"Holy shit" graphs and how to get more of them

We are likely in a situation where the most powerful models in the world are internal models, and, here's a key distinction, potentially even models that are never intended to be released to the public (i.e. a model designed to provide internal uplift on capabilities during training but that is not intended to be / not particularly suitable for commercial purposes.)

Internal-external information and capability gaps can distort both public and regulatory perception of frontier capabilities and dangers. If the internal frontier is three or six months ahead, such a gap is worrisome if the labs are close to automating various bottlenecks to increased capabilities development, like automating AI R&D.

Some ways of mitigating these visibility gaps and communicating the existing data we do have are implementable over the short term and confer minimal cost for many actors relative to the advantages of visibility for policymakers.

What is a "holy shit" graph?

Many of my intuitions about the coming impacts of AI are shaped by "Holy shit" graphs. The METR time horizon graph is one such graph, but has been discussed to death, so let's talk about another three:

Internal spending by OpenAI researchers (at API prices)

This spending graph is so bananas that when I try to google the source, The Gemini overview sometimes gets "AI token spending" and "Internal AI token spending by OpenAI researchers" confused and asserts that the thing I'm asking about could not possibly be true.

Epoch concludes:

Such rapid growth is probably unsustainable: if the recent growth in our preferred models continued for another year, annualized run rates would reach about $185 million for the median researcher and $2.3 billion for the 90th percentile researcher.

What the fuck.

General AI (as opposed to narrow AI) model revenues

This graph shows the growth in frontier AI company revenues over time. It's unclear from the data labeling but this number is growing at 3-4x per year for the industry. To get a sense for the growth estimates, we can look at a prettier but more out of date graph from Epoch instead.

There seem to be some industry leader effects; Anthropic was growing faster when it was the second-highest company in revenue than now, and do note that this number may currently be slowing, and is itself an endogenous variable downstream of policy. Nevertheless, the industry as a whole is growing at a rate that would make it larger than the 2026 world economy by 2031-33. The point is not that this will continue forever, it's that if it doesn't stop continuing soon, the world must start looking very strange by the early 2030s, if not earlier.

Extent of AI R&D automation

It's hard to see from this graph, but it seems like it is currently taking about 15 months from a category to go from low single digits to fully saturated, and about seven months for the category to reach roughly the same level that the graph was at. This would seem to predict, again absent regulation, ~8 months from the end of the graph (6 months from now) to AL4-5 hitting ~99% and ~22 months from the end of the graph (20 months from now) for the AL5 category to hit 100%. Indeed, Anthropic co-founder Jack Clark estimated back in the halcyon days of May 2026, a 60% chance that "an AI system powerful enough that it could plausibly autonomously build its own successor – happens by the end of 2028."

We can and likely should discount this from its source, but discounting it and the many related arguments all the way to "nope, not going to happen" seems hopeful.

I think communicating these "Holy shit" graphs, insofar as they are accurate, and we should have ways of verifying their accuracy externally, might be one of the most impactful pieces of communication the labs can give policymakers and the public.

Indeed, the accuracy of the first graph is weird. If we look back at the first few months of the graph, it would seem to imply that even though Codex launched in early 2025, it took until well into 2026 for Codex use to become widespread at the company. This is reasonable-ish speed of diffusion? But certainly non-obvious.

In any case, I'm not going to justify these claims here, but Anthropic reaching 100% AL5, and indeed even "only" substantial fractions of model R&D work being at AL5, seems like a four to five-alarm fire loss of control risk. This graph is a good starting point in evaluating the case that we are a couple of years away from full AI R&D on the default path, instead of 10 or more.

So it seems robust under many worldviews that independent evaluators and regulators being able to audit this information more comprehensively would be good thing. The graphs ought to be trustable, and indeed we should be cautious of those that are impossible to audit.

So what makes a "Holy shit" graph, in this definition? Three things:

0th: Your graph encodes a model of reality that is reasonably accurate, to your best understanding. Your graph being honest is not a pre-requisite for it going viral, but I will be very sad if you take this post as encouragement to make irresponsibly misleading graphs :(

1. Sustained trend over a non-trivial number of data points, and the trend doesn't yet seem to be budging.
2. Extrapolating out the trend naively gives crazy results.
3. So there must be a trend break, surely, but the impacts of the trend break are likely to be severe as well.

I think the labs are sitting on a bunch of these data sources, and it would be good if we had more of them. Here's a few suggestions:

- Increased training metric transparency: Regularly updated series showing number of current training runs above some FLOP threshold (potentially , in concordance with EU AI Act), FLOP count of largest active training run, FLOP count of largest ever training run.

- Increased deployment metric transparency: Time since end of training, total revenue breakdown for each model

- Agent metrics: Current count, 99.9th percentile length session (actual max might fail to preserve privacy, but might still be an important indicator of a rogue agent), number of websites

The high-level point is having a set of measurements which we can reference and which serve as high-level primary-source indicators of the 0th, 1st and 2nd derivatives of capabilities growth.

