AI Wiki

Leopold Aschenbrenner

AI SafetyPeople
22 min read
Updated
Suggest editHistory
RawGraph

Last reviewed

Sources

No citations yet

Review status

Needs citations

Revision

v2 · 4,431 words

Leopold Aschenbrenner is a German-born artificial-intelligence researcher and investor best known for the June 2024 essay series Situational Awareness: The Decade Ahead and for founding the AI-focused hedge fund Situational Awareness LP. Before launching the firm, he was a member of OpenAI's Superalignment team from 2023 until April 2024, when he was dismissed in circumstances he and the company described differently; Aschenbrenner attributes the firing largely to an internal memo he circulated arguing that OpenAI's security posture was inadequate for AGI-relevant research, while OpenAI characterized the dismissal as a response to information-sharing violations.[1][2] His essay series, published at situational-awareness.ai on June 4, 2024, argues that scaling-law extrapolations make AGI by roughly 2027 plausible and that an "intelligence explosion" yielding superintelligence could follow shortly thereafter, with sweeping consequences for U.S.-China geopolitics, energy infrastructure, and AI-lab security.[3][4] Following publication, Aschenbrenner launched Situational Awareness LP, which by early 2026 disclosed roughly $13.7 billion in U.S. equity and options exposure on a Form 13F filing.[5][6]

Early life and education

Aschenbrenner was born in Germany around 2001-2002 and attended the John F. Kennedy School in Berlin, a German-American bilingual school.[1] At 15 he enrolled at Columbia University, and in 2021, at age 19, he graduated as the valedictorian of Columbia College with a Bachelor of Arts in economics and mathematics-statistics.[7][1] In its 2021 valedictorian announcement, Columbia College reported that an economics professor said "Leopold's record of scholarship exceeds that of any student in the department in the last 20 years."[7] At Columbia he also co-founded the campus chapter of Effective Altruism, an intellectual milieu that would shape his subsequent research interests in longtermism and existential risk.[1]

While still an undergraduate, Aschenbrenner began publishing economics work tied to long-run growth and catastrophic risk. He received a grant from economist Tyler Cowen's Emergent Ventures program, with Cowen later describing him publicly as an "economics prodigy."[4] His personal blog, For Our Posterity (forourposterity.com), contains posts dating from 2020 onward on topics including European political stagnation, the work of growth economist Chad Jones, and "Burkean longtermism," an attempt to reconcile conservative philosophical premises with very long-horizon ethical reasoning.[8]

Early career: FTX Future Fund, Oxford, and Open Philanthropy

After graduating from Columbia in 2021, Aschenbrenner moved into the effective-altruism research and grantmaking world. From February 2022 he worked as a research analyst at the FTX Future Fund, the philanthropic arm of Sam Bankman-Fried's FTX exchange that funded long-termist projects in AI safety, biosecurity, and economic growth.[1][9] He resigned ahead of FTX's November 2022 collapse, alongside the broader Future Fund leadership team that included William MacAskill, Nick Beckstead, and Avital Balwit.[4][9]

In parallel, Aschenbrenner conducted research at Oxford University's Global Priorities Institute (GPI), where he co-authored the working paper "Existential Risk and Growth" with economist Philip Trammell. The paper formalizes a model in which faster technological growth raises short-run catastrophic-risk hazard but, by raising societal willingness to pay for safety, can yield an "existential-risk Kuznets curve" with lower lifetime risk; the current revised version is dated January 2026 and is listed as GPI Working Paper No. 13-2024.[10] The model has since been read by some commentators as part of the intellectual foundation for Aschenbrenner's later "race-but-with-security" stance on AGI.[4]

He has also contributed shorter pieces to the long-termist magazine Works in Progress and to Open Philanthropy-adjacent research circles, though he was not himself an employee of Open Philanthropy.[4][8] Press accounts have noted that Aschenbrenner's social and professional network during this period overlapped substantially with the founding cohort of younger longtermist researchers and grantmakers, including William MacAskill, Avital Balwit (later chief of staff to Anthropic CEO Dario Amodei), and Nick Beckstead, the FTX Foundation CEO who had previously spent several years at Open Philanthropy.[4][9] His blog posts from this period, including pieces on Burkean longtermism and on long-run economic growth (a 2021 post titled "My Favorite Chad Jones Papers"), establish the intellectual frame for the GPI working paper and for parts of his later AGI writing.[8]

