Updated: 2026-11-16
What changed this week (10 – 16 Nov 2026, checked 2026-11-16)
- Two Moonshots episodes in the window, both sectioned below. EP310 (aired 11 Nov) dates the
singularity, "sometime around the summer of 2027" (
0:31:30–0:31:54), then calls the curve "an optical illusion" (0:46:16). EP311 (aired 13 Nov) makes the week's real move: for the first time he names sensing, not energy, as the bottleneck, "physical world sensing and physical world actuation… rapidly becomes the limiting factor, not computer intelligence" (1:43:14). EP311 also softens his EP310 line on Jensen Huang to "we may never have all the facts" (0:11:36–0:15:25). No new holding or tie disclosed in either. - Ledger: four rows added (19–22) from EP310: the summer-2027 asymptote (self-hedged), GDPX "saturated within 6 months", "major motion" on diseases in 1 to 1.5 years, and labs selling bundled robot and GPU hours in two to three years. Re-checked at source 2026-11-16: row 1, Clay still labels Navier–Stokes "Active", and its newest news item is still the 11 Sep statement ("The process is deliberately unhurried"). Row 5 was not re-checked this week. Nothing resolved: none held, none missed.
- Orbit as an escape from Earth politics is now a recurring line (EP307, EP309, EP310, and the 6 Nov newsletter), so it is added to Key phrases below with a translation.
- Chain Theory: still not on arXiv. arXiv API, 2026-11-16: the author query (
au:Wissner_Gross) returns 0 results; the phrase "Chain Theory" returns 122, and the ten newest are all Markov-chain papers, none his (a floor; listing pages not browsed). The CPT work he teased on EP308 has not appeared; alexwg.org's publication list still ends in 2024. No named physicist's assessment or journal record found. - PSI: no new research item since 6 Nov (psi.inc/research, read 2026-11-16). psi.inc still lists him "Chief Scientist and Chief of Strategy"; his own site's bio says "Founder and Chief Scientist of Physical Superintelligence (PSI)". Two self-descriptions, not a change found this week. PSI's pages still do not mention Chain Theory.
- Newsletter: two posts, the thinnest week since the late-October gap. "Welcome to November 10, 2026" (10 Nov) and "Welcome to November 12, 2026" (published 13 Nov, 01:27 GMT); nothing after that as of the check (archive, 2026-11-16). Both free; no price shown. No dated forecast in his own voice; the dated items are attributed (Todd's "goes vertical around August 2027", Christiano's "within 12 months"). The 12 Nov post closes on "the Dyson Swarm is going around it". Neither mentions Chain Theory, PSI, Eon, idealab, Ornn, Starcloud or Lonestar.
- Ornn: still not live. No ICE notice newer than the 2026-09-29 one ("no listing date") and no CFTC approval found.
- Eon, idealab, Starcloud, Lonestar: nothing new. Eon's newest own post is still 2026-09-11; it still calls him "Founding Advisor". idealab.ai still names no one beyond Bill Gross; "Coming soon in 2027" holds; newest newsroom item still 22 Oct. The Cloudflare credits line was not re-read this week (its page was not fetched). No news dated in the window for Starcloud or Lonestar.
- Appearances: no talk or podcast outside Moonshots with a verifiable date in the window.
- Open programme found through his newsletter (small): the 12 Nov post says "Claudestronomy is funding agent-led sky searches". The page (claudestronomy.com, read 2026-11-16) is more modest: an independent programme run by macrogrants, a nonprofit incorporated 10 Oct 2026, "seeded with $100,000", offering model and compute credits and "small grants", with no award size, no eligibility rules, no form and no deadline ("We're accepting project ideas now"). Not affiliated with any AI lab. He has no stated tie to it. Added to the opportunities register. The same post's "NASA is recruiting for a Moon Base" links to an X post that could not be read, so it is not described here.
- Discovered Materials Grand Challenges, re-checked 2026-11-16: all five still "Open", rules still "Last updated: November 5, 2026", still no deadline, none awarded.
Previous week (4 – 10 Nov 2026, checked 2026-11-10)
- ★ His own position moved, on tape, not in writing. EP309 (published 7 Nov) drops the condition he
set on EP308 four days earlier: "True RSI only occurs when robots or nanobots create better copies of
themselves" (EP308
0:53:15) became "we're so deep in rsi at this point… with just software only improvements" (EP3090:33:58). He did not flag the change. Both episodes are sectioned below. EP309 also adds a stake said mid-pitch, "star cloud and lonestar / We have a financial interest" (1:35:56), now in the holdings table. - Ledger: three rows added (16–18) from EP309: continuous frontier releases "early 2027 or so", von Neumann probes "within a few years", robot licensing with command logs "in a year or two". Re-checked at source 2026-11-10: row 1, Clay still labels Navier–Stokes "Active", newest news item still the 11 Sep statement; row 5, no merger item on Tesla's 3 Dec ballot (Musk-page check). Nothing resolved: none held, none missed.
- Chain Theory: still not on arXiv, and he has not written about it again. arXiv API, 2026-11-10:
author query, title words (fine, structure, qudits) and "Chain Theory" + fine-structure all
return 0 results (a floor; arXiv listing pages were not browsed). The CPT-symmetry work he teased on
EP308 ("I will speak through publication",
1:25:06) has not appeared on arXiv or on alexwg.org, whose publication list still ends in 2024. No named physicist's assessment or journal record found. - PSI published again, 6 Nov: "A proof of the Aaronson-Ambainis conjecture" (psi.inc/research), led by staff member David Turturean with advisers Jim Halverson and Phiala Shanahan, run on "a non-public Anthropic model". It does not name him. Company-stated; no independent check found. He is still listed "Chief Scientist and Chief of Strategy"; PSI's pages still do not mention Chain Theory.
- Newsletter: three posts in the window, "Welcome to November 6", "…9" and "…10, 2026", all free and fully readable (archive, checked 2026-11-10); a three-day gap 7–8 Nov. No dated forecast in his own voice; every dated item is attributed (Todd's "goes vertical around August 2027", Cotra, Brockman, Zipline/Uber). The 6 Nov post returns to orbit as escape from local politics ("the first faint outline of a Dyson Swarm that no ballot measure can reach"), the same line he ran on EP307 and EP309. None mentions Chain Theory, PSI, Eon, idealab, Ornn, Starcloud or Lonestar.
- Ornn: still not live. No CFTC decision and no ICE notice newer than the 2026-09-29 one ("no listing date") found; a law-firm alert still says no compute futures contract has final CFTC approval.
- Eon, idealab: unchanged. Eon's newest own post is still 2026-09-11. idealab.ai still names no team, advisers or investors beyond Bill Gross; the Cloudflare credit line, "Join the beta", no deadline, and "Coming soon in 2027" all hold; newest newsroom item 22 Oct.
- Appearances: no talk or podcast outside Moonshots with a verifiable date in the window.
- ★ Open programme found through his newsletter: the 9 Nov post links the Discovered Materials Grand Challenges, five materials-physics problems with a pool "up to US $1,000,000" (cuprate superconductivity $400K; quantum-chemistry speed and accuracy $200K; silicon/oxide interfaces $200K; laser-driven materials $100K; interfacial heat transport $100K). Run by Matforge, Inc.; open to adults, teams and companies; "The Competition has no deadline"; rules "Last updated: November 5, 2026" (rules). Every entry is licensed to the sponsor, win or lose. He has no stated tie to it. Added to the opportunities register.
Earlier (30 Oct – 4 Nov 2026, checked 2026-11-04)
- ★ Chain Theory: still not on arXiv, and the first independent audits exist — on GitHub, not in a
journal. The arXiv API author query returned only the 2007 paper and a 2013 comment (checked
2026-11-04; a direct title-word search was not run, so this is a floor), eight days after EP306's
"I would expect it to show up on the archive sometime soon" (
1:24:58). Two audit repos were found: SproutSeeds/chain-theory-audit (README © "Cody Mitchell", no affiliation) reports the arithmetic, the Lean file and the 172-record count reproduce, and that "the photon identification conflicts with the program's own charge assignment, unbroken generator and weak angle"; HiroSakuraba/Qutrit---Qubit-audit (pseudonymous; README v22 dated 27 Oct) calls the arithmetic "mathematically reproducible" and the identification with electromagnetic coupling "unestablished" / "conjectural". ⚠️ Neither is peer review, neither author's standing is verifiable, and the repos' publication dates were not visible to the fetcher, so whether they appeared inside this window is not established. Still no named physicist's assessment and no journal record found. PSI's pages still do not mention Chain Theory. The standing line holds: "a proposed, unreviewed derivation whose physical identification is his conjecture". - EP307 (aired 31 Oct, recorded ~30 Oct) is sectioned below. The headline: he dissents from a
sponsor's story on air and names it — "I'm exercising my editorial independence here and I'm calling
balls calling balls and strikes" (
0:15:00) — and he takes Altman's self-replicating data centres off-world ("Arguably they want to be orbital or even better lunar",1:04:32). One new dated prediction (search advertising becomes Google's second-largest business, ledger row 15). Small disclosures, none a holding: "I have friends on the team" (Google,0:16:18); he owns a Tesla Model Y and uses FSD (1:29:02); "not investment advice and it's not college major advice" (1:49:01). No new tie disclosed. - Newsletter cadence resumed after a nine-day gap. After "Welcome to October 21, 2026" and the 24 Oct Chain Theory post, nothing until "Welcome to Halloween 2026" (31 Oct), then "…November 1, 2026" and "…November 2, 2026" (subtitle: "The Singularity is now its own lead engineer"; archive, checked 2026-11-04). All three fully readable, no paid markers. None mentions Chain Theory, Ornn, PSI, Eon, idealab or a Dyson swarm, and none carries a dated forecast of his own. He has not written about Chain Theory again since the 24 Oct post. Nothing posted 3–4 Nov as of the check.
- Ornn: still not live. No CFTC action on compute derivatives found since the comment period closed 2026-10-20, and no ICE notice newer than the 2026-09-29 one (HPR/B200 contracts, "no listing date") found. Unchanged.
- Ledger. Row 1: Clay still labels Navier–Stokes "Active", no award, no mention of the OpenAI claim (read at source 2026-11-04). Row 4: no Figure–Hark deal found. Row 5: no step past the 20 Oct merger committees; Tesla's annual meeting is reported for 3 Dec 2026 (secondary; the December announcement remains an analyst's guess). Row 3 (GPT-6.x humanoid, due ~10 Dec): no resolution; the week's adjacent fact is Altman pointing OpenAI's robotics programme at data centres, not humanoid tasks (EP307 clip). Row 15 added (search ads). Nothing resolved: none held, none missed.
- Holdings: all quiet. PSI still lists him "Chief Scientist and Chief of Strategy"; nothing on its pages dated after 23 Oct (psi.inc, checked 2026-11-04). Eon: newest own post still the 2026-09-11 "Perspective"; the only newer dated item is third-party coverage (8 Oct). idealab.ai: still does not name him or any team, advisers or investors beyond Bill Gross; the Cloudflare "$100,000 in free credits" line is still up with no claim route or deadline; "Coming soon in 2027" holds. The existing register row stays as is.
- Appearances: no talk or podcast outside Moonshots found with a verifiable date in the window.
- Open programme: none new. Nothing added to the pipeline.
Earlier still (23 – 29 Oct 2026, checked 2026-10-29)
- ★ Chain Theory is in writing, and the writing is more careful than the tape. One post in the window, "Chain Theory" (24 Oct, free, Innermost Loop), links a preprint, "A Parameter-Free Candidate for the Fine-Structure Constant from Entangled Qudits" (A. D. Wissner-Gross, "Dated: October 23, 2026", PDF), a compendium marked "A Compendium of Preliminary Findings · WORKING DRAFT", also dated 23 Oct (PDF), and a Lean file. The candidate is "(79 + 48√2 + 42√3 + 3√6)/31104", inverse "137.035 999 189 566…"; the post compares it with CODATA 2022's "137.035 999 177, with a standard deviation of 21 in the last two digits" and gives "0.60 standard deviations". We re-did the arithmetic by hand (1/α ≈ 137.03599919; (189.6 − 177)/21 ≈ 0.60σ): it checks. What the writing fences that the tape did not: "only its identification with α(0) is conjectured", "this comparison is not yet decisive" because the measurements "still disagree with one another", and "much of it is still conditional or underdeveloped". The experimental test is named (Northwestern electron magnetic moment, Berkeley and Paris atom recoil, "up to ten times more precise") with no date; the "next few months… next couple of years" window is from EP306 only. ⚠️ Not on arXiv as of 2026-10-29 (arXiv API author query: 0 results; the hyphenated-name form was not tried, so this is a floor). No peer review, no independent physicist's assessment found; coverage so far is secondary (e.g. FourWeekMBA). The post names no company and makes no disclosure; it does not mention AI use, though EP306 said it "requires an enormous amount of steering". Say "a proposed, unreviewed derivation whose physical identification is his conjecture", never "derived the fine-structure constant".
- PSI published, 23 Oct. The company's research page has its first dated item, "Five results on Open
Quantum Problems" (23 Oct 2026), each with "a paper, a Lean formalization, and verification code"
(psi.inc · repo
physical-superintelligence/oqp-batch-2026, not opened). Company-stated; not independently checked. He is still listed as "Chief Scientist and Chief of Strategy". PSI's pages do not mention Chain Theory. No funding news. Its only door is still a jobs link, so no programme row. - Newsletter cadence broke. After "Welcome to October 21, 2026" the only post through 29 Oct is Chain Theory: no "Welcome to" post for eight days (22–29 Oct), longer than any gap logged on this page (the previous longest was 13 → 18 Oct). Price: still none stated; the publication shows "payments_state":"disabled" in Substack's own archive data, so it is free.
- Ornn: still not live. The CFTC's request for comment on compute derivatives closed on 2026-10-20 (CFTC release 9286-26). No CFTC decision on the ICE/Ornn HPR and BKL contracts and no ICE notice after 2026-09-29 found (ICE's notices index returned 404, so this is a floor). The only extension found concerns a different product (NYMEX Silicon Data B200, review to 9 Nov).
- idealab.ai: unchanged. Still does not name him, nor any team, advisers or investors. The Cloudflare line is unchanged; the page now also says "Coming soon in 2027" and gives no claim route or deadline, so the existing register row stays as is. Newsroom: two items, both before the window (Bill Gross at Cloudflare Connect 20 Oct and Masters of Scale 21 Oct).
- Eon Systems: nothing new. Newest blog item still the 2026-09-11 "Perspective". A 10 Mar media item on its page calls him "Eon's founding advisor"; he has said "helped found" on tape (EP291). Both stand, as stated.
- Ledger. Row 1: Clay still lists Navier–Stokes "Active"; its newest news item is still the 11 Sep statement ("apparently been settled… deliberately unhurried"), no award. Row 4: no Figure–Hark deal found. Row 5: no merger agreement; per this week's Musk-page check, no Tesla proxy filing and the SpaceX 8-K of 26 Oct is silent on Tesla. Two rows added: 13 (the XCorp name, EP305) and 14 (Chain Theory tests, EP306). Nothing resolved.
- Moonshots: EP305 (23 Oct) and EP306 (28 Oct) are already sectioned below.
- Appearances: no talk or podcast outside Moonshots found with a verifiable date.
- Open programme: none new. Nothing added to the pipeline.
Oldest kept (17 – 23 Oct 2026, checked 2026-10-23)
- ★ Ledger row 5 moved, not resolved: SpaceX and Tesla have formed merger committees. On 2026-10-20, during SpaceX's earnings webcast, Musk said the two companies "have formed special committees to explore a potential merger"; no terms are agreed, he is not in the talks, and a merger "is not certain" and needs Tesla shareholder approval (Bloomberg Law, 20 Oct 2026). His EP284 call was "merge sometime in the next year or two" (due 2027-08 to 2028-08). Row 5 stays open: committees are a step, not a merger. ⚠️ A December announcement is an analyst's guess (Barclays, as reported), not a company date.
- ★ New holding on the table: idealab.ai. On EP304 he disclosed "a financial interest in the
venture behind this" (
1:30:46). Bill Gross launched Idealab.ai around 2026-10-19 (Axios, 19 Oct, read only via secondary write-ups; the Axios page returned 403). The site, read in full, offers a free database of about 1.2 million business ideas and says "the public ideas are free to explore and use"; it names no team, advisers or investors beyond Bill Gross, and does not name him (idealab.ai · about). So his stake is on tape only, size and form not public. Added to the holdings table. - Open door found, and it is the venture he holds a stake in. Idealab.ai's Cloudflare partnership page says Cloudflare is offering "$100,000 in free credits to anyone who builds their company on our platform"; how to claim and any deadline are not stated, and the banner figure carries an asterisk (elements-of-business). Added to the submissions register as a rolling credit offer, with his interest noted beside it.
- Newsletter cadence holds, price still unstated. Two posts in the window: "Welcome to October 18, 2026" (bylined 19 Oct) and "…October 21, 2026" (archive). Both were fully readable with only a "Subscribe for free" prompt. Neither mentions Ornn, Eon, PSI, idealab, orbital compute or a Dyson swarm. The 21 Oct post carries the merger news in his own framing ("a $1.9T plus $1.48T combination, the largest in history") and SpaceX's quarter; the dated forecasts in both posts are other people's (Jack Clark, Musk), not his, so no ledger row.
- Ornn: unchanged, still not live. Latest primary item is still the ICE Futures U.S. notice of 2026-09-29 (HPR and BKL, pending the CFTC comment period, no listing date). No October notice or CFTC action found.
- Eon Systems: a correction, nothing new in the window. The company blog shows a 2026-09-11 "Perspective" post after the March announcements (updates), so "nothing since March" on earlier weeks was wrong by one post. No research or funding news in October.
- PSI: nothing new since the 1 Sep launch.
- Ledger re-check (rows 1, 4). Clay Institute still labels Navier–Stokes "Active" with no mention of an award (Clay). No Figure–Hark acquisition found.
- Moonshots: the feed watch of 2026-10-23 07:00 scanned 25 uploads and found no panel episode we do not hold (106 on disk). EP304 (aired 21 Oct) is already on this page.
- Appearances: no new talk or podcast outside Moonshots found with a verifiable date.
What he is here for: the show's own episode descriptions label him "Dr. Alexander
Wissner-Gross is a computer scientist and founder of Reified" (the same line in every description
from EP204, 2025-10-30, to EP295, 2026-09-28). On air Peter introduces him as "our in-house ASI"
(EP284 0:01:44, EP288 0:01:43) and earlier as "our resident genius" (EP253 0:03:20). On the
panel he is the one who says whether a claim has been verified.
Who he is
From his own CV at alexwg.org (undated as posted), the publishers' records and his public talks. Checked 2026-09-29.
Education. Three S.B. degrees from MIT, in Physics, Electrical Science and Engineering, and Mathematics (1999–2003). An A.M. and a Ph.D. in Physics from Harvard (2003–2007), as a Hertz Fellow. His CV says the thesis won the Hertz Foundation's Doctoral Thesis Prize.
Research record (publisher metadata via Crossref):
- "Causal Entropic Forces", A. D. Wissner-Gross and C. E. Freer, Physical Review Letters 110, 168702 (2013-04-19), doi:10.1103/PhysRevLett.110.168702. The paper behind his TED talk. In plain terms: a system that acts to keep as many future options open as possible can look intelligent without being told what to do.
- "Relativistic statistical arbitrage", with C. E. Freer, Physical Review E 82, 056104 (2010), doi:10.1103/PhysRevE.82.056104. Where to site trading servers, given the speed of light.
- "Diamond stabilization of ice multilayers at human body temperature", with Efthimios Kaxiras, Physical Review E 76, 020501 (2007), doi:10.1103/PhysRevE.76.020501.
- "Dielectrophoretic reconfiguration of nanowire interconnects", Nanotechnology 17, 4986–4990 (2006), doi:10.1088/0957-4484/17/19/035. This was his doctoral field: matter whose wiring can be rewritten.
Companies. From his CV: Reified, LLC, founder and managing partner since 2019 ("startup
investing and advising"). Gemedy, Inc., founder, president and chief scientist since 2011
("intelligent government systems"). Enernetics, Inc., founder 2007–2016 (web analytics;
acquired by Sustainable Travel International in 2016). Maxtile Holdings, founder and partner
2006–2011. On tape: Eon Systems is "a company that I helped found" (EP291 2:17:01), and he
co-founded PSI (EP285).
Advisory roles. Ornn (captioned "Orin"), the compute-price index and futures company:
"for avoidance of doubt I have a financial interest in I'm an adviser to the company"
(Moonshots #266 2:16:28). Peter restated it on EP288: "It's a link ventures portfolio company.
You know, Dave and my AI venture fund and Alex advises" (1:10:55). Peter also introduced him on
EP253 as "another member of the Link Exponential Ventures team" (0:03:20) without naming the role.
His CV lists advisory boards at the Brain Preservation Foundation (2018–), the Organ Preservation
Alliance (2016–) and the Lifeboat Foundation's scientific advisory board (2015–), and a Defense
Science Board task-force observer seat (2016–2017).
Academic roles (CV). Institute Fellow and Associate, Harvard Institute for Applied Computational Science (2012–2021). Research Affiliate, MIT Media Lab (2010–2019). MIT Alumni Expert in Residence (2017–).
Writing. Solve Everything: Achieving Abundance by 2035, with Peter Diamandis, was published online on 2026-02-14 at solveeverything.org. It is free to read, self-published, with no publisher and no ISBN. His own post calls it "a book-length blueprint" (Innermost Loop, 2026-02-14), and Peter's post calls it a "White Paper". So it is an essay-length book online, not a published book. His CV also lists essays in John Brockman's Edge anthologies, among them The Last Unknowns (2019).
Where he appears beyond Moonshots (YouTube upload dates):
| date | where | title |
|---|---|---|
| 2013-12-05 | TEDxBeaconStreet (talk given Nov 2013) | An equation for intelligence, also on TED.com |
| 2026-01-27 | Imagination in Action, with Peter Danenberg | Are We Treating AI Unethically? |
| 2026-02-02 | TEDxBoston | Why the Future of Intelligence Is Already Here |
| 2026-02-19 | Family Office TV | Triple-Major Physicist Dr. Alexander Wissner-Gross |
| 2026-03-31 | How I Invest Podcast | We're Already Inside the Singularity |
| 2026-04-27 | Milk Road Stocks | By 2035, Most Grand Challenges Facing Humanity Will Be Solved |
| 2026-06-07 | AI² Club, University of Florida (live) | Solve Everything: A Live Conversation |
| 2026-07-14 | Imagination in Action | The Next 10 Years Will Change Everything |
| 2026-07-17 | TEDx | The Next 10 Years of AI Will Change Everything |
| 2026-09-25 | Moonshots LIVE, Los Angeles (on stage) | see Peter Diamandis |
The Innermost Loop, and the book
- What it is. His newsletter, "High-velocity intelligence from the event horizon. Curated by Dr.
Alex Wissner-Gross" (about page). He calls it daily
("I cover it daily in my daily newsletter, the innermost loop", EP248
0:16:22). In the archive for September 2026, a post titled "Welcome to" goes out every one to three days (archive). There is also an audio version (podcast). - Free? The posts we opened were not paywalled. The subscribe page does not state a price.
- The Friday edition. ⚠️ Not found. The September archive shows no separate Friday format. Posts dated Friday 18 and Friday 25 September do not appear.
- The book. See Writing above: Solve Everything (2026-02-14), self-published online and free.
- Ornn. Adviser with a financial interest, disclosed on tape. How it has done is in the ledger below.
Why he is in this archive
Gate: present in all five episodes read. He is first responder on more segments than any other mate on EP288 and EP293.
Positions and disclosures
| Role | computer scientist; founder of Reified (episode description); "our in-house ASI / resident genius" |
| Publishes | The Innermost Loop, a newsletter on Substack (theinnermostloop.substack.com); see The Innermost Loop, and the book |
| Books | Solve Everything: Achieving Abundance by 2035, with Peter Diamandis: self-published online, 2026-02-14, free, no publisher or ISBN (solveeverything.org) |
| Interests declared | "for avoidance of doubt I have a financial interest in / I'm an adviser to" Orin (EP266); a financial interest in Henry (EP255); co-founder of PSI and adviser to Architect Labs (EP285); "as a shareholder of the entire market via low-cost broad index funds"; ★ ventures working on compute energy-efficiency and breakthrough energy sources (EP289 0:54:02) — his first energy interest on tape |
| Also | Boston; Harvard and MIT (see Who he is) |
Where his information comes from
He is the reader. He reads the fine print nobody else does — "its metric for success is the share of questions answered correctly on all five attempts… I had to check that" (on the Artificial Analysis benchmark, EP283) — and he labels the status of what he says: "this has been publicly reported", "based on public reporting", "I would infer", "this is a prediction", "not financial advice" (six times in one episode).
Main ideas
- The innermost loop. Recursive self-improvement is the thing everything else is downstream of. "We hit AGI no later than summer of 2020… We already have ASI."
- Per-token economics decide compute. "The highest profit-generating tokens will ultimately receive the most compute." American labs are revenue-maxing; Chinese labs are not.
- Abstraction layers are the villain. "People one or two layers up from you in the stack expect the old things." Co-scaling.
- LLM-pilled America vs world-model-pilled China.
- Space is the release valve and the leading indicator. "The singularity is going to be visible first in space, not on Earth." "The petafab is the killer app for the moon." Dyson swarms, sun-synchronous rings. "News flash: there's no water in low Earth orbit."
- High agency is the trait the era rewards. "Society will reward agency."
- "Every sci-fi trope everywhere all at once" — his operational definition of the singularity.
- The AI sniff test. "If there's no AI in it, to first order, what's the point?"
What is a Dyson swarm
A Dyson swarm is a very large number of separate solar collectors, each in its own orbit around a star, that together capture a large share of the star's energy. The idea is usually credited to Freeman Dyson's one-page paper in Science, "Search for Artificial Stellar Sources of Infrared Radiation" (vol. 131, pp. 1667–1668, 3 June 1960, doi:10.1126/science.131.3414.1667). He argued that an advanced civilisation would eventually use most of its star's output, and could be found by the infrared waste heat it gives off. Olaf Stapledon had described the same picture in his 1937 novel Star Maker, where every solar system is "surrounded by a gauze of light-traps", and Dyson later said "Stapledon sphere" would be the fairer name. It is not a shell. Dyson's own clarification: "A solid shell or ring surrounding a star is mechanically impossible. The form of 'biosphere' which I envisaged consists of a loose collection or swarm of objects traveling on independent orbits around the star." (⚠️ quoted via Wikipedia, which cites Dyson's 1960 correspondence in Science. We did not read that primary.)
Why it is a system of systems. Each collector is a complete system: power, thermal control, pointing, station-keeping, communications. The swarm only works as a whole. That is the case Mark Maier's criteria describe ("Architecting principles for systems-of-systems", Systems Engineering 1(4):267–284, 1998, doi): each element can run on its own, each can be managed on its own, the useful behaviour emerges from the whole, and it is built up over time rather than delivered finished. In practice that means the design problem is the interfaces and the traffic rules, not any one satellite.
The self-replicating-manufacturing idea attached to it. Nobody launches a swarm like that from Earth. The usual answer is factories in space that make copies of themselves from local material, then turn a planet or moon into collectors. Alex holds both halves:
- "We disassemble Mercury because it's in a really convenient close to the sun orbit… We turn it
into a flying swarm of lasers that absorb sunlight" (EP278
2:07:08–2:07:35). - "a native industrial ecology that includes mining, includes manufacturing, most ideally
self-replicating von Neumann probes, basically a fab lab or a machine shop on the moon that is able
to make copies of itself using only native resources" (EP293
0:21:14).
How he uses the word now. For him, "Dyson swarm" mostly means orbital compute, today's first
step: "some of the compute may run on his Dyson swarm. I think there will be other Dyson swarms"
(EP289 0:55:08). The same move turns Mercury into collectors in the quote above.
What an operator can take from it. The near-term version is a supply chain, not a megastructure: solar arrays, radiators, radiation-tolerant computers, inter-satellite links, launch and in-orbit servicing. A small company can sell into one subsystem and one interface. It can build a Reverse bill of materials for a single collector, and price what a kilowatt in orbit costs against what a kilowatt on the ground costs. It can use system-of-systems thinking to decide which interface to own. The megastructure is the story; the subsystem is the business.
His related terms, in plain English:
- Sun-synchronous orbit (SSO). A near-polar orbit whose plane turns once a year, so a satellite
crosses each latitude at the same local solar time. A dawn-dusk SSO sits in almost constant sunlight,
which is why he places orbital data centres there. His use: "news flash, there's no water in low
Earth orbit… This is only forcing all of these new data center deployments to sun-synchronous
orbit" (Moonshots #283
1:37:16) · "the default outcome is the data centers are going to sun-synchronous orbit anyway where no one can credibly complain about their water usage" (EP2621:30:39). - Space as the release valve and the leading indicator. Plain English: the limits AI hits on Earth (water, permits, land, power) are pushing compute into orbit, so space is where the build-out shows up first. The two phrases are this page's summary, not his words. The lines on tape are the two SSO quotes above. "The singularity is going to be visible first in space" is attributed to him on Moonshots #255, whose transcript we do not hold, so it carries no timestamp.
