Straight lines on graphs

People who do not watch AI developments closely are often suspicious of an intuition that I’ll call “straight lines on graphs,” which holds that progress in AI is (not only rapid but) remarkably regular over time.¹

In my experience, most people (including today’s straight-liners) start out suspicious of “straight lines on graphs,” but eventually have the intuition beaten into them. It doesn’t come from rational debate; they don’t wake up one day and have a change of heart about the viability of extrapolating across orders of magnitude. Instead they make implicit guesses of how the future might go, then notice that “straight lines on graphs” predict the future better than their guesses.²

Unfortunately I don’t think I can easily convince you of the intuition via a blog post. But hopefully I can at least share some mental models that come out of it.

Jagged AI capabilities

Much has been written about the “jagged frontier” of AI capabilities — the thought that the capability profiles of AIs appear very uneven relative to some notion of human capabilities.

I’m not sure I totally buy the frame, but I definitely do agree that AI capabilities are totally remarkable in some domains and far less impressive in others.

Still, one “straight lines on graphs” intuition I have is that, although the relative “level” (i.e. current bar) of AI capabilities differs greatly across domains, those capabilities are all on the same “slope” (i.e. they progress at a similar rate). I don’t literally think this is true, but it’s a helpful starting place.

This is all well-illustrated in this post attempting to measure how AI time horizon varies across domains from my colleagues Thomas Kwa and Vincent Cheng.³

Relative to my expectations beforehand, I would say that the rate of increase over time is strikingly similar across task distributions, even while AIs are much less performant at some types of tasks (here: operating computers via their vision capabilities) at a given point in time.

Similarly, my starting belief is that progress on writing quality, research taste, use of external resources, fluid intelligence, etc. has a similar pace to other AI capabilities trends (doubling every 3-7 months), even if AIs are today in a much less impressive place than they are on, say, competitive programming problems.

Gains from RL

Many people speculate that the apparent kink in METR’s time horizon trend around 2024 is due to reinforcement learning advances triggering a faster pace of AI capabilities progress.

Personally, I’m skeptical. I suspect that some combination of the following play more of a role:

1. The tasks that we measure AIs on which give most signal for time horizon have become increasingly correlated with the tasks which AIs are trained on (driven by similar factors, e.g. need for cheap/automatic verification that make it easier to build eval tasks or RL environments). So although the measured progress is “real,” it is capturing an increasingly narrow segment of the task distribution; capabilities measured on a more representative task distribution would show shallower progress.

2. Tasks in the more difficult part of METR’s task distribution disproportionately favor AIs in some sense (e.g. because they have a scoring function which the AI can call frequently, as is true for RE-Bench).

3. Dumb luck — at ~20 effective datapoints, we lack data to reliably tell trend breaks apart from noisy continuations. (Although of course this becomes less plausible as an explanation as more data comes in.)

Still, I think “straight lines on graphs” sets the terms of the debate correctly.

If I were to lean into the perspective that RL has genuinely caused a step change in the pace of progress (as I do at least think is true for the segment of the task distribution that METR is measuring, per (1) above), I would very informally split progress into two types:

1. Progress from pre-training, which “lifts all boats” across task distributions (primarily because training data is less selected in some sense), and

2. Progress from post-training, which is targeted on specific types of problems (because of the expense of creating RL environments for all types of tasks, combined with imperfect generalization from RL).

General progress then results at some pace (say, 6 month doubling), although major AI companies can direct some additional progress in particular fields of their choosing. (They will choose in roughly descending order of cost-effectiveness, which explains why they have made concentrated efforts in software engineering and AI research, but note that they could redirect this focus later.)

Compute slowdowns

I have a paper with Parker Whitfill and Ben Snodin walking through a simple model of how AI capabilities progress might respond to slowdowns in the growth of (experimental and/or training) compute.⁴

It seems natural to think that if key inputs to AI R&D slow down, outputs will slow down in turn. If you draw out the lines using our model, you might then expect substantial delays in AI capabilities milestones due to compute slowdowns.

There’s much to be said about the ways in which this model might underestimate progress. (In particular: that some fraction of algorithmic progress is not bottlenecked on compute, as well as the possibility that AI labor will provide a dramatic increase in inputs, overpowering the compute slowdown.) In fact nowadays I’m inclined to think that capabilities are so impressive and proceeding so rapidly that remarkable capability milestones will already be hit before compute slowdowns have any time to bite.

But as in earlier cases, I think that “straight lines on graphs” helpfully clarify the terms of discussion.

1. This intuition drives the narrative in AI 2027, the goals/forecasts of major AI companies, and of course the interest in our time horizon work.

2. At least, I think this was true for me.

3. Note that I think of this post as mostly illustrative, rather than directly providing much evidence for “straight lines on graphs.”

4. We assume (perhaps overly aggressively) that all algorithmic progress is ultimately bottlenecked on compute. You don’t need to take that literally; you might instead want to say, for instance, that this is true only for some fraction of algorithmic progress. But either way, your starting belief should probably be that there’s been some contribution to AI capabilities progress from exponentially increasing compute and some from becoming more effective at using compute.