Who he is
Affiliation: Founder and portfolio manager, Situational Awareness LP (AI-focused hedge fund, launched 2024 with a reported ~$225M). Role check run 2026-09-21: still running the fund — but not the fund the May file described. H1 2026 +439% net (investor letter reviewed by the FT, 2026-07-24); July 2026 −67% at up to ~4x leverage (borrowed money multiplying the bets); the entire public stock book sold to Citadel in one block before the open on 2026-07-30 (Reuters). About $10B remains, a third to a half of it one private Anthropic stake. To investors: "We let you down this month … I take full responsibility," and the public book will now run "on a fully-paid-for basis" (letter via Business Insider, 2026-07-31). JPMorgan cut off lending (FT, 2026-09-11); Clear Street is the new prime broker (the bank that lends a fund money and holds its stock); back in the market via fully-paid flex options since 2026-09-02 (CNBC, 2026-09-11). Previously OpenAI Superalignment (2023–2024-04, fired — he says over a security memo, OpenAI says a leak); FTX Future Fund (2022); Columbia valedictorian (2021). Married Avital Balwit, chief of staff to Anthropic CEO Dario Amodei, the weekend of 2026-08-01 (Fortune). Age ~24–25.
One-line position: unchanged in words — AGI plausibly by 2027, an intelligence explosion within ~2 years, and the only safe path is a US-government-led national project that out-builds and out-secures China. But he has said nothing in public since 2024-06-14. Filings and two leaked letters are the only new evidence, and they say: still all-in on the physical layer, now without borrowed money.
Discipline & technical bet
AGI forecaster who treats progress as a trendline — count the OOMs (orders of magnitude, 10x steps) and extrapolate — turned concentrated, formerly leveraged AI-infrastructure fund manager. The wager: effective compute (~0.5 OOM/yr physical + ~0.5 OOM/yr algorithmic + "unhobbling") keeps compounding, so the binding physical constraint migrates down the stack — chips → power → (per the Q2 2026 book) memory and storage — and value migrates with it. The fund is that wager as capital: Q2 2026 book $20.24B across 26 names (13F-HR, period 2026-06-30, accession 0000935836-26-000418, cover page verified this cycle), 56% in SanDisk + Micron, the Q1 $8.46B semiconductor put book closed; private stakes in Anthropic (~20% of the fund, bought ~Mar 2025 at ~$61.5B valuation, WSJ), MatX (co-led the $500M Series B, 2026-02-24) and Fluidstack. July proved the thesis at 4x cannot survive a one-month 29–47% drop in memory names plus a software-short squeeze. September shows the same names re-bought unlevered (AMD, Bloom Energy, CoreWeave, SK Hynix, SanDisk, a memory ETF) plus a 21.1% stake in SharonAI Holdings (SHAZ), $523.9M cost, on a Schedule 13D (filed when you own more than 5% and might act; event 2026-08-27), topped up by warrant exercise (13D/A + Form 4, 2026-09-17). Says and Does still point the same way; only the leverage changed.
Key claims (Says)
- Empirical: effective compute compounds ~1 OOM/yr through 2027 (~0.5 physical × ~0.5 algorithmic; unhobbling on top). Held — Epoch: compute for a given language-model result halves every ~8 months (≈0.45 OOM/yr); the two-year scorecard (EA Forum, 2026-06-10) calls the pace "roughly supported." Caveat: pretraining slowed relative to RL post-training (Dubach, 2026-05) — total held, mix shifted.
- Empirical: models "smarter than most college graduates" by 2025/26. Corrected at migration — v1 wrote "drop-in remote workers by 2025–26"; the essay ties drop-in workers to 2027 and assigns no earlier year (checked against situational-awareness.ai). Held — GPT-5.4 ~83% on GDPval knowledge-work tasks, ~80% SWE-Bench Pro (scorecard, 2026-06).
- Empirical: AGI "strikingly plausible" by 2027 — systems doing the work of an AI researcher/engineer. Open — check below.
- Empirical: post-AGI, recursive self-improvement compresses AGI→ASI to ~1–2 years. Conditional — per-branch check below.