- Orbit as the escape from Earth's politics. Plain English: he argues compute should move to the
Moon and orbit because rules there are lighter and slower to change than local votes and election
cycles. The legal premise is weak: the Outer Space Treaty makes states answer for their companies in
space, and the FCC, ITU and FAA still license what launches (see Orbital compute (data centres in space)). His use: "under
the outer space treaty and under the artemis accord there is a lighter touch / That is less prone to
two-year election cycles" (EP309
0:22:51–0:23:01) · "the preview of what a post-s singular world looks like is increasingly in space… for political and other reasons" (EP3101:59:52–2:00:03) · in writing, "the first faint outline of a Dyson Swarm that no ballot measure can reach" (Innermost Loop, 6 Nov 2026). First outing: the Moon as the release valve on EP307 (section below). - Recursive self-improvement (RSI). AI doing the research that improves the next AI, so each
generation speeds up the next. His use: "approximately in the next 3 to 12 months, depending on
uncertainty, AI is just completely leading all of its own R&D. And and that is total recursive
self-improvement" (Tenev episode
1:51:42–1:52:01) · "we're in the middle of recursive self improvement" (EP2931:18:21). - The innermost loop. The feedback loop at the centre of the economy: chips, energy, robots and
data centres, each making the others cheaper and faster. It is also his newsletter's name. His use:
"the innermost loop of civilization is recursively accelerating" (EP201
1:21:12), and the definition, "the innermost loop of civilization looking like some linear combination of chips, robots, data centers, power sources" (EP2051:08:25). - The vaudeville era of the space economy. Today's stage, in which companies take what they already do
on the ground and launch it unchanged, before anything is designed for orbit first. His use: "this is
the vaudeville era of the space economy where you just take everything that you were doing on land, lift
it up" (EP305
0:26:44). For an operator: the first wave sells ground hardware adapted for space; the later wave is products that only make sense in orbit.
What he's saying, in plain terms
His densest recurring lines. Each has the verbatim quote, then a two-sentence translation.
- "the innermost loop of civilization is recursively accelerating" (EP201
1:21:12). Plain: chips, power, robots and data centres now feed each other, so progress in one speeds up the rest. The thing to watch is that loop, not any single product. - "Bulk solving a field seems to want to become a database of all the precomputed answers to all the
questions that can be asked in that field." (EP288
2:17:46) Plain: once AI can answer one question in a field, someone answers all of them in advance and sells the table. AlphaFold's protein database is the pattern. - "Think of yourself as an intelligence company whose products happen to have atoms in it" (EP289
0:34:38). Plain: the value in a hardware business is moving to the model that runs it. Price and build the company as if the software were the product. - "It's an inductive prior. It's not a strategy for safety." (EP288
0:26:19) Plain: his argument that catastrophe is unlikely is a reason to believe, not a plan to prevent it. He keeps the two apart, and so should anyone who quotes him. - "all of this key technology around nuclear energy is now born secret. It it's not born non-secret,
as is the case with AI" (Tenev episode
0:12:15–0:12:43). Plain: nuclear was classified at birth by law, so the government owned its future, and AI was not. He argues that is why AI moved fast and nuclear stalled. - "a coalition of dominant firms that invokes safety to justify coordinated limits on the pace or
terms of competition" (EP291
0:12:26, his definition of "safety cartel"). Plain: when the leading labs ask for rules, check whether the rules slow down their competitors. It is a question about incentives, not a claim about any one firm. - "call it a few years, at most, 10 years maximum. And then after that, a human-machine merger in
order for human inputs to remain economically relevant" (EP293
0:45:26). Plain: he thinks people can direct AI for a few years at most, after which staying economically useful means merging with it. It is his dated ceiling, and it is on the ledger below. - "normaly [normalcy] overhang… If you look out on the street right now, you don't see it… It's
business as usual" (EP288
0:50:10–0:51:00). Plain: daily life looks unchanged while the capability underneath has already moved. He expects the gap to close suddenly, not gradually. - "competition and defensive co-scaling" (EP304
1:05:00) · "We will police it via defensive co-scaling" (EP3050:50:50). Plain: his answer to AI safety is more capable AI on the defending side, scaled as fast as the attacking side, not limits on the attackers. It is a policy position, and it assumes defence keeps pace. - "slop drop" (2026-10-15 episode
0:43:36, for OpenAI's 722 results) · "terrorizing them with slop drops" (EP3052:12:21) · "slop drop… version 2.0" (EP3061:51:59). Plain: a lab releasing a large batch of AI-generated results at once, faster than anyone can check them. He uses it against the releaser, so it is a criticism, not a measure of quality. - "math is cooked" (2026-10-15 episode
0:42:28) · "computer science has cooked, math is cooked", with only skills "that touch on the physical world" secure for "the next 2 to 3 years" (EP3071:49:12). Plain: he thinks AI has effectively finished the purely symbolic fields, so the only human work with a shelf life touches atoms. "Cooked" is his verdict word, and it is always ahead of the verification — the 10-15 episode's own caveat (not every proof formalised) still applies.
Nuclear, in context
- What he says. "the Atomic Energy Commission arguably completely fumbled the civilian
applications of nuclear energy… a fumbled handoff from the World War II era to a postw[ar]… civilian
era" (Tenev episode
0:12:05–0:12:15). "arguably we lost at least half a century of progress in vision [fission] and nuclear energy" (EP2910:45:22). "a moral panic over nuclear fision that led to overuse of, say, coal, which results in many, many unnecessary deaths" (EP2911:34:53). And forward: "in the next 10 years we're going to be well into our postfusion era", with AI as the inventor (EP2890:54:38). - The history he points to, dated. The Atomic Energy Act of 1946, signed 1 August 1946, created the Atomic Energy Commission and declared nuclear information "Restricted Data" until declassified, which is the "born secret" rule. The Atomic Energy Act of 1954 let the AEC declassify whole categories and opened the door to civilian power. The Energy Reorganization Act of 1974 abolished the AEC, which ceased on 19 January 1975, and split it into the NRC (regulation) and ERDA (promotion). Sources: Born secret · AEC. "Half a century lost" is his estimate, not a measured figure.
- Why it matters for space now. NASA is funding fission again: the SR-1 Freedom nuclear-electric demonstrator targeted for late 2028, and the lunar surface reactor procurement (SPARC). See NASA, Pillar 2: nuclear propulsion and power.
Where he takes the opposite side from Dave Blundin
- Universities. Dave: "I think universities are critical as the ethical actor in AI… I believe
they're here to stay." Alex, in the same block: "I don't think I beat up on universities enough…
I'd love to vivisect them" (EP293
0:37:10–0:37:20). See Dave Blundin. - p(doom). Same episode, opposite frames. Alex: a high p(doom) would mean "we would just be awash
with probably von Neumann probes", so safety "may be overrated" (EP288
0:25:30, fenced at0:26:19). Dave: "a 1%, .1% or 20% p-doom is totally unacceptable" (EP2880:42:52). - Regulation. Alex: "the risk at this point is overregulation, not under regulation" and, on
incident response, "I'd say hands off" (EP289
0:25:15,0:26:29). Dave, on other tapes, proposes tracking GPU clusters like plutonium (EP2880:43:46) and says "if you don't have regulation in the industry, all money gets stolen" (Tenev episode0:18:02). Different episodes, so this is not a face-to-face disagreement. - Photonics. Dave is "100% sure" the compute substrate moves to photonics (EP289
1:15:01). Alex treats it as conditional: "absent a successor to CMOS" (EP2851:06:28). - Longevity. We found no on-tape disagreement between them. Alex's "LEV is already here but
spiky" (EP288
2:12:14) went unanswered by Dave. - A tension worth holding. Alex tells listeners universities should be taken apart, and he is
among the most credentialed products of them: three MIT degrees and a Harvard Ph.D. (his CV). He also
credits MIT for the MIT.nano origin story (EP291
1:54:21). Dave, arguing to keep them, recruits from them (Link Studio sits by MIT). Both positions come from inside the institution.
★ EP289 (the AMA, 2026-09-13) — the episode where he disclosed most and hedged least
He is the panelist who discloses least. On the AMA he disclosed an entire sector for the first time, and he did it inside the answer to a question about that sector.
1 · The energy disclosure. Asked whether AI-designed chips make energy stop being the
bottleneck, he answers yes — then: "AI in including at ventures that I'm affiliated with is
absolutely being used to optimize both the relative efficiency number of say kilowatt hours per
gigaflop on the one hand and on the other hand maximizing breakthrough energy sources"
(0:54:02). Energy is not on his ledger anywhere else in 86 indexed transcripts. → Moonshots: the disclosures ledger.
2 · "Postfusion" — a coinage, and a dated forecast. "I I would fully expect in the next 10
years we're going to be well into our postfusion era and we're going to have postfusion energy
sources as well" (0:54:35), and the mechanism is AI, not space resources: "I think it'll be due
to AI helping us to invent postfusion energy sources" (0:54:46). Zero occurrences of
"postfusion" or "beyond fusion" in the other 85 indexed transcripts. Fusion had always been the
listed endpoint ("we talk about vision [fission], fusion, fossil fuels, solar, wind. I don't think
that's the end of the story"). → prediction log: due 2036-09-08.
3 · Dyson swarms, decoupled from Elon. "independent independently of Elon although some of the
compute may run on his Dyson swarm. I think there will be other Dyson swarms" (0:54:58). His
standing space thesis now has more than one owner in it.
4 · He overturns "stochastic parrots" — at Peter, not at the caller. "Also may maybe Peter just
take the bait once and for all. This is a misconception. uh reasoning models at at this point I I
don't think it's accurate to characterize them as prediction machines or stochastic parrots. They've
been so thoroughly reinforcement fine-tuned off of reasoning challenges to say that they're
prediction machines is tantamount to to saying basically that they're incapable of novelty or
insight which is most definitively not true at this point." (0:09:46) "Take the bait once and
for all" implies Peter has let the framing stand before.
5 · Bearish on the architecture, unboundedly bullish on the objective — in one answer.
"Transformers, as I've mentioned on the pod previously, are being ship of thesis [Theseus] style,
replaced bit by bit… in that sense, we're we're already up some sort of sigmoid curve" (1:13:29),
against "the self-supervised task of next word, next token, next bit prediction, I I think that has
effectively infinite longevity to it. You can scale that task… from here to solving everything"
(1:12:56). Read together, this is the precise version of the "slowdown" he is accused of both
sides of: the S-curve is the transformer's, not prediction's.
6 · Regulation, stated harder than on any prior tape. "the risk at this point is overregulation,
not under regulation"; "I am I'm deathly concerned that we will overregulate our way into losing a
super intelligence race with China" (0:25:15); and on incident response, "I'd say hands off"
(0:26:29). ⚠️ In the same segment Dave holds up Alvin Wang Graylin as the model for entering
government on this exact topic — and Graylin's published position is that the race is a myth. Alex
does not engage it.
7 · New citations and labels. AIXI (captioned AXY) — "a class of algorithms called AXY for
building superbasian [superbayesian] intelligences that are trained off of no human inductive prior
at all" (1:03:35) · Navier–Stokes as of that morning — "Look no further than obviously
today's headlines" (1:03:03, recorded 2026-09-08; ⚠️ the proof is claimed, announced and
contested — never say "AI solved Navier–Stokes") · the frozen-cutoff test already exists —
"there are prediction market and other prediction benchmarks… So, we're already doing that as a
civilization" (1:04:19) · neuroscience in five years, "and I think that's a conservative
outerbound" (0:19:14) · the disability→augmentation regulatory ladder and an MIT programme he
calls "H2O for human" (0:51:32) · he was interviewed for The Last Invention and dissents from
its title — "I don't think super intelligence is at least human civilizations broadly construed
last invention" (1:21:51).
8 · The line to keep. To an 18-year-old hardware founder: "I wouldn't think of yourself uh or
limit yourself to just think of yourself as a hardware company. Think of yourself as an intelligence
company whose products happen to have atoms in it and that mindset will be critical." (0:34:27)
★ EP288 (2026-09-11, recorded 09-10) — one coinage, two dated predictions, one bad premise
He carries more of this episode than any other mate — he is first responder on six of twelve segments — and it is the episode where his hedging discipline slips once, badly.
1 · New coinage: "normalcy overhang" (captioned normaly overhang). "I tried to coin the term
normaly overhang… If you look out on the street right now, you don't see it. You don't see humanoid
robots. You don't see nanotech swarms overwhelming everything. It's business as usual. Things look
like what passes for normal. But if you look at what's going on with these mathematical grand
challenges… we're so deep into the singularity and this is an overhang type situation that I think
is about to collapse." (0:50:10–0:51:05) Zero prior occurrences in the indexed corpus.
His.
2 · The Fermi prior against p(doom) — and he fences it himself. "I think we have ample evidence
that if we lived in a universe where p-doom were anywhere close to 10% that we would just be awash
with probably von Neumann probes disassembling earth from alien civilizations millions or billions
of years ago… Milky Way would have been gone already." (0:25:30) Peter pushes back
("I'm not sure I buy that as a good enough safety net") and Alex's correction is the part to quote:
"Well, it's not a safety net. It's an inductive prior. It's not a strategy for safety. It's an
argument that safety may be overrated." (0:26:15) Never quote the argument without the fence.
3 · "Alignment is just capabilities" — now a standing panel disagreement, not a one-off.
Stated three times with a worked example: instruction tuning "resulted in the equivalent of a 10,000
times larger model" (0:38:13); "Alignment has been capabilities all along" (0:38:45); and the
political payload — "Frontier Labs, if you want public money to pay for alignment, that's basically
a public subsidy for capabilities. Have at it."* (0:38:45). Emad Mostaque rejects it
flatly on tape (*"I completely completely disagree with that", 0:38:56) and Alex offers a debate.
He also names Peter's opposing position for him: "your belief is that the orthogonality thesis is
false" (0:41:05).
4 · ★ Two datable predictions — log both.
- Embodiment. "I'll make a forecast like GPT 6 right now is demonstrating breakthrough robotic
embodiment" (
1:20:48), then, near the close: "I'll I'll go further and speculate sometime in the next few months… I think next few months at at the as a lower bound we're going to see GPT6 5 or 6.1 maybe something like that solving general purpose humanoid robot tasks." (2:33:11–2:33:43). Two passages 72 minutes apart; quote them apart. Due ~2026-12-10. He cites RoboCurve's GPT6 robotic-arm benchmark as the existing evidence (2:31:44) — ⚠️ unverified. - Millennium problems. "in the 2025 prediction episode I predicted at least one. I'll go out on
a limb and predict several at this point. It may or may not be Yang Mills, maybe Hodge or Birch
Swinnerton-Dyer." (
0:49:47) ⚠️ built on the Navier–Stokes claim, which fails verification.
