Research: Leopold Aschenbrenner — The Fired Safety Researcher Who Wrote the AGI-Race Thesis, Then Traded It

Companion pages: Watching the Watchers and The EA to AI-Safety Funding Pipeline.

Leopold Aschenbrenner is a living private person who became a public figure through his own essay, interviews and fund. His firing from OpenAI is contested: both accounts are given below, and neither is adopted. Fund figures come from press reports and 13F filings, move fast, and are dated. Any argument about what the fund’s positions say about the essay is ours, built on his own statements and his fund’s public filings; it is not a claim about his motives.

1. Background

He arrived at Columbia from Germany at fifteen. Asked by the College’s own magazine what he was like when he arrived, he answered: “I was just 15, and I was coming from Germany, an ocean away.” He graduated in 2021 as valedictorian, at nineteen (Columbia College Today, Apr 23 2021).

Before OpenAI he worked at the FTX Future Fund, Sam Bankman-Fried’s philanthropic arm, alongside William MacAskill and Avital Balwit; Fortune notes “there is no evidence that anyone who worked at the FTX” fund “knew the money was stolen or did anything illegal.” Balwit, his partner, is chief of staff to Anthropic’s CEO (Fortune, Oct 8 2025).

2. OpenAI and the superalignment team

He joined OpenAI’s Superalignment team, led by Jan Leike and Ilya Sutskever, and is a co-author of its “Weak-to-Strong Generalization” paper (December 2023, ICML 2024), whose lead author, Collin Burns, was later tapped for the US government’s AI evaluation center and removed during onboarding (arXiv:2312.09390; Watching the Watchers). Transformer reported he “was one of very few OpenAI employees who did not sign a letter calling for Sam Altman’s return after the board fired him” (Transformer, Jun 4 2024).

3. The firing, April 2024: two accounts

That he was fired in April 2024 is undisputed. Why is not.

His account (Dwarkesh Podcast, June 4, 2024, reported the same day by Transformer): he had written an internal memo arguing OpenAI’s security was “egregiously insufficient to protect against the theft of model weights or key algorithmic secrets from foreign actors” and, after a security incident, shared it with board members. He says leadership told him “the reason this is a firing and not a warning is because of the security memo,” and: “It was very made explicit that the security memo was a major reason for my being fired.” The trigger OpenAI cited was a “brainstorming document on preparedness, safety, and security measures on the path to AGI” he had shared with three outside researchers; he calls it benign, and notes OpenAI had publicly disclosed similar timelines. He says he was offered vested equity worth close to $1 million on condition of signing exit paperwork with a non-disparagement clause, and refused (Transformer; Dwarkesh Podcast).

OpenAI’s account: a spokesperson said the security concerns he raised “did not lead to his separation,” that OpenAI “disagree[s] with many of the claims he has since made” about its security, and tied the departure to the information-sharing incident (Transformer).

4. Situational Awareness: The Decade Ahead (June 2024)

A self-published essay of about 165 pages (situational-awareness.ai). In his words:

• “AGI by 2027 is strikingly plausible.”
• “We would rapidly go from human-level to vastly superhuman AI systems.”
• “Reliably controlling AI systems much smarter than we are is an unsolved technical problem,” and “failure could easily be catastrophic.”
• “If we’re lucky, we’ll be in an all-out race with the CCP; if we’re unlucky, an all-out war.”
• “Superintelligence will give a decisive economic and military advantage.” A lead of “a year or two or three” could be “as utterly decisive a military advantage as the US coalition had against Iraq in the Gulf War,” and “the advantage conferred by superintelligence would be decisive enough even to preemptively take out an adversary’s nuclear deterrent.”
• “I find it an insane proposition that the US government will let a random SF startup develop superintelligence.”
• On the labs: they “treat security as an afterthought” and are “basically handing the key secrets for AGI to the CCP on a silver platter.”

(Quotes from the essay; compiled with Zvi Mowshowitz’s verbatim extracts.)

5. Situational Awareness LP

After the essay he founded an AI-focused hedge fund with the same name, as founder and chief investment officer. Fortune calls Carl Shulman its “director of research” and names backers Nat Friedman, Daniel Gross, Patrick and John Collison, and Graham Duncan; the strategy was long the AI build-out (semiconductors, power, data centers: Fortune named Intel, Broadcom, Vistra and Core Scientific) with a large options overlay (Fortune, Oct 8 2025).

A 13F “reportable value” counts options at notional, so it is not cash under management. The filer record is SEC CIK 0002045724.

6. The July 2026 unwind

• Up 439 percent in the first half of 2026 and reported at as much as $45 billion at the start of July, the fund’s leveraged long on AI infrastructure and its software shorts moved against it together: SanDisk fell 56 percent in about a month, Nebius about 48 percent from its peak, and shorts such as Adobe rallied (CNBC, Jul 30 2026).
• On July 24 an investor letter said the fund had “not been immune” and called the sell-off one of the best buying windows since early 2025, inviting clients to add cash (The Next Web).
• Prime brokers (Bank of America, Goldman Sachs, JPMorgan) issued margin calls. Reuters: Ken Griffin “pulled an all-nighter” with senior Citadel staff “to analyze the trading book positions of Situational Awareness and how liquid the bets were”; the portfolio “lost 67% of its value” in July (Reuters via The Star, Aug 1 2026). On July 30 the fund sold its entire public equity book to Citadel in one trade; Bloomberg put the remaining assets at about $10 billion (Bloomberg; Quartz).
• Griffin’s August 21 letter: Citadel cut its exposure to the purchased book by more than 80 percent through more than 100 block trades worth over $4 billion; “Our ability to distribute this risk was central to our investment thesis” (Quartz via Yahoo Finance).
• Days after the forced sale the fund “completed a $400 million private investment” (Investing.com, citing Bloomberg).
• What survived is the private book, chiefly a stake in Anthropic the FT valued at about $5 billion. A spokesman told CNBC that reports the fund was shopping the stake were “not accurate.”

Figures conflict across outlets (a peak “past $20 billion” against “$45 billion”; a loss of “$35 billion” against “67 percent in July”). Every outlet agrees on the forced sale of the whole public book to Citadel and on the Anthropic stake kept.

7. The Anthropic exposure

Anthropic has “confidentially submitted a draft registration statement on Form S-1 … for a proposed initial public offering” (Anthropic release, via StorageNewsletter, Jun 9 2026). The fund’s surviving core is a private Anthropic stake, and its founder’s partner is chief of staff to Anthropic’s chief executive. A listing would bring that stake, and that relationship, under public-market disclosure and trading rules.

The argument

The essay’s case is that AI capability will outrun control, that the race with China makes the outcome a matter of national survival, and that the labs are not secure. The fund was a leveraged long on the same build-out the essay warns about. That is not a contradiction he hid: the essay argues the build-out is coming either way and that the United States should win it. It is a position: the man who wrote the warning was positioned to profit from the thing it warned about, until the leverage met a market that did not wait for 2027.

Related

• Watching the Watchers · The EA to AI-Safety Funding Pipeline · What the Data Shows
• Revolving Door · AI Governance Tracker