- Empirical: the cluster ladder — ~1 GW / $10s of billions by 2026; ~10 GW / $100s of billions by 2028; ~100 GW / $1T+ by ~2030 needing ">20% of US electricity production"; total AI investment "north of $1T annually by 2027" (essay, "Racing to the Trillion-Dollar Cluster"). Corrected at migration — v1's "$1T cluster by 2027–28; ~10% of US electricity by 2030" were both transcription errors. On the essay's own rows: GW-class clusters by 2026 Held (Abilene/Colossus-class sites at ~1M H100-equivalents; scorecard 2026-06). $1T/yr by 2027 Held (tracking) — big-four 2026 capex guidance ~$725–732B; 2027 street forecasts ~$1.0T (Goldman $1.01T, JPMorgan $1T). 100 GW / >20% by 2030 Drifted — LBNL's June-2026 update puts all US data centers at 9.5–15.3% of US electricity in 2030 (reference 11.8%); direction right, magnitude ahead of every official forecast.
- Empirical: power is the binding physical constraint. Held — gas-turbine lead times ~6 years by May 2026 (Wood Mackenzie), data-center grid demand +22% in 2025 (S&P), both via Dubach (secondary; flagged).
- Empirical: frontier-lab security is inadequate — the PRC will exfiltrate frontier weights within years absent USG-grade security. Drifted — NSA/CISA/FBI advisory AA26-251A (2026-09-08) documents "industrial-scale" distillation (training on another model's answers) of Claude, GPT, Gemini and Grok by six PRC firms since late 2024, billions of tokens through the API front door; no confirmed weight theft. Mechanism wrong so far, valence right.
- Empirical (essay / Dwarkesh, 2024): open-weight models fall behind the proprietary frontier as training costs scale. Wrong — DeepSeek V4 and Qwen 3.7 Max sit ~3–6 months behind at a fraction of the price (scorecard, 2026-06); the essay's clearest miss, and it removes the "durable moat" leg of his geopolitics.
- Empirical / political: whichever great power reaches ASI first secures a durable, possibly century-long military advantage. Unfalsifiable — rewrite or retire.
- Normative: USG should run a Manhattan/Apollo-scale national AGI project, private labs as contractors, under classified security. Not graded; the paired prediction ("The Project" by 2027/28) is graded below.
- Normative: the free world must win the race; authoritarian-controlled ASI is the central x-risk. Not graded.
- Mixed: alignment is technically solvable on aggressive timelines if the strongest researchers work on it in a serious institutional setting; "muddling through" is plausible, not guaranteed. Drifted — alignment faking (Anthropic, 2024-12, arXiv 2412.14093) and scheming evaluations (Apollo, 2024-12) weakened the "generalizes benignly" premise; no public update from him.
- Fund claim (letter, 2026-07-24): the selloff is "one of the best buying opportunities since early last year"; investors invited to add capital from 2026-08-01. Wrong on his own horizon — six days later the public book was force-sold and July closed −67% (Reuters; TechCrunch).
Notable predictions — with falsifiable checks
- (2024-06) AGI by 2027. Open (EA Forum: 18 months left; Dubach: provisional through 2028). Check: by 2027-12-31 an independent evaluator (METR, Epoch) or a lab documents an AI system completing an end-to-end ML research project without human intervention; his "2027/28" hedge sets a hard stop of 2028-12-31, after which this is Wrong.
- (2024-06) Drop-in remote worker by ~2027 (corrected from v1's 2025–26). On track — METR 50% time horizon 5.3 h (Claude Opus 4.5, 2026-01-29), 80% horizon 12.2 h, recent doubling ~3 months; Cotra (2026-08-14) puts ~10% on human-parity AI R&D by end-2026. Check: METR 50% horizon ≥160 h (one work-month) by 2027-12-31.
- (2024-06) $1T/yr AI investment by 2027 (v1: "trillion-dollar cluster by 2027–28"). Held (tracking). Check: 2027 capex at Q4-2027 earnings (Feb 2028) ≥ $1T across hyperscalers plus neoclouds.
- (2024-06) 100 GW cluster, >20% of US electricity, ~2030 (v1: "~10% by 2030"). Drifted. Check: any single site ≥20 GW under construction by 2028-12-31; LBNL/EIA 2030 actuals in 2031.
- (2024-06) PRC steals US frontier weights. Drifted (distillation confirmed, theft not). Check: a lab or USG confirms weight exfiltration by 2027-12-31.
- (2024-06) Intelligence explosion 1–2 years post-AGI. Conditional. Per-branch check: if the AGI check passes, AI-research automation >10x the human rate by 2029-12-31.
- (2024-06) Algorithmic efficiency ~0.5–1 OOM/yr. Held (Epoch ~8-month halving; LessWrong one-year retrospective ~0.55 OOM/yr). Check: Epoch's next language-model estimate ≥0.4 OOM/yr through 2027.
- (2024-06) "The Project" — a USG-led AGI program by 2027/28. Open, tracking Wrong — the Genesis Mission EO (2025-11-24) is a DOE science platform with Manhattan-Project rhetoric, not a classified AGI program; no lab consolidation; export controls loosened (AI Diffusion Rule scrapped 2025-05; H200 sales to China at a 25% cut approved 2025-12 — secondary, via Dubach). Check: by 2028-12-31 a USG-run program with lab consolidation or nationalized compute; Genesis does not count.
- (2024-06) Open source falls behind. Wrong. Check: a proprietary-to-open gap ≤6 months at 2027-12-31 keeps it Wrong.
- (2026-07-24) "Best buying opportunity since early last year." Wrong on the stated horizon. (2026-09-02→10) Unlevered options re-entry — too early; the six names fell 5–8% on 2026-09-14 (Protos). Check: positions visible in the Q3 13F (~2026-11-16) and their mark by 2026-12-31.
Revealed behavior (Does)
- Money: ran the public book at up to ~4x gross leverage into a 56% memory concentration (13F, 2026-06-30) — the most concentrated expression of the physical-layer thesis on the bench, unwound in one month. Kept the private book (Anthropic — $3.5B per Fortune, $5B per Bloomberg; MatX; Fluidstack) when everything public went. First disclosed post-collapse move: crossing 20% of SharonAI (13D, event 2026-08-27, $523.9M cost) — hosting-layer conviction, unlevered. Sep 2–10: flex options on AMD, Bloom, CoreWeave, SK Hynix, SanDisk and the Roundhill Memory ETF (CNBC via Protos).
- Time: zero public output since 2024-06-14 (forourposterity.com last post; the Dwarkesh episode is 2024-06-04 with no sequel — checked this cycle). Twenty-seven months of letting filings speak, then two leaked letters in one week: July 24 asking for capital, ~July 30 "we let you down."
- Hires: Carl Shulman as director of research (co-reporting person on the SharonAI 13D); Graham Duncan as advisor (Fortune, 2025-10-08).
- Bought the DeepSeek selloff (Jan 2025) and reportedly stopped at least one large tech fund from dumping — "Leopold says it's fine" (Duncan, via Fortune).
- Blamed short sellers for exacerbating losses (Reuters, 2026-07-30), then took "full responsibility" a day later — both on the record.
Feels
Certainty plus a wound. The OpenAI firing and the FTX collapse ("incredibly tough") taught him institutions fail people; the fund was built so no institution could stop him, and July taught him leverage can. "We came closer to permanent capital impairment than is acceptable to us" is the most self-aware sentence he has published. Wants to be true: the OOM line is destiny and July was noise.
Hears
An EA-forecasting inner circle (Shulman; Cotra-adjacent timelines), his LPs (the Collisons; Gross and Friedman, both now inside Meta's superintelligence push; Jane Street), Anthropic through his wife's office, and the prime-broker and short-seller loop he now blames. No policy world since 2024; no press he answers.
Sees
Where the physical constraint binds next, before the market — power in 2024, memory in 2026 — plus frontier-lab economics from the inside (OpenAI) and the household (Anthropic). What he did not see: a 4x book in illiquid mid-caps is the constraint that binds first.
Incentive map
Re-derived this cycle: maximal and now asymmetric. A ~$10B fund with a third to a half in one private Anthropic position (IPO expected October 2026, TechCrunch); a spouse who is chief of staff to Anthropic's CEO (engaged by 2025-10, married 2026-08); LPs who include two Meta superintelligence executives; prime-broker access on probation (JPMorgan out, Morgan Stanley declined at launch, Clear Street in). He cannot say: Anthropic is overvalued; the intelligence explosion is slipping; memory and power are not the bottleneck; leverage was the thesis. The essay predated the fund and its technical half has scored well on independent cards — moderate discount there. On anything Anthropic-, memory- or timeline-related the discount is maximal. His letters are now the only channel for his views, and letters exist to keep capital in the fund. Rule-25 weight: coincidence discount on every forecast his book is long.
