Affiliation: Founder, Situational Awareness LP (AI-focused hedge fund, 2024–); previously OpenAI Superalignment team (2023–04/2024); Forethought Foundation; Columbia University valedictorian One-line position: AGI is plausibly here by 2027, an intelligence explosion follows within ~2 years, and the only safe path is a US-government-led national AGI project that out-builds and out-secures China.
What he's reacting against
- "Scaling will plateau" skepticism — argues the trend lines are obvious to anyone who reads the curves
- Lab-led AGI development without national-security-grade weight security — argues current frontier labs cannot defend model weights from PRC exfiltration
- Slow-takeoff framings — argues recursive self-improvement compresses ASI emergence into ~1–2 years after AGI
- Decoupling AI from geopolitics — argues AGI is a "decisive military advantage" question, not a tech-industry question
- "Soft" governance (voluntary commitments, RSPs) at frontier scale — argues only USG can provide the security perimeter and the resource mobilization needed
- Pause / moratorium framings (Yudkowsky) — argues stopping is impossible and free-world build is the safer strategy
Key claims
- Empirical: effective compute = physical compute × algorithmic efficiency; both grow ~0.5 OOM/year, so effective compute compounds ~1 OOM/year through 2027
- Empirical: "drop-in remote workers" — AGI agents performing knowledge work end-to-end — emerge by 2025–2026 [TBD: precise milestone definition]
- Empirical: AGI is "strikingly plausible" by 2027 — defined as systems that can do the cognitive work of a top AI researcher
- Empirical: post-AGI, recursive self-improvement compresses the AGI→ASI transition to ~1–2 years (intelligence explosion)
- Empirical: a ~$1T training cluster will exist by 2027–2028; AI will consume ~10% of US electricity by 2030
- Empirical: current frontier-lab security is inadequate — PRC will successfully exfiltrate frontier weights within years absent USG-grade security
- Empirical / political: whichever great power gets to ASI first secures a durable, possibly century-long military advantage
- Normative: USG should run a Manhattan/Apollo-scale national AGI project, with private labs as contractors, under classified security
- Normative: the free world must win the race; authoritarian-controlled ASI is the central x-risk
- Mixed: alignment is technically solvable on aggressive timelines but only if the strongest researchers work on it in a serious institutional setting — "muddling through" is plausible, not guaranteed
Theories aligned with
- AI alignment / x-risk — national-security / "race to the top with USG" variant, distinct from MIRI / Yudkowsky and from Amodei's lab-led version
- Adjacent to techno-optimism — but conditional on USG-led build, not market-led
- Effectively defines a new "national-AGI-project" position that doesn't have a clean theory file yet
Where he overlaps / splits
- Overlaps with Amodei on aggressive timeline (~2027) and on race-to-the-top framing; splits on who leads — Aschenbrenner says USG, Amodei says safety-focused labs
- Overlaps with Altman on US-vs-China framing and on trillion-dollar cluster scale; splits on whether OpenAI's current governance is adequate (Aschenbrenner left OpenAI 2024-04 amid security/safety disputes)
- Overlaps with Hassabis on aggressive capability timeline; splits on geopolitical framing — Hassabis less hawkish on US-China
- Splits sharply with Yudkowsky, Bengio on pause vs. build — Aschenbrenner says building faster (with USG security) is the safety strategy
- Splits sharply with Andreessen on USG involvement — Andreessen treats regulation as capture; Aschenbrenner treats USG as the only entity capable of doing it right
- Splits with Mostaque on open-source frontier — Aschenbrenner treats open frontier weights as catastrophic for security
- Splits with Acemoglu, Autor, Brynjolfsson, Cowen on framing — labor questions are downstream of the AGI race; their timeline assumptions don't survive 2027 AGI
- Splits with Frey on adjustment horizon — Aschenbrenner's timelines collapse the multi-decade transition Frey describes
Notable predictions
- (2024-06) AGI "strikingly plausible" by 2027 — outcome: TBD; central live falsifier
