I am trying to wrap my head around the high-level implications of this statement. I can come up with two interpretations:

1. What LLMs are doing is similar to what people do as they go about their day. When I walk down the street, I am simultaneously using visual and other input to assess the state of the world around me ("that looks like a car"), running a world model based on that assessment ("the car is coming this way"), and then using some other internal mechanism to decide what to do ("I'd better move to the sidewalk").
2. What LLMs are doing is harder than what people do. When I converse with someone, I have some internal state, and I run some process in my head – based on that state – to generate my side of the conversation. When an LLM converses with someone, instead of maintaining internal state, needs to maintain a probability distribution over possible states, make next-token predictions according to that distribution, and simultaneously update the distribution.

(2) seems more technically correct, but my intuition dislikes the conclusion, for reasons I am struggling to articulate. ...aha, I think this may be what is bothering me: I have glossed over the distinction between input and output tokens. When an LLM is processing input tokens, it is working to synchronize its state to the state of the generator. Once it switches to output mode, there is no functional benefit to continuing to synchronize state (what is it synchronizing to?), so ideally we'd move to a simpler neural net that does not carry the weight of needing to maintain and update a probability distribution of possible states. (Glossing over the fact that LLMs as used in practice sometimes need to repeatedly transition between input and output modes.) LLMs need the capability to ease themselves into any conversation without knowing the complete history of the participant they are emulating, while people have (in principle) access to their own complete history and so don't need to be able to jump into a random point in their life and synchronize state on the fly.

So the implication is that the computational task faced by an LLM which can emulate Einstein is harder than the computational task of being Einstein... is that right? If so, that in turn leads to the question of whether there are alternative modalities for AI which have the advantages of LLMs (lots of high-quality training data) but don't impose this extra burden. It also raises the question of how substantial this burden is in practice, in particular for leading-edge models.

All of this is plausible, but I'd encourage you to go through the exercise of working out these ideas in more detail. It'd be interesting reading and you might encounter some surprises / discover some things along the way.

Note, for example, that the AGIs would be unlikely to focus on AI research and self-improvement if there were more economically valuable things for them to be doing, and if (very plausibly!) there were not more economically valuable things for them to be doing, why wouldn't a big chunk of the 8 billion humans have been working on AI research already (such that an additional 1.6 million agents working on this might not be an immediate game changer)? There might be good arguments to be made that the AGIs would make an important difference, but I think it's worth spelling them out.

Can you elaborate? This might be true but I don't think it's self-evidently obvious.

In fact it could in some ways be a disadvantage; as Cole Wyeth notes in a separate top-level comment, "There are probably substantial gains from diversity among humans". 1.6 million identical twins might all share certain weaknesses or blind spots.

Assuming we require a performance of 40 tokens/s, the training cluster can run $\frac{2000}{30}\times 24000=1,600,000$ concurrent instances of the resulting 70B model

Nit: you mixed up 30 and 40 here (should both be 30 or both be 40).

I will assume that the above ratios hold for an AGI level model.

If you train a model with 10x as many parameters, but use the same training data, then it will cost 10x as much to train and 10x as much to operate, so the ratios will hold.

In practice, I believe it is universal to use more training data when training larger models? Implying that the ratio would actually increase (which further supports your thesis).

On the other hand, the world already contains over 8 billion human intelligences. So I think you are assuming that a few million AGIs, possibly running at several times human speed (and able to work 24/7, exchange information electronically, etc.), will be able to significantly "outcompete" (in some fashion) 8 billion humans? This seems worth further exploration / justification.

They do mention a justification for the restrictions – "to maintain consistency across cells". One needn't agree with the approach, but it seems at least to be within the realm of reasonable tradeoffs.

Nowadays of course textbooks are generally available online as well. They don't indicate whether paid materials are within scope, but of course that would be a question for paper textbooks as well.