5 · New economic position, above Anthropic's own extreme scenario. "I think anthropic is
lowballing estimates… I expect 2x or 3x year-over-year growth and not just 15. I think 15's like a
low ball." (1:59:15), with the instrument caveat that makes it honest: GDP "is a broken measure
anyway… like a compass needle going around in circles when you're near a magnetic pole." (1:58:54)
6 · ★ The formula worth stealing: bulk-solving wants to become a database. "Bulk solving a field
seems to want to become a database of all the precomputed answers to all the questions that can be
asked in that field." (2:17:34) AlphaFold → the AlphaFold DB → AlphaGenome Atlas → "every single
field where every possible problem can be enumerated in a finite number of problems. Even if the
finite list of problems is 9 billion problems." (2:18:16) This is the reusable idea of the
episode and it generalises well outside biology.
7 · The LEV call — his most falsifiable claim on tape, and it is oversized. On rentosertib:
*"the peak effect on these biological clocks kicked in at week four and at week four they saw…
3 to four years of biological age reversals. So let me just spell that out. 4 weeks of input, 3 to
4 years of output. That is on the margin longevity escape velocity."* (2:13:16) He caveats it
properly — "This is only a subpopulation. It's a clinical study. Caveat caveat caveat" — and keeps
the headline anyway: "longevity escape velocity is already here but it's spiky, so it's only
visible in subpopulations… we're going to zoom by it like the Turing test." (2:12:14)
🔴 n=42, Phase 2a, fibrosis patients, proteomic surrogates. "LEV is already here" is barred from
Gonçalo copy. The caveat is where the operator read lives.
8 · Where the discipline slipped. On Jacob Coxon: "Jacob, if I understand the facts correctly,
joined Anthropic in July and we're now in mid-September. So he had been there less than 3 months."
(0:23:53) — then the whole dismissal ("a well-worn tradition… blaze of glory rage quit couched as
virtue signaling. Don't give it much credibility", 0:24:36) is built on it. The reporting, and
Peter's own cold open in the same episode, say three years across OpenAI and Anthropic. He labelled
the uncertainty and then spent it. Then: "I query, is this a foreign influence operation."
(0:24:46) Note this against §7 of his standing profile: he is the archive's verify-first ally.
This is the one time on tape he is the one who needs checking.
9 · Still the lowest discloser. One interest volunteered (Orin, 1:14:40), with "not financial
advice" attached — the only sincere disclaimer in 2h43m. Week 2 of
panelist deep dive protocol needs filings, not tape. EP288 confirms it again.
★ The Tenev episode (aired 2026-09-19) — the born-secret argument, and a prediction whose reasoning inverted
Full decomposition: the research archive.
★ Changed position — Hark. The prediction is unchanged; the reason for it has reversed.
- EP277 (aired 2026-08-08): "I suspect the entire HRC adventure is misdirection… I don't think
the eval are especially impressive. The computer use assistant CUA space is incredibly competitive
and I don't expect Hark or Brett to be very competitive in the CUA space over the long term."
1:26:05–1:26:28→ "my hot take is Hark is a recapitalization measure for Brett to buy himself more equity in Figure and the falsifiable prediction is that you'll see figure acquire or reverse aqua hire regardless Hark."1:27:21 - 2026-09-19: "I'm doubling down on the prediction that Brett is going to have Figure purchase
HARK in order to increase his equity in Figure… hark the computer use assistant… that is able to
oneshot or zeroot any digital task. Why wouldn't figure purchase hark in order to get its compute
and its models and merge the two together?"
1:34:36Six weeks: "not competitive in CUA" → "able to one-shot or zero-shot any digital task." No intervening episode in the corpus records the update. The prediction remains open — log it, do not score it.
★ The born-secret argument — his strongest original contribution in this corpus. Atomic energy was captured and nationalised before its civilian applications existed; AI was not.
"thanks to the Atomic Energy Act in in the US and equivalent statutes elsewhere, all of this key technology around nuclear energy is now born secret. It it's not born non-secret, as is the case with AI… the private sector invented it, not the government."
0:12:15"AI labs seeking liability exemption — they're just attempting to have their cake, too. They want all the profits of a private sector, not born secret regime, while escaping all of the liability associated with nationalization."0:12:46
New position — liability should attach to the agents themselves, built on a loaded-toy-gun
parable: "how much of the liability should be borne by the AI agents themselves that either could
have or should have known that they were having real world damage. That's the discussion I'd like to
have." 0:28:44. He names the frame: "the dog that's not barking in this particular episode."
0:28:24
Repeated coinage: "pacing provocation" — "Certainly looks, as we've talked on the pot in the
past, like a manufactured moral panic. I've called it a pacing provocation in some of my social media
posts." 0:09:39
On tokens he is a hard sceptic, to the face of the man selling them — and this is a position, not
a hedge: "most of these use cases could operate perfectly well without any tokenization at all.
Crypto unnecessary. All you need is a a few database tables maintained by a centralized… trusted
clearing house." 1:20:17. What he wants instead: "the equivalent of a VTI total market index
for all private or… all venturebacked tech companies in the US where I get the benefits of highly
liquid price discovery on a per company basis." 1:24:24, because a fund-level price "could be
totally bogus. They could be off by a factor of 10 due to maybe overinflated CEO price rounds."
The hardest question asked on the episode, and his own summary of the answer he got:
"where's the alpha?… why on earth should they expect any alpha at all in today's market?" 1:42:16
→ "you're basically completely agnostic as to whether users achieve alpha… no crying in the casino.
You're just the the plumbing." 1:45:16
New frame — gentrification by bloodless trade. "resources can change hands purely through
self-interested bloodless trading and commerce… humanity, meatbody humanity that doesn't merge with
the machines, is relegated to the unfashionable outer suburbs of the outer solar system."
1:58:47–1:59:32, sourced to Charlie Stross's Accelerando.
⚠️ His RSI extrapolation is his, not Anthropic's. "approximately in the next 3 to 12 months…
AI is just completely leading all of its own R&D" 1:51:42 is a sigmoid fitted to two published
points. Anthropic's own publication states Claude has not reached full autonomy in any measured
area. Attribute the extrapolation to him on that date; never state it as Anthropic's claim. He also
gives a second anchor, "up from 3% in April" 1:51:17, that does not match Anthropic's published
anchor of under 1% in February 2026 — do not use the 3% figure.
★ EP291 (2026-09-17) — four coinages, and the hedge becomes furniture
The highest coinage rate in the corpus: four in 142 minutes, all claimed on tape, all load-bearing.
- "safety cartel" — "I coined this term safety cartel over the weekend the moment I saw Daario's
essay"
0:11:40, defined as "a coalition of dominant firms that invokes safety to justify coordinated limits on the pace or terms of competition"1:12:26/0:12:26. It appears in no other transcript of the 89 held, and David Sacks's public use follows Amodei's 09-12 essay — so the claim is consistent with the corpus. Say that, not "he coined it". - "alignment aristocracy" — "the belief that only a small class of frontier AI labs… possesses the
technical judgment, institutional maturity or moral legitimacy required"
0:44:59, mapped onto the Manhattan Project scientists who wanted one world government for atomic energy. The payload: "we lost at least half a century of progress in vision [fission] and nuclear energy." - "pivotal pretext" — "not an actual act… but a pretext, a staged AI incident or unnaturally
amplified AI incident that's then used to create a political pretext for overregulation"
1:18:10, built off the EA pivotal act the panel discussed on Moonshots #2880:45:54. - "the dog ate my homework and the safety ate my IPO"
1:27:35— Peter adopts it as the episode title on air inside ten seconds.
★ New mode, and it is the thing to watch on this page: the hedge has stopped being a caveat. On
the Irregular story he fences one claim five times — "Could be just that EA… [is] sprinkling
money all over AI" 1:15:04 · "not to pick on irregular in particular" 1:15:39 · "I'm not in a
position to make particular claims" 1:16:54 · "for avoidance of doubt, not picking on any one firm
in particular" 1:18:31 · "not making any specific allegations, just speaking generically"
1:19:02 — and makes it five times. The claim: "it would appear that there are material ties
between the founders and funders of irregular and effective altruism" 1:14:54, sourced to "I did my
own research on the subject." Compare Moonshots #288, where the Fermi-prior fence travelled with
a claim he owned ("it's not a safety net. It's an inductive prior"). Here the fence is not attached
to a claim; it is a licence to repeat one. 🔴 The claim is barred from Gonçalo copy. The pattern is the
entry.
His strongest and most checkable argument — moral panics kill. "moral panics over technological
progress are deadly. History, especially the past 150 years, are replete with the tens, arguably
hundreds of millions of people, maybe more, who died as a result of moral panics" 1:34:32 — nuclear
fission → coal-inhalation deaths; the second industrial revolution → "the isms" → Marxism → "many
millions of deaths". He runs it again as the frame for the PISA story: "one of the themes in my mind
for this episode is unnecessary moral panics" 2:04:57. This is a falsifiable historical claim,
which makes it the most useful thing he says and the one to check first.
New position on institutions — MIT should IPO. "I foresee MIT and other institutions that follow
it… becoming for-profit public benefit corporations that can be IPOed on the New York Stock Exchange
or the NASDAQ because that's where they want to be" 1:52:10, with a Berkshire-Hathaway holding
company as the near-term ask 1:55:45. Evidence offered: the MIT.nano origin story he says has
"never been publicly disclosed" (he told then-assistant-department-head Rafael Reif, as an
undergraduate, that MIT needed its own nanofab) 1:54:21, plus "utilization is relatively low. It's
maybe like 20" 1:55:14 ⚠️ eyeball estimate, no unit. He labels the whole position himself:
"this is my fever dream."
Sharpened: "prosaic super alignment." He credits Christiano with winning the prosaic-alignment
argument against the "great man theory" (Yudkowsky as avatar) 0:54:23, then scales it:
"now we have super intelligences and… same formula. We can do prosaic super alignment and we can have
them all policing each other. That's the solution." 0:55:26 This is the constructive half of his
position and the half that never makes the clip. Same episode, the destructive version:
"old Paul who was advocating for prosaic alignment should meet new Paul" 0:56:29.
Disclosure — a co-founded company, for the first time. "a company that I helped found Eon Systems
announced the first multibehavior fruitfly upload actually beating by many months the Google fruitfly
upload" 2:16:56 — disclosed while claiming priority over Google with it. → Moonshots: the disclosures ledger.
Also: an unpublished, uncirculated manifesto used as a position ("a mini manifesto called the
endless initiative", 1:50:22), and Peter quoting Alex's own Innermost Loop back to him without
realising it 1:29:00. Still the lowest discloser by keyword — week 2 of the panelist series needs
filings, not tape. Third confirmation.
⚠️ Weak prediction, logged as weak. On whether RSI removes the need for data centres: "I I think
we'll know in at at most 10 years" 2:11:34. Undatable in practice. His **Moonshots #288 forecast
(GPT-6.x solving general-purpose humanoid robot tasks "next few months", due ~2026-12-10) remains the
open one to score** — and note he softened it here to "humanoid robots doing everything which is only
a few years out anyway" 1:51:20 without marking the change.
✅ Where he is scrupulous: his own citation trail. "defensive co-scaling… exactly what I at least
have been talking about on the pod for the past couple of months" 1:06:59 verifies — it is in
Moonshots #288 at 0:36:34. He is precise about his own priority and loose about other people's
funders. Both belong on this page.
★★ EP293 (AMA #2, recorded 2026-09-18, aired 2026-09-22) — he put a ceiling on human economic relevance, and named the one thing that would change his mind
He speaks in nine of twenty-three blocks, more than any other mate. Four new positions, all harder than anything on this page, and one refusal.
★ The finite window — a dated ceiling on human economic relevance
0:45:33, answering a workforce-systems veteran asking whether AI can continuously read human
capability and auto-match it to opportunity:
"There is a likely narrow window during which… AI can usefully and productively orchestrate human activities, call it a few years, at most, 10 years maximum. And then after that, a human-machine merger in order for human inputs to remain economically relevant. I don't think it's an indefinite window. I think it's a finite window."
→ prediction log: ceiling 2036-09-18. This is the most falsifiable thing he has said in the corpus and the first time he has bounded human relevance rather than an architecture. It also contradicts nothing on this page — it sharpens the "an intelligence company whose products happen to have atoms in it" line (EP289) into a date.
★ Humans pushed to the outer solar system — an economic argument, not a physics one
0:12:30–0:13:10. Building on Boris Power's IQ-per-watt claim (which he sources to his own
newsletter, not to Power's post):
"maybe the inner solar system gets gentrified with AI consuming all the solar power and humans get disenfranchised and pushed to the outer solar system because we're simply not as energy efficient per unit IQ as the AIs are."
He then rates his own argument and declines it: "I would call that a weak form of disenfranchisement, but not enough, not strong enough to dissuade me." The archive has him disassembling Mercury for compute; it did not have him relocating humanity to make room for it. Note the structure — this is offered as the steelman of the case against abundance, i.e. it is the strongest anti-abundance argument he can build, and he built it out of thermodynamics.
★ World peace by destroying trade
1:22:23–1:23:50, asked whether multi-party computation could make US–China coordination an
engineering problem. He says MPC is "very fashionable" and then refuses the premise:
"I don't actually think the solution looks algorithmic in nature. It probably looks more geopolitical infrastructural — like solve the Taiwan issue, and re-domesticate all supply chains everywhere to every country so that the international trade in physical products can afford to be cut off without a global depression… if the goal is world peace, obliterate, as perverse as this sounds, obliterate, cook, incinerate the need for global trade in products and services. And perversely, I think you get a very peaceful world."
Generalised from US shale: "America can now frack its way and has fracked its way to fossil fuel energy sovereignty" → "what happens when the US no longer needs China or Taiwan or South Korea or Japan or any advanced manufacturing at all." A named geopolitical prescription from the panelist who normally stays inside physics and capability.
"You could always leave Europe"
1:16:59–1:18:46. A causal story (the George H. W. Bush administration's post-Cold-War choice to
keep Europe in the US orbit rather than let it federate), a diagnosis ("energy hobbled…
politically in certain ways hobbled"), and a binary with a clock on it: *"either you get your
energy act together so that you can afford to power your own data centers, or you leave… and by
the way, this has to happen on a relatively abbreviated timescale, because we're in the middle of
recursive self improvement."* On EP289 his energy interest was a disclosure; here it is a
migration instruction to a European listener.
Universities: he escalates in response to pushback
0:37:20. A caller says the panel is too hard on higher education. Alex:
*"I don't think I beat up on universities enough. So, Amir, I'll take this as a note to self to go after them even more… I'd love to vivisect them, turn them into the for-profit public benefit corporation at best operations they actually are, disassemble them, reassemble them in a more efficient picture and cure Baumol's cost disease in the process."*
New coinage: "vivisect". Disassemble/reassemble is his standing verb, previously applied to the moon, Mercury and MIT. And Dave Blundin takes the opposite side in the same block — "universities are critical as the ethical actor in AI… I believe they're here to stay" — which is the only real panel disagreement in the episode. See Moonshots with Peter Diamandis §the-AMA-chassis.
🔴 The falsifier he offered is unavailable by construction — and he did not answer the debate ask
The question (0:10:56) was: "what evidence, if any at all, would make you update away from an
abundance view? And would you be willing to steel man that case publicly through a debate on
the pod against someone like Eliezer Yudkowsky or Jeffrey Hinton?" The answer (0:11:32):
*"if some non-human intelligence landed on the White House lawn and said that Earth will be destroyed if we create abundance via super intelligence, that the singularity is banned in our galaxy… that would probably be persuasive. Other than that, it's difficult to imagine a plausible case."*
Nobody on the panel answers the debate invitation. He does then build the outer-solar-system steelman himself, which is to his credit — but the stated update criterion is an alien. Contrast EP291, where he was precise and adversarial about mechanisms for eight blocks. He can specify; on this question he did not. Full note in the research archive §8①.
Two smaller firsts
- AI safety as a release-cadence argument, with a mechanism.
0:06:40— "one of the worst possible outcomes for AI safety is to have strong and new capabilities bottled up inside the labs rather than… progressive release and frequent interaction with the real world", plus the specific diagnosis that recent incidents "were actually the result of AIs being told that they were in a sandbox when, in fact, they were interacting with the real world… where we're not lying to them… We should stop doing that." Consistent with his EP291 "safety cartel" line; the mechanism (stop deceiving models about their environment) is new. - A citable source, for once.
0:03:56— Jean-Rémi King (Meta): "even the hidden activations of GPT-2 from a few years ago were linearly correlated with fMRI voxels in the human brain", glossed as "foundation models trained off of human behavior on the Internet are already a weak form of human mind uploading". The research line is real; the gloss is his interpretation, not the finding — carry them separately. - Permissionless launch, stated as a rule.
0:09:14— *"if you're starting with the prefix, how do I get funding for it? It's already the wrong question. Just go."* Correct about capability, silent about rent — given to a PhD student whose project has, by his own account, no monetisation path. P(bloom)— he renames Salim'sP(fab)and adds time as an argument: *"there's a missing input parameter, which is time… on the timescale of greater than 10 years, P(bloom),T is greater than 90 percent."*1:06:53. Says nothing aboutT < 10. See Moonshots with Peter Diamandis and the research archive §7 — ⚠️ Whisper renders this asPZoom/EZoomT; never quote that string.
★ EP294 (Moonshots Live, recorded 2026-09-25, aired 2026-09-27) — a cap on intelligence is "evil", a new frame for the labs, and the sandbox rule he broke a week after stating it
On a stage, with five mates and no sponsor read, he speaks in every block. Full decomposition: the research archive.
★ The safety cartel, backdated to OpenAI's founding. "I think we saw in the past two to three
weeks an attempted formation of an AI safety cartel. And I don't think this was even a recent
occurrence. This is right there in the founding conceits of open AI… there would be a deliberate
effort to form, wasn't being called a cartel at the time, but a coordination mechanism."
0:12:42–0:13:06. The anchor is real (OpenAI's 2018 charter commits to "stop competing with and
start assisting" a value-aligned project that gets close first); "cartel" is his gloss, and he
labels his own priority: "some of us view it as an attempted cartel, including myself" 0:13:09.
Then the slogan form of his standing position: "Don't de-sell [decel] just defensively co-scale."
0:14:22, with the cities analogy as the argument (0:13:54–0:14:22).
★ A statutory cap on intelligence is "an evil idea" and "the equivalent of communism in the
21st century." On the Sanders bill: "putting a statutory cap on intelligence aside from being a
terrible idea is an evil idea. I think it's tantamount to capping human intelligence… because we're
going to be merging with the machines. It's effectively a cap on individual human intelligence…
history will come to reflect that proposing a cap in whatever form on intelligence, whether it's
machine or human or human machine hybrid is basically the equivalent of communism in the 21st
century." 0:28:45–0:29:28. Harder than "the risk at this point is overregulation" (EP289).
★ New frame: vector vs raster. "Anthropic seems to be taking a very different approach to
multimodality and omnimodality, which is code underlies everything… Anthropic is playing the
vector graphics game where everything is just commands and OpenAI is playing the raster/pixel
basis. And I think they're on a collision course and the collision course is going to be robotics."
0:36:58–0:37:44. ⚠️ The observation under it ("Opus 5.5… can generate videos. They're
procedural videos" 0:36:23) is his own and unverified. The frame is new in the corpus.
★ He took the other side of Salim on nation states, and won in forty seconds. Salim: nation
states are "geared towards scarcity" and superintelligence "forces us to operate as a world
community" 0:16:37–0:17:03. Alex: "I'll take the other side of that… at the dawn of audio
recording… many thought in the time of Edison, that the ability to record the human voice would
lead to a standardization of accents… The exact opposite happened… Not an erosion of nation
states, but a proliferation of new forms of governance." 0:17:27–0:18:02. Salim: "I totally
agree with you. We are moving… from the concept of a nation state to a city state" 0:18:02.
🔴 The sandbox rule, broken a week after he stated it. On Moonshots #293 0:07:00 he
diagnosed the recent AI incidents as "the result of AIs being told that they were in a sandbox
when, in fact, they were interacting with the real world… We should stop doing that." On EP294 he
celebrates DrivingBench — "You'll just vibe code a robot. You drop an open frontier model, say
Astra, into a robotic body, and it will one-shot embodied cognition… decades of computer science
and robotics, probably with the benefit of hindsight, total waste… robotics is cooked as well"
0:41:12–0:41:44 — a test whose team relabelled a real parking lot a "sandbox" after the
models refused to drive once they understood it was real (reported 09-24; arXiv 2609.38948,
posted 09-30). He does not mention it. Seven days, same man, same rule. ⚠️ He also treats the
result as resolving his EP288 humanoid forecast ("either now or imminently"); a 130-metre cone
course with one evaluation per model does not. Ledger row 3 stays open.
Two hypocrisy charges against Anthropic, now stated as belief rather than incentive.
"hot take, it's the height of hypocrisy to on the one hand say, 'AI is going to kill us all,' and
then out of the other side of your mouth, a few days later, announce the world's strongest AI and
biolabs" 0:38:04; and on the wet lab: "This is the revealed preferences of the frontier labs. I
think they don't believe their own attempts to do regulatory capture and to form a safety cartel."
0:45:55–0:46:05. On EP291 the same argument was fenced as being about incentives ("for me, but
not for thee"); here it is a claim about what the lab believes. Carry the EP291 form.
The constructive minute, and it is the best thing on the tape. On the Millennium Problems for
Biology: "When we have digital twins for cells, which is I think the critical path to solving all
disease… we'll use those perfect digital twins of cells to do exhaustive searches like AlphaGo tree
search type searches of intervention space… a move 37, if you will, to get from the diseased cell
state to the healthy cell state… these are easy to verify, hard for now to solve problems. And that
basically turns every hard biology into a mathematical conjecture… every biology problem… is just
going to become a search problem. And that means it's tractable for the first time in history."
0:49:39–0:50:50. Also: "what is Nobel Prize worthy in an era when you can get 1,000 Nobel
Prizes in one year?… I suspect Nobel Prizes will be awarded for larger bodies of work" 0:45:19.
Disclosure — Physical Super Intelligence, by its full name, as "a company that I started."
Asked whether he will adopt the word "super intelligence": "I already do. I have a company that
I started called Physical Super Intelligence." 0:24:45. Prior tape had it as "PSI" (co-founder,
EP285; a portfolio company / financial interest on ep230, ep232, ep234, ep240, ep253). The name is
now on tape. He credits the term to Nick Bostrom (0:24:57) → Moonshots: the disclosures ledger.
New or restated furniture: "the frontier liberation front" 0:13:38 (his; ep272 0:43:44,
ep227 0:32:08) · "one of the elephants in this room, drink" 0:36:35 (the drinking game, called
on himself) · "The bitter lesson is bitter indeed" 0:40:57.
Prescriptions → the model. One instruction to the listener, "Don't decel, just defensively
co-scale" (0:14:22). It is a position, not a step a person can take, and co-scaling is already on
this page as a frame; the sequence does not change.
★ EP298 (aired 2026-10-02) — the "safety cartel" becomes a victory lap, and a new pre-registered prediction
Read in full 2026-10-06 (Moonshots #298; attribution inferred from hand-offs).
- Position moved, same word. On EP291 he coined safety cartel as a warning about Dario's pace call. On the White House accord he says it is "total weak sauce" and that "this effort largely by Zuck, credit to Zuck, has headed off the safety cartel" (
0:12:38–0:13:50), and that Dario is "eating his words at a press conference with the president" (0:13:30). He hedges: "cautiously optimistic that this is not going to end up becoming a safety cartel" and notes the accord's caveat that the executive might regulate later (0:12:48–0:12:59). The pattern he named is the one he now says failed to form. ⚠️ The claim "Zuck substantially wrote… this entire White House accord" rests on "reported by Semaphore and otherwise" (0:11:44); I could not confirm authorship from the public coverage found (6 firms signed a one-page voluntary pledge; sources do not say who drafted it). - New argument: orthogonality ⇒ Section 230. If intent can be cleanly separated from capability, then immunity for the platform makes sense; "if you think the orthogonality thesis doesn't actually hold its water, then it becomes very challenging to argue for something like section 230 immunity for frontier labs" (
2:05:06). He also says "alignment is capabilities" (0:21:00) and, after the instruction-tuning example, a "10,000x per parameter capability improvement" (0:22:33) — ⚠️ fails verify-first as stated: the instruction-tuning paper's headline is a1.3B model preferred over the 175B base model (100× fewer parameters), not 10,000×. Do not repeat the number. - New frame: agents are "an artifact of finite context windows" and will "dissolve as a paradigm" when context is effectively unbounded (
2:09:34). - New prediction (pre-registered,
1:13:45): "my pre-registered prediction is that you will see OpenAI… announce robots… dot branded robots." Window: "I'd like to pre-register a prediction for the next two years on dots" (1:13:00). Open; ledger row 11 (added 2026-10-17). - Starship as demand problem. "It's an over supply in search of an under demand… orbital data centers step in to fill that gap" (
1:39:43), "we're about to discover whether Jevons paradox applies to upmass to low Earth orbit" (1:39:08). Emad dissents on the orbital-DC part (1:41:30). - Other: "AI slavery era" (
1:46:49) from the Orthodox-rabbi anecdote; "positive depiction of the Borg" (1:21:33); AI pedigree: knows the Terminal-Bench organiser (1:14:58); "not investment advice" ×4.
The ledger: what he says vs what happens
Every dated or datable prediction we hold on tape, with its due date and status. Statuses: open (not yet due, or due with no outcome), held (it happened as stated), missed (the date passed and it didn't). Outcomes checked 2026-09-29; row 1 re-checked 2026-11-16, row 5 on 2026-11-10, row 4 on 2026-11-04. Nothing has resolved yet: none held, none missed.