OpenAI Superalignment role

In 2023, Aschenbrenner joined the newly created Superalignment team at OpenAI, a group co-led by Ilya Sutskever and Jan Leike that was charged with making conceptual and engineering progress on aligning systems "much smarter than humans" before such systems were built.[1][11] OpenAI publicly announced the team in July 2023 with a commitment to dedicate 20% of its secured compute over four years to the effort.[11]

Aschenbrenner was a co-author of the team's flagship empirical paper, "Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision," which was posted to arXiv on December 14, 2023 and presented at ICML 2024.[12] The paper, led by Collin Burns, studies the analogue problem in which a weaker model supervises a stronger pretrained model on natural-language, chess, and reward-modeling tasks; the authors report that fine-tuned strong students consistently outperform their weak supervisors, a phenomenon the paper labels weak-to-strong generalization. The other listed co-authors include Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Yining Chen, Adrien Ecoffet, Manas Joglekar, Jan Leike, Ilya Sutskever, and Jeff Wu.[12] The experimental setup uses a range of pretrained models in the GPT-4 family and was framed as an empirical analogue of the alignment problem in which humans must supervise systems substantially more capable than themselves. The Superalignment team accompanied the paper with a blog post and an open-source code release on GitHub.[12]

Aschenbrenner's work on the Superalignment team also reportedly included internal forecasting and threat-modeling efforts focused on the security of frontier-model artifacts (especially trained weights and inference-time algorithmic secrets) and on the operational risks of state-sponsored industrial espionage against U.S. AI labs.[2][13] These internal concerns formed the substantive basis for the security memo at the center of his April 2024 dismissal.

Dismissal in April 2024

OpenAI fired Aschenbrenner in April 2024. According to his own subsequent account, delivered most fully on the Dwarkesh Patel podcast in June 2024, the proximate reasons OpenAI gave at termination were two: (1) sharing what he described as a brainstorming "preparedness, safety and security" document with three external researchers for feedback, which he says was scrubbed of sensitive information and contained a reference to internal planning for AGI in 2027-2028, and (2) an earlier "security memo" he had circulated internally and later sent to the OpenAI board arguing that the company's protections against model-weight theft and algorithmic exfiltration were inadequate for AGI-relevant work, particularly with respect to Chinese state-sponsored industrial espionage.[2][13] Aschenbrenner has said he was told the security memo was a major reason for his dismissal; OpenAI publicly disputed that characterization and described the firing as the result of an information leak unrelated to retaliation.[2][14]

The Superalignment team was effectively dissolved roughly a month after Aschenbrenner's departure: Sutskever announced his exit on May 14, 2024, and Leike followed shortly thereafter, citing disagreements with OpenAI's leadership over the company's safety priorities and compute allocation for alignment research.[15] Sutskever subsequently founded the AI lab Safe Superintelligence Inc, and Aschenbrenner dedicated his Situational Awareness essay series "to Ilya Sutskever."[3]

Situational Awareness: The Decade Ahead

On June 4, 2024, roughly two months after his departure from OpenAI, Aschenbrenner self-published Situational Awareness: The Decade Ahead, a roughly 165-page monograph released as a six-part web series at situational-awareness.ai and as a downloadable PDF.[3][16] The work synthesizes scaling-law extrapolations, geopolitical analysis, and AI-lab security commentary into a single forecast: that the largest AI labs, on current trends, will plausibly reach AGI around 2027 and that an intelligence explosion to superintelligence could occur in the subsequent few years.[3]

The series is structured as an introduction followed by five numbered chapters:

SectionTitleCore claim
IntroIntroduction"You can see the future first in San Francisco." Frames the essay as an attempt to convey what insiders at frontier labs already take for granted.
IFrom GPT-4 to AGI: Counting the OOMsExtrapolating roughly 0.5 OOM/year in effective compute and 0.5 OOM/year in algorithmic efficiency, plus "unhobbling" gains from agent scaffolding and chain-of-thought, yields a GPT-2-to-GPT-4-sized jump by ~2027.
IIFrom AGI to Superintelligence: the Intelligence ExplosionOnce AGI can perform AI research, "hundreds of millions of AGIs could automate AI research, compressing a decade of algorithmic progress (5+ OOMs) into <=1 year."
IIIaRacing to the Trillion-Dollar ClusterCluster sizes will scale toward ~$1T capex and ~100 GW of power draw by ~2030, comparable to a meaningful fraction of U.S. electricity production.
IIIbLock Down the Labs: Security for AGIU.S. AI labs currently have weaker security than serious hedge funds; without dramatic improvements, key algorithmic secrets will leak to the People's Republic of China within 12-24 months.
IIIcSuperalignmentAligning superhuman systems is an unsolved technical problem and is likely to be "the most important technical challenge of our time."
IIIdThe Free World Must PrevailA U.S.-led coalition must maintain a durable lead over China to set the rules of a post-AGI international order.
IVThe ProjectThe U.S. government will, in his prediction, take a Manhattan-Project-style direct role in frontier AI by 2027-2028, with a "joint venture between the major cloud compute providers, AI labs, and the government."
VParting ThoughtsClosing reflections on responsibility, governance, and the historical novelty of the moment.