What I like about this study is that the teams are investing a relatively large amount of effort ("Each team was given a limit of seven calendar weeks and no more than 80 hours of red-teaming effort per member"), which seems much more realistic than brief attempts to get an LLM to answer a specific question. And of course they're comparing against a baseline of folks who still have Internet access.

I recently encountered a study which appears aimed at producing a more rigorous answer to the question of how much use current LLMs would be in abetting a biological attack: https://www.rand.org/pubs/research_reports/RRA2977-1.html. This is still work in progress, they do not yet have results. @1a3orn I'm curious what you think of the methodology?

Imagine someone offers you an extremely high-paying job. Unfortunately, the job involves something you find morally repulsive – say, child trafficking. But the recruiter offers you a pill that will rewrite your brain chemistry so that you'll no longer find it repulsive. Would you take the pill?

I think that pill would reasonably be categorized as "updating your goals". If you take it, you can then accept the lucrative job and presumably you'll be well positioned to satisfy your new/remaining goals, i.e. you'll be "happy". But you'd be acting against your pre-pill goal (I am glossing over exactly what that goal is, perhaps "not harming children" although I'm sure there's more to unpack here).

I pose this example in an attempt to get at the heart of "distinguishing between terminal and instrumental goals" as suggested by quetzal_rainbow. This is also my intuition, that it's a question of terminal vs. instrumental goals.

Likewise, thanks for the thoughtful and detailed response. (And I hope you aren't too impacted by current events...)

I agree that if no progress is made on long-term memory and iterative/exploratory work processes, we won't have AGI. My position is that we are already seeing significant progress in these dimensions and that we will see more significant progress in the next 1-3 years. (If 4 years from now we haven't seen such progress I'll admit I was totally wrong about something). Maybe part of the disagreement between us is that the stuff you think are mere hacky workarounds, I think might work sufficiently well (with a few years of tinkering and experimentation perhaps).

Wanna make some predictions we could bet on? Some AI capability I expect to see in the next 3 years that you expect to not see?

Sure, that'd be fun, and seems like about the only reasonable next step on this branch of the conversation. Setting good prediction targets is difficult, and as it happens I just blogged about this. Off the top of my head, predictions could be around the ability of a coding AI to work independently over an extended period of time (at which point, it is arguably an "engineering AI"). Two different ways of framing it:

1. An AI coding assistant can independently complete 80% of real-world tasks that would take X amount of time for a reasonably skilled engineer who is already familiar with the general subject matter and the project/codebase to which the task applies.
2. An AI coding assistant can usefully operate independently for X amount of time, i.e. it is often productive to assign it a task and allow it to process for X time before checking in on it.

At first glance, (1) strikes me as a better, less-ambiguous framing. Of course it becomes dramatically more or less ambitious depending on X, also the 80% could be tweaked but I think this is less interesting (low percentages allow for a fluky, unreliable AI to pass the test; very high percentages seem likely to require superhuman performance in a way that is not relevant to what we're trying to measure here).

It would be nice to have some prediction targets that more directly get at long-term memory and iterative/exploratory work processes, but as I discuss in the blog post, I don't know how to construct such a target – open to suggestions.

Coding, in the sense that GPT4 can do it, is nowhere near the top of the hierarchy of skills involved in serious software engineering. And so I believe this is a bit like saying that, because a certain robot is already pretty decent at chiseling, it will soon be able to produce works of art at the same level as any human sculptor. 

I think I just don't buy this. I work at OpenAI R&D. I see how the sausage gets made. I'm not saying the whole sausage is coding, I'm saying a significant part of it is, and moreover that many of the bits GPT4 currently can't do seem to me that they'll be doable in the next few years.

Intuitively, I struggle with this, but you have inside data and I do not. Maybe we just set this point aside for now, we have plenty of other points we can discuss.

To be clear, I do NOT think that today's systems could replace 99% of remote jobs even with a century of schlep. And in particular I don't think they are capable of massively automating AI R&D even with a century of schlep. I just think they could be producing, say, at least an OOM more economic value. ...

This, I would agree with. And on re-reading, I think I may have been mixed up as to what you and Ajeya were saying in the section I was quoting from here, so I'll drop this.

[Ege] I think when you try to use the systems in practical situations; they might lose coherence over long chains of thought, or be unable to effectively debug non-performant complex code, or not be able to have as good intuitions about which research directions would be promising, et cetera.

This was a nice answer from Ege. My follow up questions would be: Why? I have theories about what coherence is and why current models often lose it over long chains of thought (spoiler: they weren't trained to have trains of thought) and theories about why they aren't already excellent complex-code-debuggers (spoiler: they weren't trained to be) etc. What's your theory for why all the things AI labs will try between now and 2030 to make AIs good at these things will fail?

I would not confidently argue that it won't happen by 2030; I am suggesting that these problems are unlikely to be well solved in a usable-in-the-field form by 2027 (four years from now). My thinking:

1. The rapid progress in LLM capabilities has been substantially fueled by the availability of stupendous amounts of training data.
2. There is no similar abundance of low-hanging training data for extended (day/week/more) chains of thought, nor for complex debugging tasks. Hence, it will not be easy to extend LLMs (and/or train some non-LLM model) to high performance at these tasks.
3. A lot of energy will go into the attempt, which will eventually succeed. But per (2), I think some new techniques will be needed, which will take time to identify, refine, scale, and productize; a heavy lift in four years. (Basically: Hofstadter's Law.)
4. Especially because I wouldn't be surprised if complex-code-debugging turns out to be essentially "AGI-complete", i.e. it may require a sufficiently varied mix of exploration, logical reasoning, code analysis, etc. that you pretty much have to be a general AGI to be able to do it well.

I understand you might be skeptical that it can be done but I encourage you to red-team your position, and ask yourself 'how would I do it, if I were an AI lab hell-bent on winning the AGI race?' You might be able to think of some things.

In a nearby universe, I would be fundraising for a startup to do exactly that, it sounds like a hell of fun problem. :-)  And I'm sure you're right... I just wouldn't expect to get to "capable of 99% of all remote work" within four years.

I realize you’re not explicitly labeling this as a prediction, but… isn’t this precisely the sort of thought process to which Hofstadter's Law applies?

Indeed. Like I said, my timelines are based on a portfolio of different models/worlds; the very short-timelines models/worlds are basically like "look we basically already have the ingredients, we just need to assemble them, here is how to do it..." and the planning fallacy / hofstadter's law 100% applies to this. The 5-year-and-beyond worlds are not like that; they are ... looking at lines on graphs and then extrapolating them ...

So my timelines do indeed take into account Hofstadter's Law. If I wasn't accounting for it already, my median would be lower than 2027. However, I am open to the criticism that maybe I am not accounting for it enough.

To be clear, I'm only attempting to argue about the short-timeline worlds. I agree that Hofstadter's Law doesn't apply to curve extrapolation. (My intuition for 5-year-and-beyond worlds is more like Ege's, but I have nothing coherent to add to the discussion on that front.) And so, yes, I think my position boils down to "I believe that, in your short-timeline worlds, you are not accounting for Hofstadter's Law enough".

As you proposed, I think the interesting place to go from here would be some predictions. I'll noodle on this, and I'd be very interested to hear any thoughts you have – milestones along the path you envision in your default model of what rapid progress looks like; or at least, whatever implications thereof you feel comfortable talking about.

This post taught me a lot about different ways of thinking about timelines, thanks to everyone involved!

I’d like to offer some arguments that, contra Daniel’s view, AI systems are highly unlikely to be able to replace 99% of current fully remote jobs anytime in the next 4 years. As a sample task, I’ll reference software engineering projects that take a reasonably skilled human practitioner one week to complete. I imagine that, for AIs to be ready for 99% of current fully remote jobs, they would need to be able to accomplish such a task. (That specific category might be less than 1% of all remote jobs, but I imagine that the class of remote jobs requiring at least this level of cognitive ability is more than 1%.)

