGPT over previous models feels like it's just because it was in a convenient chat interface that was easy to use and free.
I don't have extensive relevant expertise, but as a personal datapoint: I used Davinci-002 multiple times to generate an interesting dialogue in order to test its capabilities. I ran several small-scale Turing tests, and the results were quite unimpressive in my opinion. When ChatGPT came out, I tried it out (on the day of its release) and very quickly felt that it was qualitatively better at dialogue. Of course, I could have simply been prompting Davinci-002 poorly, but overall I'm quite skeptical that the main reason for ChatGPT hype was that it had a more convenient chat interface than GPT-3.
I have now published a conversation between Ege Erdil and Ronny Fernandez about this post. You can find it here.
Let me restate some of my points, which can hopefully make my position clearer. Maybe state which part you disagree with:
Language models are probability distributions over finite sequences of text.
The “true distribution” of internet text refers to a probability distribution over sequences of text that you would find on the internet (including sequences found on other internets elsewhere in the multiverse, which is just meant as an abstraction).
A language model is “better” than another language model to the extent that the cross-entropy between the true distribution and the model is lower.
A human who writes a sequence of text is likely to write something with a relatively high log probability relative to the true distribution. This is because in a quite literal sense, the true distribution is just the distribution over what humans actually write.
A current SOTA model, by contrast, is likely to write something with an extremely low log probability, most likely because it will write something that lacks long-term coherence, and is inhuman, and thus, won’t be something that would ever appear in the true distribution (or if it appears, it appears very very very rarely).
The last two points provide strong evidence that humans are actually better at the long-sequence task than SOTA models, even though they’re worse at the next character task.
Intuitively, this is because the SOTA model loses a gigantic amount of log probability when it generates whole sequences that no human would ever write. This doesn’t happen on the next character prediction task because you don’t need a very good understanding of long-term coherence to predict the vast majority of next-characters, and this effect dominates the effect from a lack of long-term coherence in the next-character task.
It is true (and I didn’t think of this before) that the human’s cross entropy score will probably be really high purely because they won’t even think to have any probability on some types of sequences that appear in the true distribution. I still don’t think this makes them worse than SOTA language models because the SOTA will also have ~0 probability on nearly all actual sequences. However…
Even if you aren’t convinced by my last argument, I can simply modify what I mean by the “true distribution” to mean the “true distribution of texts that are in the reference class of things we care about”. There’s absolutely no reason to say the true distribution has to be “everything on the internet” as opposed to “all books” or even “articles written by Rohin” if that’s what we’re actually trying to model.
Thus, I don’t accept one of your premises. I expect current language models to be better than you at next-character prediction on the empirical distribution of Rohin articles, but worse than you at whole sequence prediction for Rohin articles, for reasons you seem to already accept.
Large language models are also going to be wildly superhuman by long-sequence metrics like "log probability assigned to sequences of Internet text"
I think this entirely depends on what you mean. There's a version of the claim here that I think is true, but I think the most important version of it is actually false, and I'll explain why.
I claim that if you ask a human expert to write an article (even a relatively short one) about a non-trivial topic, their output will have a higher log probability than a SOTA language model, with respect to the "true" distribution of internet articles. That is, if you were given the (entirely hypothetical) true distribution of actual internet articles (including articles that have yet to be written, and the ones that have been written in other parts of the multiverse...), a human expert is probably going to write an article that has a higher log probability of being sampled from this distribution, compared to a SOTA language model.
This claim might sound bizarre at first, because, as you noted "many such metrics are just sums over the next-character versions of the metric, which this post shows LLMs are great at". But, first maybe think about this claim from first principles: what is the "true" distribution of internet articles? Well, it's the distribution of actual internet articles that humans write. If a human writes an article, it's got to have pretty high log-probability, no? Because otherwise, what are we even sampling from?
Now, what you could mean is that instead of measuring the log probability of an article with respect to the true distribution of internet articles, we measure it with respect to the empirical distribution of internet articles. This is in fact what we use to measure the log-probability of next character predictions. But the log probability of this quantity over long sequences will actually be exactly negative infinity, both for the human-written article, and for the model-written article, assuming they're not just plagiarizing an already-existing article. That is, we aren't going to find any article in the empirical distribution that matches the articles either the human or the model wrote, so we can't tell which of the two is better from this information alone.
What you probably mean is that we could build a model of the true distribution of internet articles, and use this model to estimate the log-probability of internet articles. In that case, I agree, a SOTA language model would probably far outperform the human expert, at the task of writing internet articles, as measured by the log-probability given by another model. But, this is a flawed approach, because the model we're using to estimate the log-probability with respect to the true distribution of internet articles is likely to be biased in favor of the SOTA model, precisely because it doesn't understand things like long-sequence coherence, unlike the human.
How could we modify this approach to give a better estimate of the performance of a language model at long-sequence prediction? I think that there's a relatively simple approach that could work.
Namely, we set up a game in which humans try to distinguish between real human texts and generated articles. If the humans can't reliably distinguish between the two, then the language model being used to generate the articles has attained human-level performance (at least by this measure). This task has nice properties, as there is a simple mathematical connection between prediction ability and ability to discriminate; a good language model that can pass this test will likely only pass it because it is good at coming up with high log-probability articles. And this task also measures the thing we care about that’s missing from the predict-the-next-character task: coherence over long sequences.
Ah, I see your point. That being said, I think calling the task we train our LMs to do (learn a probabilistic model of language) "language modeling" seems quite reasonable to me - in my opinion, it seems far more unreasonable to call "generating high quality output" "language modeling".
Note that the main difference between my suggested task and the next-character-prediction task is that I'm suggesting we measure performance over a long time horizon. "Language models" are, formally, probability distributions over sequences of text, not models over next characters within sequences. It is only via a convenient application of the Markov assumption and the chain rule of probability that we use next-character-prediction during training.
The actual task, in the sense of what language models are fundamentally designed to perform well on, is to emulate sequences of human text. Thus, it is quite natural to ask when they can perform well on this task. In fact, I remain convinced that it is more natural to ask about performance on the long-sequence task than the next-character-prediction task.
We disagree that this measure is better. Our goal here isn't to compare the quality of Language Models to the quality of human-generated text; we aimed to compare LMs and humans on the metric that LMs were trained on (minimize log loss/perplexity when predicting the next token).
Your measure is great for your stated goal. That said, I feel the measure gives a misleading impression to readers. In particular I'll point to this paragraph in the conclusion,
Even current large language models are wildly superhuman at language modeling. This is important to remember when you’re doing language model interpretability, because it means that you should expect your model to have a lot of knowledge about text that you don’t have. Chris Olah draws a picture where he talks about the possibility that models become more interpretable as they get to human level, and then become less interpretable again as they become superhuman; the fact that existing LMs are already superhuman (at the task they’re trained on) is worth bearing in mind when considering this graph.
I think it's misleading to say that language models are "wildly superhuman at language modeling" by any common-sense interpretation of that claim. While the claim is technically true if one simply means that languages do better at the predict-the-next-token task, most people (I'd imagine) would not intuitively imagine that to be the best measure of general performance at language modeling. The reason, fundamentally, is that we are building language models to compete with other humans at the task of writing text, not the task of predicting the next character.
By analogy, if we train a robot to play tennis by training it to emulate human tennis players, I think most people would think that "human level performance" is reached when it can play as well as a human, not when it can predict the next muscle movement of an expert player better than humans, even if predicting the next muscle movement was the task used during training.
Building on this comment, I think it might be helpful for readers to make a few distinctions in their heads: