Leopold Aschenbrenner: Situational Awareness Two Years On - philippdubach.comSkip to main contentphilippdubach<br>Quantitative finance, AI, and the economics underneath.
Eight of Aschenbrenner’s technology and infrastructure predictions have confirmed by May 2026 (test-time compute, GPQA saturation, power as binding constraint, Marcellus gas-for-AI, Gulf chip pivot, AMD’s $1T compute TAM); eight political-economy predictions have falsified.<br>The modal-story method (a vivid dated narrative bet rather than a probability distribution) is durable on empirically-lawful substrates like scaling curves and capex aggregates, brittle on coalition-political substrates like elections and executive turnover.<br>Libertarian, Burkean hawk, and alignment optimist run as a series circuit where each premises the next; mixed evidence on a multiplicative chain is closer to disconfirmation than the headline scorecard reads.<br>Situational Awareness LP put the framework in capital: NVIDIA and gas-turbine trades paid, but the ‘big bond short’ on real interest rates above 10% has not fired and is the cleanest open test of the syllogism’s last link.
×In June 2024, Leopold Aschenbrenner published a 165-page essay called Situational Awareness and committed himself to a chain of dated forecasts. Test-time compute would be the next paradigm. Power, not chips, would be the binding constraint on the US AI buildout. The Marcellus shale would end up powering data centres, not the hyperscalers’ climate pledges. By May 2026 those calls have landed. Eight of his concrete technology and infrastructure predictions have confirmed cleanly.<br>Eight others have falsified: voluntary lab merger, Congressional trillions, a coalition of democracies, tightening export controls.<br>This split tells us something about the worldview holding both halves up. Read Situational Awareness and the four-and-a-half-hour Dwarkesh Patel interview side by side and you find something stranger than a clean libertarian, a clean hawk, or a clean alignment researcher. You find all three running at once.<br>The libertarian:<br>“I am a big believer in the American private sector, and would almost never advocate for heavy government involvement in technology or industry.”
That sentence is in Situational Awareness. Three pages later he calls for the largest peacetime industrial nationalisation in American history.<br>The hawk: he wants the US to lock down its AI labs to “B-21 bomber-grade” security, threaten nuclear retaliation against strikes on its data centres, and absorb the frontier labs into a state project by 2027/28. He calls this Burkean.<br>“American checks and balances have held for over 200 years and through crazy technological revolutions.”
The alignment researcher: he was an initial member of OpenAI’s Superalignment team. He has publicly stated, including in the Dwarkesh Patel interview, that he was dismissed in spring 2024 after sharing a security memo with the board. He has also publicly stated that he declined the company’s non-disparagement NDA and forfeited roughly a million dollars in equity as a result. “Freedom is priceless,” was his summary. He thinks the default plan (scalable oversight, weak-to-strong generalisation, interpretability) will probably work. By the standards of MIRI and Yudkowsky, he is an optimist.<br>Each of those identities, taken alone, is internally consistent. The conjunction is not. A libertarian does not nationalise. A Burkean does not impose a Manhattan Project on a peacetime economy. An empirical alignment optimist does not insist that the only viable governance structure is wartime SCIF custody by April 2028.<br>Aschenbrenner gets to all three because he runs the same forecasting machinery on three different domains (compute scaling, great-power competition, alignment tractability) and the forecasts converge on a single decisive event. The opening line of his Dwarkesh interview names the convergence directly:<br>“What will be at stake will not just be cool products, but whether liberal democracy survives, whether the CCP survives, what the world order for the next century will be.”
This essay reads Aschenbrenner the way he reads scaling laws: by decomposition. The three identities are the spine. For each, we look at the framework, today’s evidence, the strongest counter-position, and what the synthesis costs him. The point is to show what a convergent forecast does to a worldview, and what that reveals about the AI-policy debate the rest of us are still having.<br>I. The forecaster before the politics<br>Before any criticism, what he got right.<br>The signature analytical move in Situational Awareness is what he calls “counting OOMs”, orders of magnitude of effective compute. The decomposition is multiplicative: half an OOM per year of training compute, half an OOM per year of algorithmic efficiency, plus discrete unhobbling gains (RLHF,...