But insight into speed of diffusion also matters, I would argue for both understanding the mechanics of diffusion and for study of gradual disempowerment. Much has been written about gradual disempowerment , and indeed I think it's a policy area that is as important or more to understand as even the x-risk landscape. But much less has been written about how to operationalize it, though the GD paper authors do cite this as important follow-up work. Part of the reason that less work has gone into these metrics is that existing labor automation tools capture it meaningfully, but there are facets of the problem that go beyond economic and labor impacts. I'd love to see work done to measure every all three sections of the paper. Here are some thoughts:

Labor, relational and political

- Amount of time spent on claude.ai or chatgpt.com, averaged across every American or across the whole globe.

- Amount of time spent on claude.ai or chatgpt.com, averaged across every American.

- Measurements of aggregate token output, and some translation of said token output to a "human-equivalent hours worked" concept.

- Percentage of Americans that have most of their political content or discussion come from a chat application or agent.

This could be you!

Are you an individual researcher, research organization or group working on alignment? Maybe you could make some "Holy shit" graphs with datasets that are already public or could be made public if requested?

If you're a grantmaker or manager, area worth thinking about?

I think METR and Epoch are, in important ways, highly cited in part because they make "Holy shit" graphs. It also seems important to me that these graphs have distinctive visual style and branding, and that they are either hand-crafted by humans or that their AI contributions are not highly visible; if the same graph was in some sense "equally good" but designed by the visual style of even Opus 5.5, many would disengage from the graph.

This does not mean that you should pivot away from the wonderful work that you are currently doing and go make "holy shit" graphs. I do think the number of useful "Holy shit" graphs is relatively small, maybe 20 or 50? And many of the best ones require a lab to publish the relevant data product. But the more general point I'm trying to gesture at is that investing in communicating your work legibly to your audience of choice is highly cost effective. Maybe someone on your team should be making "Holy shit" graphs, or identifying future "Holy shit" graphs you could make, or in general investing in communicating your research memorably, legibly, and in a way that gets your organization's logo around the internet a few hundreds of thousands of times.

Note that though I think this work is on average net positive, I am arguing it is also a force multiplier for your work. If many people currently think your work is net negative, consider engaging effortfully with those arguments before making the cool graph.

1. The METR time horizons graph has come under intense scrutiny recently. This piece lays out a few reasons for why to distrust the graph. From most to least severe:
   
   1. The experimental design paid the benchmarked study participants per hour, which can incentivize working longer than you would if you were paid per completion of task.
   2. The sample size for the experiments was vanishingly small, with n=4 for some data points.
   3. The METR task suite systematically overestimated task length, as the participants took a surprisingly long time getting oriented on task basics.
      
   The first point is a fairly strong criticism, and the second one should make us worry about the measurement error of the task suite, but the third point I would argue has the opposite sign. If we add a fixed time to each task or more realistically a time penalty that is decreasing in task length share, shorter tasks would get overestimated more than longer tasks are underestimated. This should make your estimate of the level biased upward and your estimate of the slope on a log plot biased downward.
   
   But in any case, if we're worried about the external validity of the experiment, METR has designed other graphs with comparable results. Overall I agree that METR's time horizons graph is not literally, actually measuring "the thing we care about", but it's a close-enough correlate to the thing we care about that it's worth paying attention to. Imagine if we'd had a (non-saturated) mathematics research time horizon graph. I think it would have had enough internal validity to give slight advance warning of OpenAI's recent results.
   
   Similar, and imo equally non-fatal methodological critiques apply to the other graphs discussed here.
2. There's a more updated graph, but I think that one's paywalled :( I think it would be very socially positive if this specific graph were not paywalled because the deduplication effects are neat and useful and address a central skepticism of frontier AI revenues. But even if this graph remains paywalled, the version made from lab internals need not be.
3. I did a more detailed job of that here. Other people have done much more detailed jobs elsewhere.
4. And any other lab, to be clear.
5. 1x a week, targeting <1 month reporting lag.
6. I'm particularly fond of these because they dovetail nicely with the AIFP's additional models for estimating the date of "automated coder."
7. I know that individual instances have hard timeouts, but my understanding is, especially with API pricing, it is possible to run a scheduler on an agent setup that more or less allows sequences of instances to act like one single autonomous agent.
8. Not enough! GD is not just an outcome that we get if we avoid losing control quickly, it's also a potential precursor to losing control quickly, or to locking in trajectories where we lose control quickly.
9. When I say Americans here, I'm saying so because the frontier labs have a disproportionate number of Americans in their customer base and because frontier labs are meaningfully based in the US, but this data would also be great to have for other regions, of course.
10. Not to mention, I think the models are not quite at human expert level on visual design.
11. I also have the impression that many alignment organizations are currently smaller than would be cost-effective, and am unsure that cost-effectiveness is even the right north star to be aiming at in this situation, but as with anything happy to discuss these things / change my mind.