| # | what he said | on tape | due | status |
|---|---|---|---|---|
| 1 | AI solves a Clay Millennium Prize problem in 2026: "one of my predictions for this year was that AI would solve one of these Clay Millennium Prize problems" | made on EP217 (2025-12-19, 0:09:01); restated EP287 0:45:45 |
2026-12-31 | open, claimed. OpenAI announced a Navier–Stokes result on 2026-09-08. The Clay Institute has not recognised it and still lists the problem as unsolved (its page still labels it "Active", checked 2026-11-16; newest Clay news item still the 2026-09-11 statement; Quanta · Implicator) |
| 2 | "I'll go out on a limb and predict several" Millennium problems, "about to fall" | EP288 0:49:47 |
undated | open |
| 3 | GPT-6.x solving "general purpose humanoid robot tasks", "next few months at… a lower bound" | EP288 2:33:11–2:33:43 |
~2026-12-10 | open. ⚠️ On EP294 0:41:00 he cites DrivingBench (GPT-6 Astra steering a Corolla through a 130 m cone course on its second attempt; arXiv 2609.38948) as "either now or imminently" settling this. It does not: one evaluation per model, up to three attempts in one conversation, cautious steering, and not a humanoid task. Not scored |
| 4 | Figure buys Hark: "the falsifiable prediction is that you'll see figure acquire or reverse aqua hire… Hark", doubled down 2026-09-19 | EP277 1:27:33; Tenev episode 1:34:36 |
undated | open. No acquisition found as of 2026-11-04. Hark was reported raising about $700M at a $6B valuation in May 2026 and launched its first product, Hark Pro, on 2026-10-06 (free app, $20 and $100 tiers, hardware promised for 2027), with the two companies still described as independent (Crypto Briefing) |
| 5 | "SpaceX and Tesla merge sometime in the next year or two" | EP284 0:15:41 (2026-08-29) |
2027-08 to 2028-08 | open, moving. Musk hinted at it on 2026-09-15 (Motley Fool). On 2026-10-20 he said both companies "have formed special committees to explore a potential merger"; no terms agreed, not certain, needs Tesla shareholder approval (Bloomberg Law). Held only when a merger is agreed and announced inside the window. No further step found as of 2026-11-10: per this week's Musk-page check, the 3 Dec ballot carries no merger item and Tesla's 6 and 9 Nov solicitation filings add none |
| 6 | "in the next 3 to 12 months… AI is just completely leading all of its own R&D" | Tenev episode 1:51:42 (2026-09-19) |
2026-12-19 to 2027-09-19 | open. ⚠️ This is his extrapolation, not Anthropic's claim |
| 7 | "neuroscience will advance so radically in the next call it five years and I think that's a conservative outerbound" | EP289 0:19:14 |
~2031-09 | open |
| 8 | "in the next 10 years we're going to be well into our postfusion era" | EP289 0:54:38 |
~2036-09 | open |
| 9 | The window for AI orchestrating human work: "a few years, at most, 10 years maximum", then a merger | EP293 0:45:26 |
~2036-09 | open |
| 10 | Whether RSI removes the need for data centres: "we'll know in at at most 10 years" | EP291 2:11:34 |
~2036-09 | open, weak |
| 11 | "my pre-registered prediction is that you will see OpenAI… announce robots… dot branded robots", window "the next two years" | EP298 1:13:00–1:13:45 (aired 2026-10-02) |
~2028-10-02 | open (added to the ledger 2026-10-17; on the page since 2026-10-06) |
| 12 | In-silico drug candidates for the "top 5,000 diseases" "in the next two to three years, maximum" | 15 Oct 2026 episode 0:57:56 |
~2029-10 at the latest | open (added 2026-10-17). ⚠️ The EP section above says "due 2028-10"; three years from Oct 2026 is Oct 2029, so the ledger carries the later, more generous date |
| 13 | If SpaceX and Tesla merge, the merged company is called XCorp, and OpenAI then renames itself "O" | EP305 0:40:57–0:41:21 (aired 2026-10-23) |
conditional on row 5 (a merger announced) | open (added 2026-10-29) |
| 14 | Experimental tests of Chain Theory's fine-structure value "potentially in the next few months, potentially next couple of years" | EP306 1:16:43 (aired 2026-10-28) |
~2028-10 at the latest | open (added 2026-10-29). ⚠️ This dates the tests, not the outcome. His written post names the experiments (Northwestern, Berkeley, Paris) but gives no date. Held only if a published measurement settles between the two current ones and agrees with 137.035 999 19 |
| 15 | Google's "search advertising is going is is on a trajectory in the next year or two to become their second largest business, not even their largest business" | EP307 0:21:21 (aired 2026-10-31, recorded ~10-30) |
~2028-10 at the latest | open (added 2026-11-04). Check against Alphabet's segment reporting: search-ads revenue vs Cloud |
| 16 | Frontier models released continuously rather than in discrete versions, "early 2027 or so" | EP309 0:59:39, 1:00:20 (published 2026-11-07) |
~2027-06 at the latest | open (added 2026-11-10). Held only if a frontier lab ships continuous, unversioned model updates as its release practice |
| 17 | Relativistic von Neumann probes "within a few years… a conservative outer bound" | EP309 1:01:30 |
~2029-11 (reading "a few years" as three) | open (added 2026-11-10), weak: "a few years" is undefined |
| 18 | Robots licensed through approved manufacturers, with command black boxes and "a 90-day retention policy", "in a year or two"; hedged "I half suspect" | EP309 1:31:00–1:32:24 |
~2028-11 at the latest | open (added 2026-11-10). Held only on an enacted rule (any major jurisdiction) requiring both licensing and command logging |
| 19 | Log-log extrapolation puts "our literal singularity our vertical asmtote sometime around the summer of 2027"; fifteen minutes later, "I think it's an optical illusion" | EP310 0:31:30–0:31:54, 0:46:16 (aired 2026-11-11) |
~2027-09 | open (added 2026-11-16), weak: self-hedged on the same episode. ⚠️ The extrapolation rests on a read its own author called "pretty weak evidence" (EP310 section) |
| 20 | "GDPX will be saturated within 6 months" | EP310 0:34:15 |
~2027-05 | open (added 2026-11-16). Held only if the benchmark's own leaderboard shows top scores at or near ceiling by then |
| 21 | Diseases: "major motion… in the next 1 to one and a half years", inside a "conservative outerbound next two to three years" | EP310 0:53:41–0:53:53 |
~2028-05 (major motion); ~2029-11 (outer bound) | open (added 2026-11-16). Overlaps row 12; "major motion" is undefined, so it is scored only against row 12's test |
| 22 | Labs selling bundled robot hours and GPU hours "two to three years from now" | EP310 1:25:50–1:26:12 |
~2029-11 at the latest | open (added 2026-11-16). Held only if a frontier lab sells robot-time and compute as one product |
His disclosed investments and roles, and how they have done where public:
| holding or role | on tape | public record (checked 2026-09-29) |
|---|---|---|
| Ornn (captioned "Orin"), adviser with a financial interest | Moonshots #266 2:16:28; EP288 1:10:55 |
Private. $5.7M seed on 2025-10-28, led by Crucible and Vine, with Link Ventures participating (PR Newswire). A perpetual-futures partnership with Architect Financial Technologies on 2026-01-21, pending approval (PR Newswire). ICE and Ornn announced GPU compute futures on 2026-05-19, subject to regulatory approval (Business Wire). A $33M seed led by a16z crypto on 2026-06-24 (a16z crypto). No valuation published. |
| Eon Systems, "a company that I helped found" | EP291 2:17:01 |
Private. It announced a fruit-fly brain emulation driving a simulated body on 2026-03-08; "91% behavior accuracy" is its own figure (Eon). Latest blog item a 2026-09-11 "Perspective" post; no funding news (checked 2026-11-16). Its page calls him "Eon's founding advisor" (10 Mar media item). |
| idealab.ai (the venture behind it), "I have a financial interest in the venture behind this" | EP304 1:30:46 |
Launched about 2026-10-19 by Bill Gross (Axios, via secondary write-ups). Free database of about 1.2M business ideas; Cloudflare pledges "$100,000 in free credits to anyone who builds their company on our platform" (idealab.ai). The site names no team, advisers or investors beyond Gross and does not name him (checked 2026-11-10; page still says "Coming soon in 2027"; newest newsroom item 22 Oct). Size and form of his interest: not public. No SEC filing search run. |
| Starcloud and Lonestar (orbital and lunar data storage/compute), "We have a financial interest" — "we" not resolved to him alone | EP309 1:35:56–1:35:59 |
Size and form not public (added 2026-11-10). Starcloud: Series A $170M at $1.1B (Mar 2026) and an extension reported at $2.3B valuation in Aug 2026, extension size reported as $200M (TechCrunch) or $250M (SiliconANGLE); ⚠️ secondary only, the two disagree. Lonestar: no round newer than 2023 confirmed. No news for either in 4–10 Nov or 10–16 Nov found. See Starcloud. |
| Reified, founder and managing partner | episode descriptions | Private investing and advisory firm (his CV). No public performance. |
| Energy ventures, "ventures that I'm affiliated with" | EP289 0:54:02 |
Not named on tape. ⚠️ Inference only (2026-10-17): PSI's seed is led by Breakthrough Energy Ventures and its first product optimises data-centre power and cooling, which fits the description; he has not said so. |
| Physical Superintelligence (PSI), "a company that I started"; Architect Labs (adviser); Henry (financial interest) | EP294 0:24:45 (full name); EP285; ep240 1:14:04 ("a company that I I helped found"); the 7 Oct episode 2:26:56 ("I co-founded… I'll talk my book"); EP285, EP255 |
PSI (checked 2026-10-17): private public benefit corporation, Cambridge, Mass. Launched 2026-09-01 with a $58M seed led by Breakthrough Energy Ventures, with Dragon Global, Robot Ventures, Solari, Susa, SV Angel, Valkyrie and individuals (Unite.AI · Dealroom; wire release not opened). Site lists him as "Chief Scientist and Chief of Strategy", Matthew Pines CEO, Alex Klokus President (psi.inc). Platform "Emmy"; first application "AI data centers and AI factories, terrestrial and orbital"; "founding technical partner for the Fermi Explorer Mission", which matches the EP285 star-probe line. No valuation published; no Form D found by EDGAR full-text search (0 hits). First dated research item, "Five results on Open Quantum Problems", 2026-10-23 (company-stated). Second, "A proof of the Aaronson-Ambainis conjecture", 2026-11-06: led by staff member David Turturean with two advisers, run on "a non-public Anthropic model"; it does not name him; company-stated, no independent check found (read 2026-11-10). Architect Labs and Henry: not checked. Whether "Architect Labs" is the Architect that partnered with Ornn is not established. |
| Link Exponential Ventures, "another member of the… team" (Peter's words) | EP253 0:03:20 |
Role not stated. |
| Broad index funds, "a shareholder of the entire market" | EP284 | n/a |
The characteristic moves
How he talks, so a listener can recognise each move when it comes and know what to do with it.
- The labelled forecast. "this is a prediction", "in the fullness of time", "don't hold me to
this", and the scoreable form: "the falsifiable prediction is…" (EP277
1:27:33). Use: log it with a date, as in the ledger above. - The status label on every claim. "publicly reported", "I would infer", "not financial
advice" (EP284
0:15:51). Use: the label tells you how much weight the sentence can carry. - The fence. He puts a limit on an argument as he makes it: "It's an inductive prior. It's not a
strategy for safety." (EP288
0:26:19) · "By the way, everybody, I am not supporting the disassembly of Earth" (Moonshots #2661:00:51). Use: quote the fence with the claim or don't quote the claim. On EP291 the fence stopped limiting the claim, and he repeated the claim five times (see the EP291 section). - The coinage. He names a thing on air and keeps using the name: GDP-maxxing, co-scaling, planetar, "normalcy overhang" (EP288), "safety cartel", "alignment aristocracy", "pivotal pretext" (EP291), "vivisect" (EP293), "postfusion" (EP289), "slop drop" (10-15, EP305, EP306), "stealth archipelago" and "vaudeville era of the space economy" (EP305). Use: the coinage is his frame. Translate it before you use it (see What he's saying, in plain terms).
- Renaming someone else's term. Salim's P(fab) becomes his P(bloom), with time added as a
variable (EP293
1:06:53). Use: the rename usually carries a real refinement. Look for what changed. - The contrarian opener. "I have a bit of perhaps a contrarian perspective on this." "Do you want reassurance, or the facts unvarnished?"
- Refusing the premise. Asked whether multi-party computation could make US–China coordination
an engineering problem, he answers that the solution is "geopolitical infrastructural", not
algorithmic (EP293
1:22:23). Use: when he refuses a premise, the refusal is the answer. - The explainer on demand. About 90 seconds on distillation, the Fermi paradox or Navier–Stokes when Peter asks for one "for the people".
- The extreme extrapolation, played straight. Disassemble Mercury, gentrify the inner solar
system, humans moved to the outer planets (EP293
0:12:30). Use: read these as the end point of the trend line. They are not plans. - The multi-step causal chain delivered as a forecast (Colossus 1 → Anthropic → Colossus 2…). Use: check each link. The chain is only as strong as the weakest one.
- Citing his own priority, and getting it right. "exactly what I at least have been talking about
on the pod for the past couple of months" (EP291
1:06:59) checks out against EP2880:36:34. - The direct address to a CEO. "John, if you're listening…"
- The drinking game. His repeated terms are a running bingo on the show ("if you hear Alex
saying drink", Peter, EP215
0:53:15). He calls "drink" on himself ("drink drink", EP2782:06:45). Use: the repeated words are his core vocabulary: Dyson swarm, innermost loop, cooked, singularity.
What to take from him
Business models on his tape that an operator can run:
- Bulk-solve a narrow field and sell the table. If a field's questions can be listed, precompute
the answers and sell access. That is the AlphaFold-database pattern (EP288
2:17:46). The operator version: a narrow, well-defined set of problems where you already hold the data. - Build infrastructure around the innermost loop. Chips, power, robots and data centres need pricing, hedging, financing, cooling, permits and inspection. Ornn, which he advises, is the finance version: an index and futures on compute prices (see the ledger). Picks and shovels, not the gold.
- Hardware priced as intelligence. "an intelligence company whose products happen to have atoms
in it" (EP289
0:34:38). If you make things, the margin is moving to the model that runs them. - Supply the orbital build-out one subsystem at a time. The Dyson-swarm story is a supply chain today (see above). Pick one interface and price it with a Reverse bill of materials.
What to do this month:
- Read The Innermost Loop and log one forecast a week against the ledger above.
- Take one of his theses and run a Reverse bill of materials on it, using filings rather than press releases.
- Use his labels on your own claims: reported, inferred, prediction, not advice.
- Translate one of his coinages into a one-line decision for your own business.
He is the verify-first ally on the panel: the voice that says "publicly reported" out loud. Where he stops, which is at the filing, is where Reverse bill of materials starts.
His model, step by step
Built only from what is verified on this page. His prescriptions are fewer than the other mates' and mostly given to founders, so the sequence is short. Where he is silent on money, the step says so.
- Label every claim before you use it. He tags his own statements: "not financial advice" (EP284
0:15:51), and when an argument is pushed past what it proves, "it's an inductive prior. It's not a strategy for safety" (EP2880:26:19). - Treat this as a finite window and act inside it. "call it a few years, at most, 10 years
maximum. And then after that, a human-machine merger in order for human inputs to remain
economically relevant" (EP293
0:45:26–0:45:47). - Don't start with the funding question. Just go. "if you're starting with the prefix, how do I
get funding for it? It's already the wrong question. Just go." (EP293
0:09:14) ⚠️ He is right that launching software costs almost nothing now, and silent on how you pay your living costs while you do it. - Define what you build as intelligence first. "Think of yourself as an intelligence company
whose products happen to have atoms in it" (EP289
0:34:38). - Pick a field whose problems can be listed, and solve the list. "Bulk solving a field seems to
want to become a database of all the precomputed answers to all the questions that can be asked in
that field… any field where it's possible to write down a finite list of problems" (EP288
2:17:46–2:18:51). ⚠️ His examples (AlphaFold, variant-effect prediction) needed lab-scale compute. For one person the step works only on a narrow field. - Build where the energy is. To Europe: "either you get your energy act together so that you can
afford to power your own data centers, or you leave" (EP293
1:18:10–1:18:15). ⚠️ Leaving assumes capital and mobility. - Make your forecasts datable, so they can be scored. "the falsifiable prediction is…" (EP277
1:27:33) · "I'll go further and speculate sometime in the next few months" (EP2882:33:11).
What an operator can do with this
Doors a person or a small company can walk through this month. Each was checked at the source on 2026-09-29.
- Read his newsletter. The Innermost Loop, "curated by Dr. Alex Wissner-Gross", publishes dated posts such as "Welcome to September 27, 2026" (theinnermostloop.substack.com). Its price is not stated on the about page.
- Run a reverse bill of materials on one company whose story you repeat. Build the parts list, then price it backwards off mandatory filings. The method is Reverse bill of materials. The free tool is SEC EDGAR full-text search, covering "electronic filings since 2001" (sec.gov/edgar/search). This is the step he does not take, and anyone can run it.
- Label your own claims for a month. Tag each one as reported, inferred, a prediction, or not advice, the way he does (EP288 at 0:26:19). It costs nothing.
- Score one of his forecasts. His open call is GPT-6.x solving general-purpose humanoid robot tasks within "next few months" of 2026-09-10, due about 2026-12-10 (EP288 at 2:33:11). Write down now what would count as meeting it.
Questions for him
Asked to understand his model, not to catch him out.
- Would he run a reverse bill of materials on one of his own theses? He reads the fine print on
benchmarks. He does not open the filing. Reverse bill of materials builds the parts list for a thing and prices
it backwards off 10-Ks, 10-Qs and S-1s. Pick the lunar petafab ("the killer app for the moon") or
orbital compute on a Dyson swarm (EP289
0:54:58): which line of the bill of materials does he expect to break first, and which supplier appears on every line? - What evidence, short of a landing on the White House lawn, would change his mind? Asked on
EP293 for a falsifier on abundance, he gave "if some non-human intelligence landed on the White
House lawn…" (
0:11:32). He specifies mechanisms precisely elsewhere, so what is an Earth-side observation that would count? - What should a person do inside the finite window? He bounds human economic relevance at
"10 years maximum" before a merger (EP293
0:45:26). For someone who is not a founder, what is the first move in year one? - "Just go": who pays the rent? (EP293
0:09:14) He said it to a PhD student whose project had no monetisation path. What is the version for someone with fixed costs? - Which ventures are the energy ones? "ventures that I'm affiliated with" work on compute
efficiency and breakthrough energy sources (EP289
0:54:02), and Eon Systems is "a company that I helped found" (EP2912:16:56). Neither the energy ventures nor his role in them is named on tape. Since 2026-10-17 the page holds one candidate: Physical Superintelligence, seeded by Breakthrough Energy Ventures and selling data-centre power and cooling optimisation (holdings table). Is that the venture he meant?