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 — and, two years on, no government has adopted it
- Bottleneck-cascade investing (own the binding constraint and the next one) — his method, generalized in Rich's inner-loops work; provenance in the how-he-thinks note
What he's reacting against
- "Scaling will plateau" skepticism — argues the trend lines are obvious to anyone who reads the curves (and they have kept reading his way: test-time compute, GPQA saturation)
- 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
- NEW (2026-07): short sellers "targeting the firm's positions" (Reuters) — the first market adversary he has named; the leverage that made them dangerous was his own
Where he overlaps / splits (with Rich)
- Overlaps with Amodei on aggressive timeline (~2027) and race-to-the-top framing; splits on who leads — Aschenbrenner says USG, Amodei says safety-focused labs. NEW: the split now sits inside one household and one portfolio — Anthropic is his largest holding and Amodei's chief of staff is his wife; treat any Aschenbrenner view of Anthropic as non-independent
- Overlaps with Altman on US-vs-China framing and trillion-dollar scale; splits on whether OpenAI's governance is adequate (left OpenAI 2024-04 amid security/safety disputes)
- Overlaps with Hassabis on capability timeline; splits on geopolitics — Hassabis less hawkish on US-China
- Splits sharply with Yudkowsky, Bengio on pause vs. build — building faster (with USG security) is his 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; NEW: both were long infrastructure scarcity into the July 2026 selloff — same trade, opposite politics
- Splits with Mostaque on open-source frontier — Aschenbrenner treats open frontier weights as catastrophic for security; on capability Mostaque's side has won so far (open-source call graded Wrong)
- Splits with Acemoglu, Autor, Brynjolfsson, Cowen on framing — labor questions are downstream of the AGI race; NEW: the Canaries payroll data (2026-08) shows entry-level displacement arriving before AGI, so "downstream" is already wrong in sequence
- Splits with Frey on adjustment horizon — his timelines collapse the multi-decade transition Frey describes
- NEW — Sutskever: Ilya says the age of scaling is over (2025-11); Aschenbrenner's OOM count includes algorithmic and test-time gains and has held even as the pretraining leg slowed — they argue about the mix, not the total
- NEW — Cotra: her own 2026 posts converge on his framing — "the intelligence explosion really could be this year" (2026-08-14), milestones landing 30–50% faster than she forecast — without naming him; her scored bias runs slow, his runs fast, and 2026 sits between them
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. Arc: ~$225M (2024, reported) → $1.5B (2025-10, Fortune) → ~$20B (2026-06, WSJ; 13F $20.24B) → −67% July → ~$10B (Reuters/Bloomberg). +439% net H1 2026; +80% YTD at end-July after the collapse (Fortune, FT)
- Independent scorecards (2026): technical and infrastructure calls ~all confirmed or on track, political calls ~all unfulfilled or reversed (Dubach, 2026-05); 3 on track / 1 wrong / 2 open / 2 pending (EA Forum, 2026-06-10). The physics half is aging well, the government half badly; the fund only ever expressed the physics half
- Short public track record — first major essay 27 months old; influence disproportionate to age and résumé length; now also the reported largest single-month dollar loss in hedge-fund history (FT via Investing.com, 2026-09-11)
- Pattern: confident, specific, falsifiable — easy to score by 2027–28; the 2026 additions (leverage; a capital call six days before a forced sale) show the same confidence applied to timing, where it has not paid
- 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 — the well-scored half
- Normative: USG must lead; free world must win; alignment effort must scale with capability; lab governance is inadequate at frontier — none adopted by any government two years on
- 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. 2026 update: the conflict now includes a ~$3.5–5B private Anthropic stake, a spouse in Anthropic's C-suite, and letters written to retain capital after a 67% month
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; the checks above pin a hard stop at 2028-12-31
- 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) — and has said nothing since Sutskever's own reversal
- "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 — and open-source proliferation is the branch that actually happened
- USG-led national AGI project is politically infeasible at the scale described — proposal is normatively coherent but operationally underdeveloped; reversed export controls and Gulf compute deals say the state went the other way
- 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 27 months old; no prior multi-year forecasting record; the fund record is one spectacular year and one catastrophic month
- Doesn't engage seriously with distributional / labor consequences — treats them as downstream of the AGI race, which assumes the race resolves on his timeline
- NEW: the fund is now his only voice, and a fund's letters are marketing. Until he publishes or speaks, every "view" attributed to him is an inference from a 45-day-stale holdings report
- NEW: reported figures conflict — peak AUM $45B (CNBC) vs ~$20–24B (WSJ, Reuters, FT, and the 13F itself); Anthropic stake $3.5B (Fortune) vs $5B (Bloomberg). This file trusts the ~$20B net figure (it matches the 13F and the FT-read letter; $45B is consistent with gross-of-leverage exposure) and carries the Anthropic range
- NEW: "short sellers did it" and "I take full responsibility" were said a day apart; which one he believes decides whether the lesson was learned
Rich's take
- (your synthesis here)
Delta log
2026-09-21 — v1→v2 migration + validation (bulk run)
- Grades: 7 Held / 3 Drifted / 3 Wrong (plus 1 unfalsifiable, 4 open or conditional with checks attached, 3 normative not graded).