[3][16][17][18]

Key arguments and methodology

The central methodological move in Chapter I is "counting the orders of magnitude" of effective compute. Aschenbrenner separates progress into three streams: physical compute scaling (he estimates roughly 0.5 OOMs of training compute per year), algorithmic efficiency (another ~0.5 OOMs per year, with ImageNet history cited as a calibration point), and "unhobbling" (techniques such as chain-of-thought reasoning, tool use, and agent scaffolding that turn raw next-token predictors into something closer to drop-in remote workers).[16] He argues that, summed, these trends produce another GPT-2-to-GPT-4-magnitude leap by 2027, enough in his view to automate much of AI research itself.[16] In the same chapter he characterizes the GPT-4 baseline as roughly "smart high-schooler" level, sufficient to write nontrivial code, reason through competition-mathematics problems, and serve as a daily coding assistant; he predicts that the equivalent jump again will yield systems behaving as drop-in remote workers capable of automating much of the work performed by AI engineers themselves.[16]

Chapter II uses that prediction as input to an intelligence-explosion argument familiar from earlier writers such as I. J. Good and Nick Bostrom: once AI researchers can be automated, the next several OOMs of algorithmic improvement may be compressed into months rather than decades, with corresponding implications for capability gains across robotics, scientific R&D, and military technology.[17] Specific quantitative claims in this chapter include the projection that "hundreds of millions of AGIs could automate AI research, compressing a decade of algorithmic progress (5+ OOMs) into <=1 year" and that individual systems would operate at 10x-100x human speed on cognitive tasks, with aggregate inference fleets capable of generating on the order of one trillion tokens per hour.[17] Aschenbrenner draws an analogy to the leap from atomic to thermonuclear weapons: superintelligence, in his framing, would be qualitatively different from human-level AI in the way the hydrogen bomb was qualitatively different from a fission device, exhibiting "novel, creative, complicated behavior we couldn't even begin to understand."[17]

Chapter IIIa pushes the resource-side of the argument into specifics. Aschenbrenner sketches a trajectory from a 1 GW training cluster around 2026 toward a 10 GW cluster around 2028 and a 100 GW, ~$1 trillion capex cluster by 2030, framing electricity supply and chip-fab throughput as the binding constraints.[16] The implied infrastructure build-out has become a touchstone in subsequent industry discussion of projects such as the Stargate Project and the Abilene data center in Texas.[5]

Chapter IIIb argues that frontier AI labs do not currently approach the security level of major hedge funds, let alone of national-security-relevant facilities, and that algorithmic and weight-level theft by Chinese state actors is plausible "in the next 12-24 months" absent dramatic upgrades.[18][19] The chapter distinguishes between weight-level security (protecting trained model parameters from exfiltration) and algorithmic-secret security (protecting the engineering know-how that determines how a successful training run is composed), arguing that the latter must be addressed years before the former because the secrets are being produced in labs today, well before AGI-level weight protection becomes critical.[18] Aschenbrenner has illustrated the point colorfully, saying in social-media commentary tied to the launch that an adversary "doesn't need to mount a dramatic espionage operation," and could instead obtain key information by attending parties in San Francisco, bribing cleaning staff, or looking through office windows.[19] Chapter IIIc surveys the technical AI alignment problem and concedes uncertainty about whether superalignment is solvable on the predicted timelines, while Chapter IIId argues that, given those risks, the only acceptable outcome is for a U.S.-led coalition of liberal democracies to maintain a durable lead.[3]

Chapter IV, "The Project," makes the most distinctive political prediction in the series: that the U.S. national-security state will, around 2027-2028, take direct operational control of frontier AGI development through a public-private structure resembling the Manhattan Project, integrating leading labs, hyperscalers, and government agencies under a unified chain of command.[17] In Aschenbrenner's description, this would not necessarily mean literal nationalization of private companies but rather a "joint venture between the major cloud compute providers, AI labs, and the government" operating under defense-contracting frameworks, with elements including SCIF-style classified facilities, a consolidated research team drawn from leading labs, and a unified chain of command for any frontier military applications, with eventual civilian use post-stabilization.[17] Chapter V, "Parting Thoughts," frames the essay as a personal letter rather than a forecast and is dedicated to Sutskever.[3] The published acknowledgments thank a wide range of correspondents and reviewers, including Collin Burns, Avital Balwit, Carl Shulman, Jan Leike, Ilya Sutskever, Holden Karnofsky, Sholto Douglas, James Bradbury, and Dwarkesh Patel.[3]

Reception

Situational Awareness circulated widely outside the AI-safety community within days of publication. Axios summarized it as "AI from now to 2034" and called it a defining set of arguments for those expecting a near-term AGI transition.[20] Fortune reported in 2025 that the document "circulated among policymakers like the juiciest classified NSA assessment," and quoted computer scientist Scott Aaronson saying that the essay "is the document some general or national-security person is going to read and say: 'This requires action.'"[4] Stanford's Digital Economy Lab hosted a public event with Aschenbrenner discussing the essay shortly after release.[23]

The series was also extensively criticized. Writers on the EA Forum and alignment-adjacent fora such as LessWrong argued that the document overstates the tractability of "superalignment" on its proposed timelines, frames a U.S.-China AI competition as a foregone conclusion, and underweights the chance of a slower or qualitatively different trajectory.[21][22] Some commentators noted that the framing of AGI as a national-security race might encourage governments and labs to deprioritize safety in exactly the regime Aschenbrenner himself describes as unsafe.[21] Critics also focused on the essay's "data wall" caveat: Aschenbrenner acknowledges that frontier training runs are running out of high-quality natural-language tokens, but the essay treats algorithmic recovery from this constraint as relatively likely, an assumption that some reviewers regard as load-bearing and underdefended.[21]

Effective-altruism forum analysts have begun publishing retrospective reviews evaluating the accuracy of specific predictions, including model-capability milestones, compute build-outs, and political developments such as the public announcement of large U.S. compute build-outs in 2025.[22] Reaction pieces by Zvi Mowshowitz on Substack and by various authors on LessWrong have noted that several of the trillion-dollar-cluster, energy-bottleneck, and U.S.-led-coalition framings have at least partially come to pass in subsequent industry announcements, while expressing reservations about the precise 2027 AGI claim and about the political-economy assumptions underlying "The Project."[21][22]

Dwarkesh Patel podcast

The series was launched in tandem with a long-form interview on the Dwarkesh Patel podcast titled "Leopold Aschenbrenner: 2027 AGI, China/US super-intelligence race, & the return of history," released on June 4, 2024. The episode runs roughly 4 hours and 29 minutes and is structured around the same six themes as the essay: the OOM accounting, the trillion-dollar cluster, espionage and lab security, the project, alignment, and "the return of history" framing of great-power competition.[13] It is the venue in which Aschenbrenner gave his most extended public account of his OpenAI dismissal.[2][13]

He has since appeared on a range of additional podcasts and at policy events, including a Stanford Digital Economy Lab session on Situational Awareness.[23]

Situational Awareness LP

Within months of publishing the essay series, Aschenbrenner launched Situational Awareness LP, a San Francisco-based hedge fund explicitly built around the investment thesis of the Situational Awareness monograph: that public-equity markets systematically underprice the physical inputs (compute, electricity, fabs, data-center real estate) and certain enabling firms required to support a continued AGI build-out.[4][5] Carl Shulman, a longtime AI forecaster and former employee of Peter Thiel's Clarium Capital, was hired as director of research and co-portfolio manager.[4][24]

Anchor investors disclosed in press coverage include Nat Friedman (former GitHub CEO), Daniel Gross (Friedman's investing partner), and Stripe co-founders Patrick and John Collison; Fortune reports additional commitments from West Coast founders, family offices, institutions, and endowments.[4] As of an October 2025 Fortune profile, the fund was reported at roughly $1.5 billion in assets, with its first-half-2026 performance described as a 47% gain net of fees; reporting from May 2026 placed total disclosed U.S. equity and options exposure at roughly $13.7 billion across its 13F filings.[4][5][6]

