"Situational Awareness," scored two years later

The essay that named the decade

In June 2024 Leopold Aschenbrenner published a 165-page essay arguing that AGI was plausible by 2027, that "by 2027 AI systems will have the capacity to conduct their own AI research," that this would compress a decade of progress into a year, and that the resulting race between the United States and China would be the central fact of the decade. It gave the discourse a vocabulary it still uses. Counting the OOMs. The trillion-dollar cluster. Unhobbling. The Project.

We disliked the essay when it came out, for reasons of tone more than content, and we have been trying to correct for that while grading it. What follows leans on two scorecards written this year, Jamie Harris's on the EA Forum in March and Philipp Dubach's in May, plus what has been in the financial press since late July. We are reporting their verdicts where we cannot check a claim ourselves.

Where he was right

The physical predictions did best. Harris's scorecard finds that "infrastructure investment and algorithmic efficiency are tracking ahead of his predictions." Dubach calls the argument that power rather than chips would be the binding constraint on the American buildout the essay's "most prescient line of policy analysis," and points to heavy-frame turbine lead times reaching six years. The specific and unglamorous prediction that the clusters would run on behind-the-meter natural gas, climate pledges notwithstanding, also landed.

The test-time compute argument was the best technical call. Aschenbrenner asked what would happen if a model "could use millions of tokens to think about and work on really hard problems." OpenAI shipped o1 four months later and DeepSeek-R1 followed in January 2025. He also said GPQA Diamond would "fall in the next generation or two," and Dubach records it at 88.4 percent for GPT-5 by May 2026, roughly on the timeline the essay implied. On capabilities generally, Harris grades the benchmark performance as broadly met, while noting that the "shocking leap" the essay described did not arrive in the form described.

The security warnings also aged well. Harris marks the safety and security section as validated, because the essay's claim that labs were operating with inadequate safeguards under race pressure proved, in his word, prescient.

Where he was wrong

The economics missed. The essay expected AI revenue to hit a 100 billion dollar run rate by mid-2026. Harris reports the best figure as 60 billion, and is generous about it, since "off by less than a factor of 2" on a two-year revenue forecast is not bad. But it matters which direction the miss went, because the whole argument for the trillion-dollar cluster was that revenue would justify it.

The political predictions missed harder. Dubach summarizes the nationalisation and coalition forecasts as having "mostly remained unfulfilled or moved in opposite direction." There has been no voluntary merger of labs into a government project, no trillions from Congress, no democratic coalition of the kind the essay imagined. Harris grades the government response as behind schedule while noting the essay's own deadline for The Project has not technically elapsed. That is a fair reading and also the kind of thing that keeps a prediction alive past its useful life.

And the China chapter was half right in an instructive way. Harris finds that the narrow predictions about 7nm chips, power, and espionage held, while the essay missed both China's independent innovation and the resilience of open-weight models. The essay predicted open source would fade as proprietary algorithms became decisive. The scorecard at agiscorecard.com marks that one simply "wrong." Most of the models we use day to day at a non-profit are open weights from labs the essay expected to be locked out.

The fund

The reason this essay is in the news again in August 2026 has nothing to do with its predictions. After publishing, Aschenbrenner set up a hedge fund, Situational Awareness LP, backed by Patrick and John Collison, Daniel Gross, and Nat Friedman. According to an investor letter reviewed by the Financial Times, it returned 439 percent net in the first half of 2026, and reached roughly 45 billion dollars in assets by July. Then AI infrastructure stocks fell, and the fund, which had used borrowing to amplify returns, faced successive margin calls. It sold its public positions to Citadel and was left with its private holdings and, by press accounts, around 35 billion dollars in losses inside a month.

We raise this because it is the cleanest test the essay will ever get. The thesis was that anyone with situational awareness could see the trajectory and act on it. The author acted on it with real money, was right about the direction of the industry for two years, and was still nearly wiped out by leverage and timing. CNBC reported leverage of up to 400 percent. A correct macro forecast did not protect against the ordinary risk of being that levered into a drawdown. That is a lesson about forecasts as much as about finance.

What the essay did to the way people talk

The essay's lasting effect is on vocabulary rather than on any particular prediction. Before June 2024, labs talked about model releases. After it, they talked about gigawatts, and governments did too. The habit of describing progress as orders of magnitude of effective compute, with algorithmic gains and unhobbling folded in as extra OOMs, is now how frontier roadmaps get written and how export controls get justified. Whether or not the framing is right, it is the water everyone swims in.

From the vantage point of a small foundation, that framing has a cost. It makes compute the unit of seriousness, which quietly implies that anyone without a cluster has nothing to contribute. The two years since suggest otherwise. The open-weight ecosystem the essay wrote off is where most of the checkable science has happened, and the algorithmic efficiency gains it correctly predicted are precisely the things that a group without a trillion dollars can study.

If we were grading the essay as a forecast we would give it a pass on physics and a fail on politics. If we were grading it as a piece of persuasion we would give it full marks, and we would want the next person who writes something this influential to publish their predictions in a form that makes a scorecard easy to build. Harris and Dubach had to do that work by hand. It should not have been necessary.