Leopold Aschenbrenner and the Case for AGI by 2027

Learning what a forecast depends on

At Germany’s national Jugend forscht competition in 2016, Hentges and Aschenbrenner took fourth prize in earth and space sciences. Their system used machine learning to turn readings from homemade sensors into local forecasts of particle concentrations for the following days. In their published account, the two students described both the sensors and the limits of the existing monitoring network. The competition record credits them jointly.

At Columbia, his attention turned to growth. His undergraduate thesis asked whether an aversion to change could bring exponential economic growth to an end. The economics department lists him as a 2021 Romine Prize recipient. He graduated as valedictorian at nineteen, as he recounts on his own website. He worked on long-run growth at Oxford’s Global Priorities Institute and later co-authored a working paper with Philip Trammell. In their model, faster technological progress can raise danger in the near term yet reduce lifetime exposure by shortening the time spent at a dangerous stage. That result depends in part on whether a wealthier society actually devotes more resources to safety.

What a weaker teacher can teach

At OpenAI he joined the Superalignment team. Its problem was easy to state: how can a person supervise a system whose best answers they cannot reliably judge? In December 2023 the team published a study with twelve co-authors. It identifies Aschenbrenner as one of its primary authors. The researchers gave a stronger model training labels supplied by a weaker one. On some language tasks the student learned to answer questions its teacher got wrong. Results were much less encouraging on the task of learning to score chatbot responses. The experiment gave the team a way to study the problem under controlled conditions. It also showed how much the answer depended on the task.

OpenAI dismissed him in April 2024. In his June interview with Dwarkesh Patel, Aschenbrenner said the company had described his sharing of a preparedness document with three outside researchers as a leak. He said he had sought feedback and believed a separate memo about laboratory security also contributed to his dismissal. That is his account of the dispute. The interview does not independently establish the company’s motive.

What he meant by 2027

In June 2024 he released Situational Awareness: The Decade Ahead. The 165-page edition is dated 6 June. He wrote in a personal capacity. His starting point was a specific occupation he knew from the inside. Could a system do the whole job of an AI researcher or engineer, including working independently on a problem for weeks? In the first chapter he calls that prospect by 2027 “strikingly plausible”. Elsewhere he writes more confidently about the course the industry is on. He also allows for a much longer wait if the next expansion fails to deliver.

His argument combines three changes. A laboratory can devote more computing power to training. A better algorithm can obtain the same result with less computation. A model can also be given time, tools and memory to carry a task further. He calls that third change unhobbling. The distinction matters. A cheaper way to reach a known score is measurable. How much scientific judgement a model gains from being able to work for a week is a different question.

Then comes the feedback loop. If one system can do frontier research, a laboratory might run millions of copies and use their discoveries to improve the next generation. In his scenario a decade’s worth of algorithmic advances could fit into a year or less. It depends on experiments that can run in parallel, enough computing resources to run them, and useful findings that survive replication. Ten thousand researchers waiting for the same experiment still have to wait for its result.

Taking the forecast seriously, two years on

One unusually informative test comes from inside a laboratory. On 5 November 2026, METR published its assessment of Anthropic. It estimated that progress on an internal capability index had accelerated by a factor of 1.9 since October 2025, with a 95% confidence interval of 1.5 to 2.8. Its investigators judged that AI probably explained much of the recent acceleration. That attribution was substantially more uncertain than the index measurement. Some evidence was redacted, and people still performed essential parts of model development. The finding supports one link in Aschenbrenner’s argument: AI use can help a laboratory produce better AI more quickly. It leaves open the larger question of a laboratory’s research being carried out autonomously.

His electricity forecast admits a different test. Around 2030 his illustrative training cluster costs at least one trillion US dollars and draws about 100 gigawatts. Run continuously, that is 876 terawatt-hours per year, a little over one fifth of the US generation figure he used in 2024. Berkeley Lab’s June 2026 US data-centre report puts the entire domestic sector at 649 terawatt-hours in its 2030 reference case, with a range of 521 to 843. One hypothetical cluster at full load would exceed even the upper end of that American range. Its location and average utilisation therefore matter enormously. Both are projections. Neither figure is an observed electricity bill.

Our judgement as of 15 November 2026 is that the case for a feedback loop has gained empirical support, while the evidence examined here does not establish the 2027 milestone. A convincing test would give a system a novel research question, then document who chose the experiments, recognised a failed idea and decided what to try next. Independent researchers would need to reproduce the result and know the human help, time and computation it took. A system that wrote every line of code could still have left the decisive scientific choices to a person.

From laboratory security to public policy and investment

Aschenbrenner also made a political argument. He wanted the United States and its allies to protect frontier research against state espionage, especially by China, and expected a government-led AGI effort around 2027–28. His chapter on control raises a separate problem: keeping a model out of an adversary’s hands says little about whether its own operators can supervise it. His argument reached Washington in a traceable way. In April 2026 Yusuf Mahmood cited the essay in written testimony before a House select committee. The Genesis Mission also gave the US government a role in AI-assisted scientific discovery. Neither the testimony nor the programme establishes the creation of the specific AGI project he predicted.

His investment plans were already public on 4 June 2024, when he discussed them with Patel at the launch of the essays. A September regulatory filing bears his signature for Situational Awareness Partners LP. The venture exposed capital to his view of the world, including the demand for computing and power. In July 2026 Reuters reported from a letter to investors that the portfolio had fallen 67% that month and reported the sale of the bulk of its stock portfolio to Citadel. That episode records the risks of the positions the fund held. It cannot tell us when an autonomous researcher will exist.

A chronology

1. 2016–17 · BerlinCompetes with Max Hentges in Jugend forscht and co-authors their account of particulate pollution monitoring.
2. 2021 · ColumbiaGraduates and receives a Romine Prize for his economics thesis.
3. 2023 · OpenAICo-authors the Superalignment team’s weak-to-strong generalisation study.
4. April–June 2024Leaves OpenAI in April. Releases the essays and discusses his investment plans in June.
5. September 2024Signs the fund’s US regulatory notice.
6. 2026The essay is cited in congressional testimony in April. The fund’s public equity portfolio undergoes a major sale in July.