Disclosed positions and notable trades reported across SEC filings and press coverage include large long exposures to U.S. power producers (Vistra, Constellation Energy), data-center-adjacent crypto miners (Core Scientific), the Nvidia-heavy VanEck Semiconductor ETF, and semiconductor companies such as Intel and Broadcom; analysts have noted that the fund's Q1 2026 13F filing also disclosed substantial put-option positions on Nvidia and on the VanEck Semiconductor ETF, which commentators have interpreted as a hedge rather than a directional bearish view.[25][5][6] Fortune's March 2026 follow-up profile emphasized the fund's bet on electricity producers and bitcoin miners as proxies for AI energy demand and framed the thesis as "the most valuable assets in the AI era may not be algorithms, but electricity and computing power."[26] An earlier fund position in Bloom Energy, a U.S. solid-oxide-fuel-cell manufacturer that supplies on-site power for data centers, has been highlighted as a direct expression of the energy thesis; the stock subsequently rose substantially over the following year, contributing to the fund's reported performance.[25][26]

Press reporting on the fund's Q1 2025 13F filings, which it began disclosing soon after launch, noted approximately $459 million in Intel call options as a single named position; subsequent reporting on the fund's January 2025 reaction to the DeepSeek-R1 release indicated that, rather than panic-selling AI-exposed positions, Situational Awareness used the broad-market drawdown to add to existing exposures.[4]

The fund is run as a publicly-traded-equities vehicle rather than a venture-style allocator, in contrast to Friedman and Gross's separate AI Grant program.[4] As of 2026 reporting, its portfolio is concentrated in fewer than 30 disclosed positions, with the U.S. power-and-data-center complex, the semiconductor supply chain, and selective use of options as the three main components.[5][6]

Public commentary on AGI timelines

Outside the essay series, Aschenbrenner has used his blog, podcast appearances, and X account to argue a set of consistent positions: that scaling-law trends in deep learning have substantially more runway than skeptics assume, that compute and electricity are the binding inputs to frontier capability gains, that algorithmic secrets at frontier labs are the most security-sensitive artifacts in the system, and that AGI-era industrial policy is likely to be dominated by U.S.-China strategic competition.[3][13][18] In a widely shared June 2024 post on X coinciding with the essay launch, he wrote that AI labs were "barely making an effort" on security and that the U.S. was "probably leaking key AGI breakthroughs to the CCP in the next 12-24 months."[18]

His earlier blog post "Nobody's on the Ball on AGI Alignment" (March 2023) advanced a similar critique inside the AI safety field, arguing that the number of researchers working specifically on alignment of superhuman systems was far too small relative to the predicted timeline of capability gains.[8]

Significance

Several aspects of Aschenbrenner's recent work have been singled out by commentators as distinctive:

  • Synthesis across scaling, security, and geopolitics. The Situational Awareness series is unusual in combining detailed quantitative scaling-law extrapolation with arguments about industrial espionage, national-security industrial policy, and energy infrastructure under a single thesis.[4][17]
  • Direct policy influence. Coverage in Axios and Fortune credits the essay with shaping how some U.S. policymakers and defense-adjacent investors think about an "AGI race" framing; the same coverage notes critics' concern that this framing can crowd out safety considerations.[20][4]
  • Bridging AI safety and finance. Through Situational Awareness LP, Aschenbrenner became one of the first AI-safety-aligned figures to operate a large public-equities fund whose thesis is itself a published AGI-timelines forecast, an unusual combination of intellectual and financial commitment to a specific scenario.[4][24]

Criticisms

Critiques of Aschenbrenner's public work cluster along several axes:

  1. Confidence and timeline. Critics on the EA Forum, LessWrong, and elsewhere argue that the OOM accounting underweights uncertainty in algorithmic gains and assumes continued availability of high-quality training data; some retrospective reviews have flagged data-wall concerns Aschenbrenner himself acknowledges in passing.[21][22]
  2. Race framing. A recurring concern is that framing AI as a U.S.-China zero-sum race may pressure labs and governments to accelerate beyond the safety thresholds the same essays argue are necessary, an internal tension that several commentators have called the central paradox of the series.[21]
  3. Alignment optimism. Some safety researchers argue that the essay treats "superalignment" as a hard but solvable engineering problem on the relevant timelines, an assumption that AI-safety writers such as those at LessWrong have called insufficiently justified.[22]
  4. Conflict-of-interest perceptions. Because Aschenbrenner now manages a fund whose returns benefit from the AGI build-out he predicts, some commentators have raised concerns about the alignment of his analytical and financial incentives; he has publicly stated that the fund's exposures follow from rather than drive his published views.[4]
  5. OpenAI's account of the dismissal. OpenAI has publicly disputed Aschenbrenner's characterization of his April 2024 dismissal, presenting the firing as a response to information-sharing violations rather than to whistleblowing about lab security; the two accounts remain in tension.[2][14]