Rather than referencing scaling laws, my arguments stem from analysis of two specific mechanisms which I believe are missing from current LLMs:

1. Long-term memory. LLMs of course have no native mechanism for retaining new information beyond the scope of their token buffer. I don’t think it is possible to carry out a complex extended task, such as a week-long software engineering project, without long-term memory to manage the task, keep track of intermediate thoughts regarding design approaches, etc.
2. Iterative / exploratory work processes. The LLM training process focuses on producing final work output in a single pass, with no planning process, design exploration, intermediate drafts, revisions, etc. I don’t think it is possible to accomplish a week-long software engineering task in a single pass; at least, not without very strongly superhuman capabilities (unlikely to be reached in just four years).

Of course there are workarounds for each of these issues, such as RAG for long-term memory, and multi-prompt approaches (chain-of-thought, tree-of-thought, AutoGPT, etc.) for exploratory work processes. But I see no reason to believe that they will work sufficiently well to tackle a week-long project. Briefly, my intuitive argument is that these are old school, rigid, GOFAI, Software 1.0 sorts of approaches, the sort of thing that tends to not work out very well in messy real-world situations. Many people have observed that even in the era of GPT-4, there is a conspicuous lack of LLMs accomplishing any really meaty creative work; I think these missing capabilities lie at the heart of the problem.

Nor do I see how we could expect another round or two of scaling to introduce the missing capabilities. The core problem is that we don’t have massive amounts of training data for managing long-term memory or carrying out exploratory work processes. Generating such data at the necessary scale, if it’s even possible, seems much harder than what we’ve been doing up to this point to marshall training data for LLMs.

The upshot is that I think that we have been seeing the rapid increase in capabilities of generative AI, failing to notice that this progress is confined to a particular subclass of tasks – namely, tasks which can pretty much be accomplished using System 1 alone – and collectively fooling ourselves into thinking that the trend of increasing capabilities is going to quickly roll through the remainder of human capabilities. In other words, I believe the assertion that the recent rate of progress will continue up through AGI is based on an overgeneralization. For an extended version of this claim, see a post I wrote a few months ago: The AI Progress Paradox. I've also written at greater length about the issues of Long-term memory and Exploratory work processes.

In the remainder of this comment, I’m going to comment what I believe are some weak points in the argument for short timelines (as presented in the original post).

[Daniel] It seems to me that GPT-4 is already pretty good at coding, and a big part of accelerating AI R&D seems very much in reach -- like, it doesn't seem to me like there is a 10-year, 4-OOM-training-FLOP gap between GPT4 and a system which is basically a remote-working OpenAI engineer that thinks at 10x serial speed.

Coding, in the sense that GPT4 can do it, is nowhere near the top of the hierarchy of skills involved in serious software engineering. And so I believe this is a bit like saying that, because a certain robot is already pretty decent at chiseling, it will soon be able to produce works of art at the same level as any human sculptor. 

[Ajeya] I don't know, 4 OOM is less than two GPTs, so we're talking less than GPT-6. Given how consistently I've been wrong about how well "impressive capabilities in the lab" will translate to "high economic value" since 2020, this seems roughly right to me?

[Daniel] I disagree with this update -- I think the update should be "it takes a lot of schlep and time for the kinks to be worked out and for products to find market fit" rather than "the systems aren't actually capable of this." Like, I bet if AI progress stopped now, but people continued to make apps and widgets using fine-tunes of various GPTs, there would be OOMs more economic value being produced by AI in 2030 than today.

If the delay in real-world economic value were due to “schlep”, shouldn’t we already see one-off demonstrations of LLMs performing economically-valuable-caliber tasks in the lab? For instance, regarding software engineering, maybe it takes a long time to create a packaged product that can be deployed in the field, absorb the context of a legacy codebase, etc. and perform useful high-level work. But if that’s the only problem, shouldn’t there already be at least one demonstration of an LLM doing some meaty software engineering project in a friendly lab environment somewhere?