- Held: OOM effective-compute rate (Epoch, 2024-03; EA Forum scorecard, 2026-06-10); college-graduate-level models 2025/26 and GW-class clusters by 2026 (same scorecard); $1T/yr AI investment by 2027, tracking (I/O Fund capex roundup, 2026); power as binding constraint, no-plateau / test-time-compute call and algorithmic efficiency (Dubach scorecard, 2026-05; LessWrong retrospective, 2025).
- Drifted: 100 GW / >20% electricity by 2030 (LBNL 2025 Update, 2026-06: 9.5–15.3% for all US data centers); PRC weight theft → API distillation (NSA/CISA/FBI AA26-251A, 2026-09-08); alignment "muddling through" (Anthropic alignment faking, 2024-12; Dubach).
- Wrong: open source falls behind (EA Forum scorecard, 2026-06-10); "best buying opportunity" capital call of 2026-07-24 (TechCrunch, 2026-07-30; Reuters via Yahoo, 2026-07-30); the leveraged July expression of the memory/infra-over-software thesis — both legs moved against him: SanDisk −47%, Micron −29%, software shorts squeezed (Reuters/CNBC via Yahoo, 2026-08-21).
- Most surprising delta: the bench's fund-as-validation persona lost 67% in a month at 4x leverage and sold its entire public book to Citadel (2026-07-30) — five weeks before this review, and two weeks before the --investing twin's 2026-08-15 read of the Q2 13F, which recorded the memory pivot but not the collapse. Second surprise: the largest remaining position (Anthropic, ~20% of the fund since ~2025-03, WSJ) never appeared in either bench file because 13Fs only show public stock; the spouse link to Anthropic's CEO office was public from 2025-10 (Fortune).
- Employer/role check (mandatory): no title change. Material events since the 2026-05-17 review: H1 +439% (FT-read letter via Disruption Banking, 2026-07-30); July −67% and the Citadel block; "we let you down" (Business Insider via AOL, 2026-07-31); marriage (Fortune, 2026-07-31); SharonAI 21.1% (Schedule 13D, 2026-08-28; 13D/A, Form 4, 2026-09-17); JPMorgan out, Clear Street in (FT via Investing.com, 2026-09-11); options re-entry (Protos citing CNBC, 2026-09-14). Incentive discount raised to maximal on Anthropic, memory and timeline claims.
- Numbers corrected against the primary (Racing to the Trillion-Dollar Cluster; From GPT-4 to AGI): "$1T cluster by 2027–28" → the $1T+ cluster is the ~2030 row and $1T is annual investment by 2027; "~10% of US electricity by 2030" → essay says ">20%" for the 100 GW cluster; "drop-in remote workers by 2025–26" → 2027, with 2025/26 being "smarter than most college graduates."
- Silence check: forourposterity.com last post 2024-06-14; Dwarkesh episode 2024-06-04, no sequel; a search-engine summary claiming an "August 2025" Dwarkesh appearance was checked against the episode page and rejected. v1's open [TBD]s on 2025–26 posts and post-2025 talks resolve to: none found.
- Sibling-run lead (Cotra converges on his framing): supported from her own posts (Hurtling through 2026, 2026-08-14; AI predictions for 2026, 2026-01-14); she does not name him — logged as convergence, not citation.