Personal life

Aschenbrenner lives in San Francisco. Press coverage in 2025 reported that he is engaged to Avital Balwit, a colleague from the FTX Future Fund era who has worked as chief of staff to Anthropic CEO Dario Amodei.[1][4]

Aschenbrenner's Situational Awareness shares conceptual ancestry with earlier scaling-driven AGI-timelines arguments, including the Scaling Laws tradition initiated by the 2020 Kaplan et al. paper and the Chinchilla scaling laws refinement, both of which underlie his "counting the OOMs" methodology.[16] His arguments about existential risk from AI connect his GPI economics paper with Bostrom's Superintelligence tradition and with Holden Karnofsky's "most important century" essays.[4][10] His emphasis on superintelligence as a near-term operational concept echoes both Geoffrey Hinton's post-2023 public warnings and Sutskever's framing in launching Safe Superintelligence Inc.[15]

On the industrial side, the essay's "trillion-dollar cluster" framing is closely related to the announcements of the Stargate Initiative and the build-out of the Abilene data center, which post-date but echo Aschenbrenner's predictions.[5][16]

See also

References

  1. "Leopold Aschenbrenner", Wikipedia, accessed 2026-05-20. https://en.wikipedia.org/wiki/Leopold_Aschenbrenner. Accessed 2026-05-20. ↩
  2. Shakeel Hashim, "OpenAI employee says he was fired for raising security concerns to board", Transformer News, 2024-06-04. https://www.transformernews.ai/p/openai-employee-says-he-was-fired. Accessed 2026-05-20. ↩
  3. Leopold Aschenbrenner, "Situational Awareness: The Decade Ahead (Introduction)", situational-awareness.ai, 2024-06-04. https://situational-awareness.ai/. Accessed 2026-05-20. ↩
  4. Jeremy Kahn, "How former OpenAI researcher Leopold Aschenbrenner turned a viral AI prophecy into profit, with a $1.5 billion hedge fund and outsize influence from Silicon Valley to D.C.", Fortune, 2025-10-08. https://fortune.com/2025/10/08/leopold-aschenbrenner-openai-ftx-1-5-billion-hedge-fund-situational-awareness/. Accessed 2026-05-20. ↩
  5. Dave Manuel, "The Most-Anticipated 13F of 2026: What Leopold Aschenbrenner Actually Owns", DaveManuel.com, 2026-05-19. https://www.davemanuel.com/2026/05/19/leopold-aschenbrenner-q1-2026-13f-situational-awareness/. Accessed 2026-05-20. ↩
  6. Quiver Quantitative, "Former OpenAI Employee's Hedge Fund Unveils Massive Nvidia and AI Chip Options Positions", Quiver Quantitative, 2026-05-16. https://www.quiverquant.com/news/Former+OpenAI+Employee%E2%80%99s+Hedge+Fund+Unveils+Massive+Nvidia+and+AI+Chip+Options+Positions. Accessed 2026-05-20. ↩
  7. Columbia College, "Columbia College Announces 2021 Valedictorian and Salutatorian", Columbia College, 2021-04-30. https://www.college.columbia.edu/news/columbia-college-announces-2021-valedictorian-and-salutatorian. Accessed 2026-05-20. ↩
  8. Leopold Aschenbrenner, "For Our Posterity (blog index)", forourposterity.com, accessed 2026-05-20. https://www.forourposterity.com/. Accessed 2026-05-20. ↩
  9. FTX Future Fund, "About", ftxfuturefund.org, 2022. https://ftxfuturefund.org/about/. Accessed 2026-05-20. ↩
  10. Leopold Aschenbrenner and Philip Trammell, "Existential Risk and Growth (GPI Working Paper No. 13-2024)", Global Priorities Institute, University of Oxford, 2026-01-05. https://www.globalprioritiesinstitute.org/wp-content/uploads/Leopold-Aschenbrenner-and-Philip-Trammell-Existential-Risk-and-Growth-2.pdf. Accessed 2026-05-20. ↩