More generally, how do we define “schlep” such that the need for schlep explains the lack of visible accomplishments today, but also allows for AI systems be able to replace 99% of remote jobs within just four years?

[Daniel] And so I think that the AI labs will be using AI remote engineers much sooner than the general economy will be. (Part of my view here is that around the time it is capable of being a remote engineer, the process of working out the kinks / pushing through schlep will itself be largely automatable.)

What is your definition of “schlep”? I’d assumed it referred to the innumerable details of figuring out how to adapt and integrate a raw LLM into a finished product which can handle all of the messy requirements of real-world use cases – the “last mile” of unspoken requirements and funky edge cases. Shouldn’t we expect such things to be rather difficult to automate? Or do you mean something else by “schlep”?

[Daniel] …when I say 2027 as my median, that's kinda because I can actually quite easily see it happening in 2025, but things take longer than I expect, so I double it.

Can you see LLMs acquiring long-term memory and an expert-level, nuanced ability to carry out extended exploratory processes by 2025? If yes, how do you see that coming about? If no, does that cause you to update at all?

[Daniel] I take it that in this scenario, despite getting IMO gold etc. the systems of 2030 are not able to do the work of today's OAI engineer? Just clarifying. Can you say more about what goes wrong when you try to use them in such a role?

Anecdote: I got IMO silver (granted, not gold) twice, in my junior and senior years of high school. At that point I had already been programming for close to ten years, and spent considerably more time coding than I spent studying math, but I would not have been much of an asset to an engineering team. I had no concept of how to plan a project, organize a codebase, design maintainable code, strategize a debugging session, evaluate tradeoffs, see between the lines of a poorly written requirements document, etc. Ege described it pretty well:

I think when you try to use the systems in practical situations; they might lose coherence over long chains of thought, or be unable to effectively debug non-performant complex code, or not be able to have as good intuitions about which research directions would be promising, et cetera.

This probably underestimates the degree to which IMO-silver-winning me would have struggled. For instance, I remember really struggling to debug binary tree rotation (a fairly simple bit of data-structure-and-algorithm work) for a college class, almost 2.5 years after my first silver.

[Ajeya] I think by the time systems are transformative enough to massively accelerate AI R&D, they will still not be that close to savannah-to-boardroom level transfer, but it will be fine because they will be trained on exactly what we wanted them to do for us.

This assumes we’re able to train them on exactly what we want them to do. It’s not obvious to me how we would train a model to do, for example, high-level software engineering? (In any case, I suspect that this is not far off from being AGI-complete; I would suspect the same of high-level work in most fields; see again my earlier-linked post on the skills involved in engineering.)

[Daniel] …here's a scenario I think it would be productive to discuss:

(1) Q1 2024: A bigger, better model than GPT-4 is released by some lab. It's multimodal; it can take a screenshot as input and output not just tokens but keystrokes and mouseclicks and images. Just like with GPT-4 vs. GPT-3.5 vs. GPT-3, it turns out to have new emergent capabilities. Everything GPT-4 can do, it can do better, but there are also some qualitatively new things that it can do (though not super reliably) that GPT-4 couldn't do.

…

(6) Q3 2026 Superintelligent AGI happens, by whatever definition is your favorite. And you see it with your own eyes.

I realize you’re not explicitly labeling this as a prediction, but… isn’t this precisely the sort of thought process to which Hofstadter's Law applies?

Thanks for the thoughtful and detailed comments! I'll respond to a few points, otherwise in general I'm just nodding in agreement.

I think it's important to emphasize (a) that Davidson's model is mostly about pre-AGI takeoff (20% automation to 100%) rather than post-AGI takeoff (100% to superintelligence) but it strongly suggests that the latter will be very fast (relative to what most people naively expect) on the order of weeks probably and very likely less than a year.