- Tier: quarterly — the Q3 13F (~2026-11-16, handled by the scheduled read; not attempted here), an expected Anthropic IPO, the AGI-2027 clock and a fund on prime-broker probation make each of the next four quarters gradeable.
- Confidence unchanged at medium: fund facts rest on filings and FT/WSJ/Reuters-chain reporting, but peak-AUM and Anthropic-stake figures conflict across outlets and his own words exist only as leaked letters. Web-search budget ran out at the end of this run; two secondary citations (export-control reversal; turbine lead times) are flagged in-line rather than replaced with primaries.
Sources
- Primary — his own: "Situational Awareness: The Decade Ahead" (2024-06-04) · Racing to the Trillion-Dollar Cluster (cluster table, $1T/yr by 2027) · From GPT-4 to AGI (OOMs, the 2027 claim) · Dwarkesh Patel interview (2024-06-04) · forourposterity.com (dormant since 2024-06-14) · investor letters 2026-07-24 and ~2026-07-30 as quoted by FT, Business Insider via AOL (2026-07-31) and HedgeFundAlpha (2026-07-31)
- Primary — filings (EDGAR CIK 0002045724): 13F-HR Q2 2026 cover (26 holdings, $20,242,292,228; filed 2026-08-14) · Schedule 13D — SharonAI Holdings, 21.1%, $523.9M cost (2026-08-28) · 13D/A No. 1 — warrant exercise (2026-09-17) · Form 4 as 10% owner (2026-09-17) · EDGAR submissions index · Q1 2026 put-book detail: bench note _aschenbrenner-deep-research-2026-07-19.md (EDGAR-verified)
- Fund events 2026: CNBC via Yahoo — forced unwind, Adobe short named (2026-07-30) · Reuters via Yahoo — Citadel buys the book, ~$10B remains (2026-07-30) · TechCrunch — Anthropic, MatX, Fluidstack kept; July 24 letter quote (2026-07-30) · SpotGamma — anatomy of the margin call, 4x gross (2026-07-30) · Disruption Banking — 439% H1 letter, Aug 1 capital raise (2026-07-30) · Fortune — wedding, Balwit's title, $3.5B Anthropic slice (2026-07-31) · Reuters/CNBC via Yahoo — Griffin unwinds 80%, $45B peak claim (2026-08-21) · FT via Investing.com — JPMorgan cuts lending, Clear Street, flex options (2026-09-11) · Protos — the six options names, citing CNBC (2026-09-14) · Fortune profile — $1.5B AUM, Shulman, Duncan, DeepSeek buy, engagement (2025-10-08) · Wikipedia (biography cross-check)
- Scorecards and graders: EA Forum two-year scorecard (Edison, 2026-06-10) · Dubach — Aschenbrenner's receipts (2026-05) · EA Forum evaluation (Harris, 2026-03) · LessWrong one-year retrospective (2025) · Lawfare — obstacles to AGI by 2027
- Market-level and technical data: METR Time Horizon 1.1 (2026-01-29) · METR live page (updated 2026-05-08) · Epoch — algorithmic progress in language models (~8-month halving) · I/O Fund — 2026–27 hyperscaler capex roundup · LBNL US Data Center Energy Usage Report: 2025 Update (2026-06) · NSA/CISA/FBI advisory AA26-251A (2026-09-08) · Genesis Mission EO (2025-11-24) · Alignment faking in large language models (2024-12) · CSIS — China's open-weight challenge · Cotra — Hurtling through 2026 (2026-08-14) · Cotra — AI predictions for 2026 (2026-01-14)
- Earlier coverage carried from v1: ChinaTalk (2024-06); Bari Weiss / Honestly (2024); New York Times profile (2024); Wired, The Information on the OpenAI departure (2024-04) — dates still unconfirmed; Forethought / academic output pre-2023 limited (Weak-to-Strong Generalization, ICML 2024)
- Bench siblings (context, not modified): --investing twin (reviewed 2026-08-15; predates the collapse) · how-he-thinks note · _aschenbrenner-13f-readsheet.html · _aschenbrenner-deep-research-2026-07-19.md
migrated v1→v2 2026-09-21 · backup: _archives/leopold-aschenbrenner.html.bak-20260921 · converts-from: people/leopold-aschenbrenner.md · AI & Society domain