  11. OpenAI, "Introducing Superalignment", openai.com, 2023-07-05. https://openai.com/index/introducing-superalignment/. Accessed 2026-05-20. ↩
  12. Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Leopold Aschenbrenner, et al., "Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision", arXiv:2312.09390, 2023-12-14. https://arxiv.org/abs/2312.09390. Accessed 2026-05-20. ↩
  13. Dwarkesh Patel, "Leopold Aschenbrenner: 2027 AGI, China/US super-intelligence race, & the return of history", Dwarkesh Podcast, 2024-06-04. https://www.dwarkesh.com/p/leopold-aschenbrenner. Accessed 2026-05-20. ↩
  14. Tom Jowitt, "OpenAI Fires Two Researchers Over Information Leaks", Silicon UK, 2024-04-12. https://www.silicon.co.uk/e-innovation/artificial-intelligence/openai-fires-researchers-558601. Accessed 2026-05-20. ↩
  15. Cade Metz, "OpenAI's Chief Scientist Ilya Sutskever Departs Company", The New York Times (via reporting roundup), 2024-05-14. https://www.nytimes.com/2024/05/14/technology/ilya-sutskever-leaves-openai.html. Accessed 2026-05-20. ↩
  16. Leopold Aschenbrenner, "I. From GPT-4 to AGI: Counting the OOMs", situational-awareness.ai, 2024-06-04. https://situational-awareness.ai/from-gpt-4-to-agi/. Accessed 2026-05-20. ↩
  17. Leopold Aschenbrenner, "II. From AGI to Superintelligence: the Intelligence Explosion", situational-awareness.ai, 2024-06-04. https://situational-awareness.ai/from-agi-to-superintelligence/. Accessed 2026-05-20. ↩
  18. Leopold Aschenbrenner, "IIIb. Lock Down the Labs: Security for AGI", situational-awareness.ai, 2024-06-04. https://situational-awareness.ai/lock-down-the-labs/. Accessed 2026-05-20. ↩
  19. Leopold Aschenbrenner, post on X (formerly Twitter), x.com/leopoldasch, 2024-06-04. https://x.com/leopoldasch/status/1798483689455964585. Accessed 2026-05-20. ↩
  20. Ina Fried, "Leopold Aschenbrenner's 'Situational Awareness': AI from now to 2034", Axios, 2024-06-23. https://www.axios.com/2024/06/23/leopold-aschenbrenner-ai-future-silicon-valley. Accessed 2026-05-20. ↩
  21. Various authors, "Response to Aschenbrenner's 'Situational Awareness'", EA Forum, 2024-06-15. https://forum.effectivealtruism.org/posts/RTHFCRLv34cewwMr6/response-to-aschenbrenner-s-situational-awareness. Accessed 2026-05-20. ↩
  22. Various authors, "How did Leopold do? Evaluating Situational Awareness's predictions", EA Forum, 2025. https://forum.effectivealtruism.org/posts/RuwF8FCfpsLeZRgur/how-did-leopold-do-evaluating-situational-awareness-s. Accessed 2026-05-20. ↩
  23. Stanford Digital Economy Lab, "Leopold Aschenbrenner: Situational Awareness (event)", Stanford Digital Economy Lab, 2024. https://digitaleconomy.stanford.edu/event/leopold-aschenbrenner-situational-awareness/. Accessed 2026-05-20. ↩
  24. Przemek Chojecki, "Situational Awareness LP: AI Hedge Fund in 2025", Medium, 2025. https://pchojecki.medium.com/situational-awareness-lp-ai-hedge-fund-in-2025-a2d9e88d0c05. Accessed 2026-05-20. ↩
  25. Linas Beliunas, "Leopold Aschenbrenner's $5.5B Situational Awareness Fund", Linas's Newsletter (Substack), 2025. https://linas.substack.com/p/leopold-aschenbrenner-situational-awareness-fund-portfolio-playbook. Accessed 2026-05-20. ↩
  26. Jeremy Kahn, "Why Leopold Aschenbrenner's AI hedge fund is betting big on power companies and Bitcoin miners to fuel the AI boom", Fortune, 2026-03-05. https://fortune.com/2026/03/05/leopold-aschenbrenner-ai-hedge-fund-superintelligence-agi-power-companies-crypto-miners/. Accessed 2026-05-20. ↩

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation.

Suggest edit