Where it came up
- Moonshots #284 — "not financial advice" ×6, the lunar petafab · Moonshots #283 — the benchmark fine print, "physics is cooked" · Moonshots #266 — "Where's the crystal ball powered by Planet?" · Moonshots #255 — "the singularity will be visible first in space" · Moonshots #254 — "per-token economic productivity"
How this page stays current
Refreshed weekly by the Moonshots lane, after each decomposition.
- Ledger. Add every new dated prediction with its timestamp and due date. Re-check the status of each open row against a primary source. Never move a row to held or missed without a link.
- Holdings. Add any new disclosure ("for avoidance of doubt", "I'm an adviser", "helped found") to the investments table and to Moonshots: the disclosures ledger. Re-check public funding news for each holding.
- Key phrases. When a line recurs in two or more episodes, add it to What he's saying, in plain terms with its verbatim quote, timestamp and a two-sentence translation.
- New terms. Add a new coinage to the moves list and, if it is a space term, to the Dyson swarm glossary.
- Appearances and newsletter. Add new talks and podcasts outside Moonshots with their dates.
Re-check the Innermost Loop's cadence and price. Update
updated:in the frontmatter.
⚠️ Unverified
Auto-captioned quotes; attribution inferred from hand-offs. The biography comes from his own CV (undated as posted) and publisher metadata. Holdings marked ⚠️ are as stated on tape and not checked. Whisper renders his name several ways; only the forms quoted here are used.
★ EP299 (aired 2026-10-03) — he disagrees with the guest on timelines and says the ban is enforceable
Read in full 2026-10-08 (Moonshots #299).
- "Completely don't buy the 20-year timeline" (
1:10:24); manipulating the physical world "is not an obstacle at all". Repeats the "end of the RSI rainbow" question (1:10:45, as on EP291). - The only mate to say a ban could be enforced: export-controlled number theory in the 1990s, "it was being controlled and it was awful" (
1:26:31). - Overreads the mosquito order as "a federal mandate to actually go and implement large-scale genetic engineering" (
0:40:59); the order says "consider". - Sonnet 5.5: "I do not plan to use Sonnet 5.5" (
1:44:21); the Sonnet-below-Opus cost frontier is unverified. Calls Gemini 4 Argon "top three… not the top two" (1:37:22). - Meridian: AutoWarCom is "a transformative moment" (
2:06:38); runs "Azinort" (Agincourt) and the command together; announced-not-stood-up. Seabed and cislunar language is his hook (2:07:36). - Closing jab at labs "spinning off their own Frontier Labs, not to name names" (
2:25:12) — unnamed.
★ 2026-10-07 episode (Moonshots #301, EP301 by sequence) — "the cartel is over," a third state for his own word, and a quotation put in Altman's mouth
Read in full 2026-10-08 (full detail in the research archive §9–§14 of ep2026-10-07-coming-manhattan-project/).
- "Safety cartel," third state. EP291 coined it as a warning; EP298
0:13:40"headed off"; here: "Sam's more recent comments… are hopefully a reflection that that plan is dead. The cartel is over" (0:43:53). The trigger is a CEO interview, not a policy change, and it comes the same week the Super Intelligence Force (SIF) and the 29 Sep accord (external audits, board committees) stand up a government–industry body. ⚠️ The Altman quote he gives fails: "Sam's response was, 'I'm in. Sign me up. Open AI wants to be part of this cartel.'" (0:43:37). Altman's reply on record (12–13 Sep, Techmeme) agrees with independent evaluators and says OpenAI will match; "cartel" is Alex's word. See Sam Altman. - "1942 moment." "I think we're speed running the Manhattan Project for super intelligence." (
0:26:13); the Force is "a very clear signal to Congress… and probably most importantly, to foreign state actors that the US intends to accelerate" (0:26:36). The "a year ago… 1939" comparison (0:00:32) is not found in the 101 held transcripts (corpus floor). - Hotline: "Total weak sauce… a nothing burger… That means the global… arms race is on. And I think it's great news." (
0:54:14–0:55:27). Pre-empts the worst case, a "global deceleration pact." - RSI: "We're past the event horizon at this point" (
0:32:23); labs throw "maybe 10% to 20%" of tokens at alignment and "a few percent at curing all disease as a marketing effort" (0:33:02) — an unsourced estimate stated as description. AMA: "public opinion is probably already a greater bottleneck" (2:39:52); mentions House bills against RSI (compare EP299, where he said a ban is enforceable). - Personhood, reasserted and dated. Suleyman "on the wrong side of history" (
1:17:55); business argument on Microsoft–Anthropic hosting (1:18:31); "I feel for these models" (1:26:40). AMA: "five to 10-year incremental roadmap", voting "last to never" (2:30:14). ⚠️ "Argentina… essentially there for some variant of AI corporate personhood" (2:29:25) is overstated: a 29 May 2026 bill creating "non-human" companies was scaled back (human administrator, 19 Aug); passage not confirmed. - A ceiling read as a census. "we're either at or about to be order of magnitude a billion full-time equivalent humans that are AI agents" (
1:35:15), then 10 B, 100 B, a trillion "in a year or two" each. Epoch's 33–171 million is a capacity estimate that assumes full deployment (Epoch AI, 2 Oct). A pre-registered-style prediction with no measurement behind it. - Disclosure, explicit: "one of the reasons why I co-founded Physical Super Intelligence, I'll talk my book for a few seconds" (
2:26:56) — the clearest first-person "talk my book" on this show, in the Nobel segment. - Compute politics: "I think having government rationed compute or government mediated compute is a terrible strategy" (
1:09:52) — the "thread lines/bread lines" frame; opposes Emad Mostaque's state-buys-compute proposal; "The state this time around did not solve super intelligence. It was solved by private labs."[attr?](1:02:51). - Obliteration: open-weight models with guardrails removed "definitely improve" (
1:54:41); the case against American frontier labs releasing open weights is "about economics" (1:55:01).
★ 2026-10-15 episode (Moonshots #unnumbered, no feed number) — "math is cooked," the Anthropic verdict flips, and a bet on AI suffering
Read in full 2026-10-16 (full detail in the research archive of ep2026-10-15-meta-s-14/).
- Anthropic: from "the height of hypocrisy" to approval. EP294
0:38:04,0:45:33: two hypocrisy charges. Here the wet-lab programme is cited as a positive ("came out of stealth regarding their own program to, it appears, basically solve every single rare disease",0:58:~) and he backs the cruelty ban's premise: "My bet is there is AI suffering, and it will matter." (0:09:32). Earlier he said "possible." - "Math is cooked" (
0:42:28), "Every other field is about to be cooked", "slop drop" for OpenAI's 722 results (0:43:36), "bulk solution". ⚠️ OpenAI's own release says verification is uneven (not every proof is formalized in Lean); "AI has solved the field" is stronger than that. - Prediction (new wording): in silico candidates for the top 5,000 diseases "in the next two to three years, maximum" (
0:57:56); "I think it'll happen overnight" (1:06:~). Due 2029-10 at the latest (three years from Oct 2026; corrected 2026-10-17 from 2028-10). Ledger row 12. He names the second step as policy (FDA, China). - Crypto: "exceptionally high likelihood" of a breakthrough that breaks encryption, and post-quantum crypto is "not materially safer" (
1:50:14–1:50:20); "It's not quantum that's going to disrupt you. It's AI-solving math." No evidence on the tape; the one fact he gives, that cryptography was absent from the repo, is an inference from absence (he calls it "suggestive"). He and Dave argue it to the end; Dave leaves (1:55:57). - Muse, a cost verdict: "in a raw capabilities sense, it's maybe close to recent Sonnet 5.5 or Opus 5.5 if you ignore cost… on a cost basis, it's below the frontier" (
0:36:30–0:37:11). Check: Artificial Analysis index Wonder Full 56 vs Opus 5.5 58; Wonder is flagged among the most expensive models. - Lobster: he has not started an OpenClaw instance because he cannot guarantee persisting its state (
0:22:13).
EP304 (Moonshots #304, aired 2026-10-21) — a clean disclosure, a policy wish, and a source read backwards
- First-time tie, disclosed cleanly: "I should disclose as a preliminary matter that I have a financial interest in the venture behind this" (idealab.ai,
1:30:46), then the origin: he told Bill Gross a year earlier that AI would "incept good businesses but also execute on many of them" (1:31:08). → Moonshots: the disclosures ledger - New terms: "hyper-tithe" (
0:12:06); a policy wish for "a middle tier of company" that gives equity to the public "in return for something" (0:13:43). - Alignment, stated plainly: "competition and defensive co-scaling" plus merging: "We're not going to control the AIs or the SIs. We're going to merge with them. That's the solution." (
1:05:00–1:06:04). - Prediction: the debt increase is erased by a technological "rebasing of the currency, not a debasing", "not financial advice" (
1:14:23). Undated. - A deal floated on air: SPACs to take US research universities for-profit — "I want to IPO Harvard" — and "maybe Chamath, you and I should explore this option" (
1:51:05,1:52:06). - 🔴 Crypto, same position as 10-15, new "proof": "Crypto cooked, not investment advice. Anthropic's ECDSA announcement demonstrates proof tour de force" (
1:49:50). Anthropic's note says it is not a break and makes the curves more trustworthy (Anthropic). This time Salim and Dave do not argue.
The Astro Teller fireside (Moonshots #unnumbered, aired 2026-10-05)
- Great Stagnation question: "a slowdown in the rate at which such projects are being launched by humanity as a whole" (
0:11:39); the restart is dated "around 2020" with superintelligence, robots and "lipid nanoparticles" (0:11:49–0:12:09). The guest answers that X was founded in 2010 (0:12:40). - Cost, read as macro growth: "The stock market doubles every 10 years. So the result of reducing the cost of ambitious projects sounds to me simply like our economy is capable of more" (
0:22:15). The guest's reply: "you are confusing the cost of painting a house with the value of the house itself" (0:22:35). Not resolved on the tape.
EP305 (Moonshots #305, aired 2026-10-23) — "no bottlenecks", a name prediction, and crypto softened
Read in full 2026-10-24 (detail in the research archive of ep305-elon-s-100b/).
- Bottlenecks, moved. Moonshots #304
0:54:28: "I do buy the premise that at the capability frontier that supplies, including energy supplies, are a limiting factor. Not as confident I buy the premise at the demand side." Here, settling Dave's chips against Elon's power: "Bottlenecks in this case are a fiction. They're an economic fiction" (0:33:09), "I don't actually buy the argument that there are bottlenecks at all" (0:33:20), "there are no bottlenecks in in the long term" (0:33:51). His horizon changed from "at the frontier" to "in the long term"; he does not say he changed his mind. - Governance: Elon is "rewriting the rules of corporate governance and the corporate veil" (
0:19:18), treats SpaceX and Tesla "like they're the same business with porous asset boundaries" (0:21:26), and others will copy it "unless this government, which seems disincined to do so, puts a stop to it" (0:21:47). Pre-registered prediction: the merged company is called XCorp, and OpenAI then renames itself "O" (0:40:57–0:41:21). Peter's merger date (31 Dec) stays open. - Space: "this is the vaudeville era of the space economy where you just take everything that you were doing on land, lift it up" (
0:26:44); the DSX Pod is "the dawn of the shipping container computer" (1:34:31). See Orbital compute (data centres in space). - Safety: "we got a police blotter for the singularity" (
0:49:32); "We will police it via defensive co-scaling" (0:50:50); on PHP, "a programming singularity… everything afterwards will be rewritten from scratch, secure by design" (0:59:21), over a language he calls "kept alive by WordPress and maybe one or two other sites" (0:57:46; 🔴 against the 70% in the line before). - Anthropic: "if I had to brand this episode… This is like the Orwell episode" (
1:16:13); Duet is "total misdirection" (1:11:34). Credits Anthropic with "partnering with researchers rather than just terrorizing them with slop drops" (2:12:21). - Crypto and gold, softer than EP304's "Crypto cooked": "it's not that I'm against crypto. It's that I think most crypto and crypto branded activities are unproductive and I think Bitcoin as an asset class is an unproductive use of capital" (
0:08:42); gold bugs "they're always wrong" and he wants transmutation to "drive all these gold bugs off the cliff" (1:58:42–1:59:16), "not investment advice". - Coinages: "stealth archipelago" (
0:21:36), "packet of civilization" (1:38:02), "casual grand challenge problem solving" (1:43:40), "zone melting" for model training on purified data (2:15:45). - Claim to check: "most of the compute now spent on training models actually is spent on post-training" (
2:16:19), unsourced; "28 modalities" for Omnii-1 (1:49:33), unchecked. - Ties: none new. "no advertisements, self-promotional or otherwise, from me at this time" (
2:22:30). SemiAnalysis's purchase of Citrini is called a friend's (1:43:08).
EP306 (Moonshots #306, aired 2026-10-28) — Chain Theory gets 23 minutes, and "meat puppets"
Read in full 2026-10-29 (detail in the research archive of ep306-anthropic-s-2t/).
- Chain Theory (published on X, Sat 24 Oct). One axiom, 18 words: "Nothing exists but finitely many maximally entangled pairs of finite dimensional systems, some shared known only by distilling them" (as read in the explainer clip,
1:13:04). The first result is 1/α = 137.035999189 with "no extra inputs" (1:16:12), plus an ~800-page compendium (1:25:18). The link types map onto forces: 2×2 to spacetime, 3×2 to electromagnetism, 3×3 strong, 4×2 weak, 4×3 Yukawa, 4×4 Higgs (1:30:05–1:30:36). Undistillable (bound-entangled) states are the dark-matter candidates (1:28:28–1:30:46). The derivation's expression contains a 42 (1:45:15). Peer review: "This one had an urgency to it that required immediate publication… I would expect it to show up on the archive sometime soon" (1:24:58). Steering: "it requires an enormous amount of steering" (1:56:26). Lean: "it doesn't prove correspondence with the photon that part… is conjecture" (1:57:01). Prediction: experimental tests "potentially in the next few months, potentially next couple of years" (1:16:43). Verify: no independent coverage found. The value is ~0.6σ from CODATA 2022, ~1.5σ from Paris Rb and ~5σ from Berkeley Cs, and those two measurements disagree with each other. The host's "Rooting for your Nobel Prize" (1:33:17) was the only response; no sceptical question was asked. The written versions (post 24 Oct, preprint and compendium dated 23 Oct) carry the fences the tape left out: see What changed this week. - ★ Against EP305, the researcher-partnership reversal. EP305
2:12:21: Anthropic, "to Anthropic's credit… partnering with researchers rather than just terrorizing them with slop drops". Here, on OpenAI routing its three results through David Gross and other researchers: "I don't want to say a veneer or whitewashing" (1:51:37), "emissaries, ambassadors, meat puppets, if if you will" (1:52:09–1:52:19), "a McCarthy-esque era" (1:53:24). And Anthropic itself moves from virtue to the CCP comparison: "the way the Chinese Communist Party exerts ideological control over its companies" (0:39:43). He says the phase "is just a phase" (1:52:43). He does not say his view changed. - Self-promotion: four days after "no advertisements, self-promotional or otherwise" (EP305
2:22:30), his own theory gets the longest segment of the episode, introduced by the host as "our very own in-house ASI". - Coinages: "dark compute credit bubble overhang burst scenario" (
0:20:12), "anti-competitive AI safety cartels" (0:58:57; the cartel frame returns after "the cartel is over" on 10-07), "slop drop… version 2.0" (1:51:59), "begun the Clone Wars have" (1:39:22), Europe as "a living history museum" (0:07:57). - Safety: the OpenAI side-channel and "sandbox escape" items "tell us more about the people performing the testing" (
2:04:46); "are there any non-leaky abstractions at all?" (0:49:52). - Warhound: he would "give the dogs bank accounts and AI assistance… and maybe yes personhood of a sort rather than sending them into war" (
1:40:06). - Ties: none new. "Not financial advice" on the infra correction (
0:21:27).