And it's a good model, so we need to take this seriously. My only quibble would be to raise again the possibility (only a possibility!) that progress becomes more difficult around the point where we reach AGI, because that is the point where we'd be outgrowing human training data. I haven't tried to play with the model and see whether that would significantly affect the post-AGI takeoff timeline.

(Oh, and now that I think about it more, I'd guess that Davidson's model significantly underestimates the speed of post-AGI takeoff, because it might just treat anything above AGI as merely 100% automation, whereas actually there are different degrees of 100% automation corresponding to different levels of quality intelligence; 100% automation by ASI will be significantly more research-oomph than 100% automation by AGI. But I'd need to reread the model to decide whether this is true or not. You've read it recently, what do you think?)

I want to say that he models this by equating the contribution of one ASI to more than one AGI, i.e. treating additional intelligence as equivalent to a speed boost. But I could be mis-remembering, and I certainly don't remember how he translates intelligence into speed. If it's just that each post-AGI factor of two in algorithm / silicon improvements is modeled as yielding twice as many AGIs per dollar, then I'd agree that might be an underestimate (because one IQ 300 AI might be worth a very large number of IQ 150 AIs, or whatever).

And (b) Davidson's model says that while there is significant uncertainty over how fast takeoff will be if it happens in the 30's or beyond, if it happens in the 20's -- i.e. if AGI is achieved in the 20's -- then it's pretty much gotta be pretty fast. Again this can be seen by playing around with the widget on takeoffspeeds.com.

Yeah, even without consulting any models, I would expect that any scenario where we achieve AGI in the 20s is a very scary scenario for many reasons.

--I work at OpenAI and I see how the sausage gets made. Already things like Copilot and ChatGPT are (barely, but noticeably) accelerating AI R&D. I can see a clear path to automating more and more parts of the research process, and my estimate is that going 10x faster is something like a lower bound on what would happen if we had AGI (e.g. if AutoGPT worked well enough that we could basically use it as a virtual engineer + scientist) and my central estimate would be "it's probably about 10x when we first reach AGI, but then it quickly becomes 100x, 1000x, etc. as qualitative improvements kick in." There's a related issue of how much 'room to grow' is there, i.e. how much low-hanging fruit is there to pick that would improve our algorithms, supposing we started from something like "It's AutoGPT but good, as good as an OAI employee." My answer is "Several OOMs at least." So my nose-to-the-ground impression is if anything more bullish/fast-takeoff-y than Davidson's model predicts.

What is your feeling regarding the importance of other inputs, i.e. training data and compute?

> I think of AI progress as being driven by a mix of cognitive input, training data, training FLOPs, and inference FLOPs. Davidson models the impact of cognitive input and inference FLOPs, but I didn't see training data or training FLOPs taken into account. ("Doesn’t model data/environment inputs to AI development.") My expectation that as RSI drives an increase in cognitive input, training data and training FLOPs will be a drag on progress. (Training FLOPs will be increasing, but not as quickly as cognitive inputs.)

Training FLOPs is literally the most important and prominent variable in the model, it's the "AGI training requirements" variable. I agree that possible data bottlenecks are ignored; if it turns out that data is the bottleneck, timelines to AGI will be longer (and possibly takeoff slower? Depends on how the data problem eventually gets solved; takeoff could be faster in some scenarios...) Personally I don't think the data bottleneck will slow us down much, but I could be wrong.

Ugh! This was a big miss on my part, thank you for calling it out. I skimmed too rapidly through the introduction. I saw references to biological anchors and I think I assumed that meant the model was starting from an estimate of FLOPS performed by the brain (i.e. during "inference") and projecting when the combination of more-efficient algorithms and larger FLOPS budgets (due to more $$$ plus better hardware) would cross that threshold. But on re-read, of course you are correct and the model does focus on training FLOPS.