- (2024-06) "Drop-in remote workers" doing end-to-end knowledge work by 2025–2026 — outcome: TBD; partial agentic capability emerging in 2025 (Claude Code, ChatGPT Agents, etc.) but not full drop-in knowledge-worker yet [TBD: rigorous benchmark]
- (2024-06) Trillion-dollar individual training cluster by 2027–2028 — outcome: TBD; directionally consistent with Stargate (~$500B announced 2025-01) and Microsoft/Meta/Google capex
- (2024-06) ~10% of US electricity to AI by 2030 — outcome: TBD; grid build-out and data-center power demand consistent with trajectory
- (2024-06) PRC succeeds in stealing US frontier weights absent USG-grade security — outcome: TBD; not publicly verifiable
- (2024-06) Intelligence explosion (AGI → ASI) compresses to ~1–2 years post-AGI via recursive self-improvement — outcome: TBD; conditional on AGI being reached
- (2024-06) Algorithmic efficiency continues at ~0.5–1 OOM/year — outcome: TBD; consistent with frontier model efficiency trends through 2025
Track record
- Forethought Foundation research → OpenAI Superalignment team (2023); fired 2024-04 (he claims for raising security concerns, OpenAI characterized differently)
- "Situational Awareness: The Decade Ahead" (2024-06-04) — ~165-page essay; became one of the most-discussed AI documents of 2024; cited by Altman, Amodei, policy figures
- Founded Situational Awareness LP (2024) — long-AI-build trade; investor backing real
- Short public track record — first major essay only 24 months old; influence disproportionate to age and résumé length
- Pattern: confident, specific, falsifiable — easy to score by 2027–28
- Prediction style sits in the middle of his fund's portfolio thesis — coincidence weight per CONVENTIONS rule 25
Empirical vs normative
- Empirical: scaling trajectory, capability milestones, compute and power demand, geopolitical security gaps
- Normative: USG must lead; free world must win; alignment effort must scale with capability; lab governance is inadequate at frontier
- Conflict-of-interest note: Situational Awareness LP is an AI-bull fund; his timeline forecasts have direct portfolio implications (CONVENTIONS rule 25). Weight accordingly — not dismissively given the essay predated public fund disclosures, but skeptically.
Sources
- Essays: "Situational Awareness: The Decade Ahead" (2024-06-04) — primary source, ~165 pages, free PDF
- Substack / blog: situational-awareness.ai (irregular updates 2024–) [TBD: 2025–26 posts]
- Podcasts: Dwarkesh Patel Podcast (2024-06) — ~4-hour interview, primary long-form source; ChinaTalk (2024-06) [TBD: confirm ep dates]; Bari Weiss / Honestly podcast [TBD: date]
- Talks: limited public talks pre-2025 [TBD: post-2025 conference appearances]
- Press coverage: New York Times profile (2024) [TBD: exact date]; Wired, The Information coverage of his OpenAI departure (2024-04)
- Original Forethought / academic output: limited public footprint pre-2023 [TBD]
Weak spots / open questions
- "Plausibly by 2027" is rhetorically vivid; the actual claim has wide error bars he hedges with — moving-goalpost watch per CONVENTIONS rule 18
- Treats scaling continuation as near-certain; doesn't seriously engage steel-manned scaling-wall arguments (LeCun, Gary Marcus, capability-plateau evidence on some benchmarks 2024–25)
- "Intelligence explosion" depends on recursive self-improvement working as theorized; current evidence is suggestive but not load-bearing
- National-security frame compresses the policy space; alternative frames (multilateral coordination, open-source proliferation as deterrent, civilian-led governance) get short treatment
- USG-led national AGI project is politically infeasible at the scale described — proposal is normatively coherent but operationally underdeveloped
- Conflict of interest is significant — fund's positioning is bullish on AI cluster build; forecasts and investment thesis align too neatly for fully independent reading
- Track record is too short to weight heavily — first major essay only 24 months old; no prior multi-year forecasting record
- Doesn't engage seriously with distributional / labor consequences — treats them as downstream of the AGI race, which assumes the race resolves on his timeline
Rich's take
- (your synthesis here)
converts-from: personas/leopold-aschenbrenner.md · schema v1 · AI & Society domain