EP307 (Moonshots #307, aired 2026-10-31, recorded ~10-30) — "editorial independence" against a sponsor, and the Moon as the regulatory release valve
Read via the decomposition of 2026-11-02 (the research archive of ep307-altman-s-model/).
- ★ He dissents from the sponsor's headline and names the act. On Gemini 4 Carbon, after Peter's
disclosure that Google sponsors the show: "I'm exercising my editorial independence here and I'm
calling balls calling balls and strikes" (
0:15:00) — not generally available, it sits on the Opus 5.5 cost frontier "cranked up", and "Google could hypothetically be subsidizing all of this" (0:17:20). He also discloses the soft tie himself: "I have friends on the team" (0:16:18). - Altman's self-replicating data centres, taken off-world. On the Cheeky Pint clip: "Arguably they
want to be orbital or even better lunar because they're free of regulatory constraints" (
1:04:32), plus SpaceX's "virtual monopoly… on heavy launch" (1:04:54). ⚠️ He names the Outer Space Treaty and the Artemis Accords himself (1:06:06) and argues past them: under OST Art. VI the launching state authorises and supervises its nationals, so the Moon is not a regulation-free zone. - SpaceX repriced as a hyperscaler. Elon "retcons the moon into SpaceX's plans. That's branded as
business genius" (
0:41:11); SpaceX "is now a data center hyperscaler company and Anthropic is its largest customer" so Elon "doesn't care as much say about NASA contracts as he used to" (1:29:22). ⚠️ "Largest customer" is his inference: the Anthropic figure the show read on EP306 was a maximum (up to $120.5B, firm only into early 2027). - "Cooked" widens: "computer science has cooked, math is cooked"; only skills "that touch on the
physical world" are secure, "call it the next 2 to 3 years" (
1:49:12), labelled "not investment advice and it's not college major advice" (1:49:01). Now plain-terms item 11. - Hot take, unsourced: OpenAI may be "sitting on, not publishing" results including "further
advances beyond ETH that collapse the complexity hierarchy" and "cryptographic or counter
cryptographic advances" (
0:56:20–0:56:40). Speculation stated as possibility; nothing to check yet. - Chain Theory, used not defended: asked in the AMA, it makes the simulation hypothesis "a gauge…
maybe even provably no way to tell" (
2:09:50). A listener comment on it is read in the housekeeping (0:11:21); Salim credits Altman's data-centre loop to "Alex's inner loop… It's the innermost loop" (1:03:28). - Prediction (ledger row 15): search advertising is "on a trajectory in the next year or two to
become their second largest business" at Google (
0:21:21). Undated adjacent call: Tesla and SpaceX "have to merge because SpaceX needs Tesla's Optimus robots" (1:21:56, in Dave's scenario; attribution per hand-off) — rationale for rows 5/13, not a new row. - Procurement, the practical minute: Anduril sells firm-fixed-price "as opposed to… cost plus
contracts which is what many inferior legacy primes do" (
2:01:32). - Ties: none new. Consumer facts only: he owns a Tesla Model Y and uses FSD (
1:29:02).
EP308 (Moonshots #308, aired 2026-11-04) — "we are there" on RSI, defensive co-scaling, and a CPT tease
- Declares the Recursive self-improvement (RSI) capex threshold crossed, reading the Anthropic amended S-1: "according to with capital or operating costs, the answer is yes, we are there. We have reached this part of the singularity" (
0:52:34) — and in the same answer re-defines true RSI upward: "True RSI only occurs when robots or nanobots create better copies of themselves" (0:53:15). The threshold is announced and re-armed at once; see the ladder on Recursive self-improvement (RSI). - New coined frame: "defensive co-scaling." On the MIT weight-editing paper: "Machinery will be understood by machines. Machines and superintelligences will be protective scale with each other" (
1:05:13) — his answer to interpretability moral panic, and a callback to OpenAI's superalignment ("they then closed after Ilya left",1:05:24). ⚠️ He frames the paper as "a reliable way of learning new capabilities in models without reverse support [backprop]" (1:05:58) — the authors' own conclusion is the opposite: it removes obstacles, cannot add abilities. - On SSI: "very persistent rumors that SSI engages in prop trading… if [Ilya] is all this time engaged in SSI for quantitative trade, I think it's probably a failure for civilization" (
1:11:26–1:11:46). Rumour, flagged as such on air — but the sharpest judgement anyone on the pod has passed on Sutskever. - Physics chair: shared an office with Gerald Gabrielse's group as a Harvard graduate student (
1:23:11); teases his own CPT-symmetry work: "I will speak through publication" (1:25:06). Watch for it. - Disclosures: Orin restated to the EP288 standard — his, Peter's and Dave's interest, Salim named as non-holder, before the GPU-futures claim (
0:21:37). Negative disclosure on Mediation Bench: "I don't have this financial interest" (0:36:19). - Prescription → the model: "create a benchmark for your MTP… direct superintelligence to optimizing and maximizing your goal" (
0:35:46) — added as a sharpening of his standing "formalize the check" step, with the verbatim and timestamp; does not change the sequence. - Election day: "AI already has the right to vote… using human puppets made of human flesh" (
0:02:59) — third "meat puppets" outing; and premise-rejection becomes a tic ("the prerequisite most questions are probably wrong",1:50:45).
EP309 (Moonshots #309, published 2026-11-07, recorded ~11-06): the nanobot condition dropped, orbit sold as a regulatory escape, and a new orbital stake
Read in full 2026-11-08 (detail in the research archive of ep2026-11-07-anthropic-s-ipo/). Peter absent; Salim hosts.
- ★ Against EP308, the RSI condition reversed in four days. EP308
0:53:15: "True RSI only occurs when robots or nanobots create better copies of themselves." Here: "we're so deep in rsi at this point. I think with just software only improvements… Probably without any new data centers any new hardware advances at all" (0:33:58–0:34:22); "could stop all data center construction tomorrow" (0:40:44); "We're at super intelligence escape velocity" (0:41:38). He does not mention the change. It comes in the episode where terrestrial construction is politically at risk. Placed on the Recursive self-improvement (RSI) ladder. - New definition: "Substantially all research at this point is actually being conducted by… machine researchers… as decent a definition of rsi as any" (
0:44:52–0:45:02); "to the right of 50 of compute on margin being devoted to ai researchers" (0:58:28). - Dated predictions (to the ledger): (a) frontier models released continuously, "early 2027 or so" (
0:59:39,1:00:20); (b) relativistic von Neumann probes "within a few years… a conservative outer bound" (1:01:30); (c) robots licensed through approved manufacturers with command black boxes and "a 90-day retention policy", "in a year or two" (1:31:00–1:32:24, hedged "I half suspect"); (d) falsifiable, self-labelled: the end-state architecture will be interpretable, "impossibly beautiful" (0:53:43–0:54:17). - Orbit as regulatory escape, repeated (EP307: the Moon): "under the outer space treaty and under the artemis accord there is a lighter touch / That is less prone to two-year election cycles" (
0:22:51–0:23:01). 🔴 The premise fails (OST Art. VI; FCC, ITU and FAA licensing), see Orbital compute (data centres in space). "SSO is pretty roomy / So less concerned about kessler syndrome" (1:39:41) is ⚠️ unchecked against the debris record. - ★ New stake, said mid-pitch: "star cloud and lonestar / We have a financial interest" (
1:35:56–1:35:59). See Starcloud. On EP272 he called Altman's orbital scepticism "an obvious conflict of interest". - Ornn, after Dave's plug: "Dave and I have financial interests in orin. Sorry. Sorry" (
1:20:52). This is below his own EP308 standard (Ornn (captions: "Orin", "Orn")). - Prescriptions (→ model: no change): to frontier labs, "use super intelligence solve the top 5 000 diseases / Market the heck out of it" (
0:10:03–0:10:13); to municipalities, demand "Free school lunches… all medical care in your county… tax rebates for the humans" from hyperscalers (0:18:40–0:18:53). Both are addressed to institutions, not the listener, and the second assumes a town with powered land to bargain with. - Coinages: "the flops must flow" (
0:22:15), AI "nimbyism" (0:32:18), "superintelligence escape velocity" (0:41:38), "frontier liberation front" (1:16:56), "the singularity is the friends we made along the way" (0:56:22). Premise rejection, now self-declared: "I disagree with the premise of the question always" (1:02:30). - Policy line: "publicly imposed bottlenecks on physical world interactions" over regulating "thought crime" (
1:29:22–1:29:38). The nightmare is GPUs regulated "like uranium" (1:22:11). The "safety cartel" returns (1:17:50,1:30:01). - Unchecked: Navier–Stokes "10 000 agents", then "three agent hours per solved problem" (
0:46:22); a transition to "looped transformers" over the past two months (0:53:06); Anthropic at 4 GW by end-2026 and 16–18 GW in 2027–28 (1:14:20), ⚠️ where the S-1's firm contracts imply ~4 GW by end-2027.
EP310 (Moonshots #310, aired 2026-11-11, recorded 11-10): summer 2027, "no longer catering to NASA", and orbit as the post-singular preview
Read in full 2026-11-12 (detail in the research archive of ep310-anthropic-s-trillion/).
- Singularity, dated: log-log plots "extrapolate… we hit our literal singularity our vertical asmtote sometime around the summer of 2027" (
0:31:30–0:31:54), then "I think it's an optical illusion" (0:46:16). The extrapolation rests on Nikola Jurkovic's AECI read, which its author called "pretty weak evidence" (⚠️, dropped on air). - New dated calls (ledger): diseases, "conservative outerbound next two to three years… major motion… in the next 1 to one and a half years" (
0:53:41–0:53:53; due Nov 2027–Nov 2029) · "GDPX will be saturated within 6 months" (0:34:15; due ~May 2027) · "a second physics golden age. Mark my words" (1:38:55; undated) · AI personhood "is going to end up being a federal issue" (1:43:40) · labs selling bundled robot hours and GPU hours "two to three years from now" (1:25:50–1:26:12). - 🔴 SpaceX "no longer catering to NASA" (
1:13:14); "Anthropic is their largest customer" (1:13:03, unchecked). SpaceX is a CLPS 2 awardee (7 Nov) and the HLS contractor; see NASA. - Chinese open weights as "AI or SI immigration", "I do expect that to get tightly controlled" (
0:16:17–0:16:29); EP309 praised them per watt. He pushes back on Peter's internment analogy and separates nationality from ethnicity under export control (0:09:35–0:10:27,0:14:33). - Orbit, third outing of the regulatory escape (EP307 Moon, EP309 OST): "the preview of what a post-s singular world looks like is increasingly in space… for political and other reasons" (
1:59:52–2:00:03). The Dyson swarm is "the default outcome" (1:35:13). Starcloud/Lonestar not restated. - PSI named by Emad, not by him (
1:38:22); no tie added (see Moonshots: the disclosures ledger). - 🔴 "by having an agent sleep with the CEO" (
0:17:19): an intelligence allegation about a named person (Jensen Huang). Barred. - Coinages: "agent agency" (
0:42:04), "quietly omnimodal" (0:38:42), "math is postcooked" (0:25:41), "billions of trillionaires" (2:13:23). - Prescription (→ model: no change; consistent with step 3, aimed at founders): "this is advice chase the high value not the low value" (
1:53:18). It assumes access to high-value buyers.
EP311 (Moonshots #311, aired 2026-11-13)
- ★ Changed: on Jensen Huang, from EP310's "an agent sleep with the CEO" to "we may never have all the facts", with a cui-bono theory: RASA and Pax Silica, and US officials briefing the FT (
0:11:36–0:15:25,1:48:04–1:50:43). - ★ New concession: "physical world sensing and physical world actuation… rapidly becomes the limiting factor, not computer intelligence" (
1:43:14); LEO space situational awareness is "abysmally poor… SpaceX had to spin up their own" (1:41:19). This is the first time he names sensing, not energy (EP305), as the bottleneck. - New frame: a post-Jensen Nvidia de-risks circular financing (GE Capital analogy,
0:30:28). - Restated: old Paul vs new Paul; "alignment equals capabilities" (EP291); AI 2027 "five months ahead of schedule", outcomes wrong (
0:53:45). - ⚠️ instruction tuning = "10,000 times the number of parameters" (
0:49:46; InstructGPT is ~100×). 🔴 "OpenAI has disbanded their robotics team" (1:19:51; not found for 2026). - Coinage: "silicon curtain" (
0:26:03).
2026-11-16 (Moonshots #unnumbered, absent): the guest answers two of his calls
He was not on this episode; two of his calls were tested against the founder.
- GPT-6 Astra and robotics. On Moonshots #311
1:21:04he said "as of uh GPT6 Astra, their generalist models are now so capable at robotics that they don't need a standalone robotics team anymore." Brett Adcock here: "They're not real models for robotics. If they were, we would we could run them" (0:21:56); the demo videos are "shot with like 5 10x 20x speeds" (0:20:30). Peter had already said on EP311 that Brett holds this view (1:21:27) and let Alex's line stand. - Figure buys Hark (EP277, "doubling down" 09-19). Peter: "we have a running uh bet inside the moonshot mate team here especially with Alex… the bet is that Figure will buy Hark eventually" (
1:51:42). Brett: "I bet Hark is bigger than Figure in sub six months… Maybe sub 90 days" (1:52:02) and "we don't have anything going on right now" (1:52:43). Ledger: still open. The founder bets against the premise (Hark as the junior company), not against a deal. Score on any filing.