Max Tegmark

Who he is

Affiliation: Professor of Physics, MIT (MIT Physics faculty page: "the main focus of his current research is on the physics of intelligence"; FLI's bio adds the Institute for AI & Fundamental Interactions and the Center for Brains, Minds and Machines). Founder and Chair, Future of Life Institute (FLI) — role check run 2026-09-21: title changed since the last review and the change was missed. FLI's own quote line on the 2025-10-22 superintelligence statement still reads "Future of Life President Max Tegmark" with Anthony Aguirre as "co-founder and Executive Director" (FLI press release); the FY2024 Form 990 (the annual tax return a US nonprofit files) lists him as President (ProPublica, filed 2025-11-07). By FLI's 2026-02-27 statement he is "Founder and Chair of the Future of Life Institute" (FLI), and by 2026-03-04 Aguirre is "FLI President and CEO" (FLI). No announcement was found; the handoff sits inside that window. FLI's site is still inconsistent — the Summer-2026 Safety Index page labels his video "FLI's President" (FLI) while the People page and his profile say Founder and Chair. He is also Board President of the Lightcone Foundation and a Director of the Future of Life Foundation — the two related entities that hold the endowment (see Revealed behavior). Serial nonprofit founder: Foundational Questions Institute (with Aguirre), Improve the News Foundation (FLI bio). Author of Our Mathematical Universe (2014) and Life 3.0: Being Human in the Age of Artificial Intelligence (2017). Born 1967-05-05, Stockholm; PhD Berkeley 1994 (Wikipedia). TIME100 AI in 2023 (FLI bio) and again in 2026 (TIME, 2026-08-27, listed as "Founder and Chair, Future of Life Institute; MIT professor"). MIT status: no leave, retirement or emeritus notice found (dated search 2026-09-21).
One-line position (v1, carried): AI poses a genuine existential risk that requires organized, mainstream advocacy — including, if necessary, pausing frontier development until safety catches up — and physicists have a responsibility to engage because we understand exponential processes.
2026 revision of the ask: the demand escalated from a six-month pause (2023) to a prohibition on superintelligence "not lifted before there is broad scientific consensus that it will be done safely and controllably, and strong public buy-in" (Statement on Superintelligence, 2025-10-22), and the positive program is now "tools, not beings" — "If we can continue AI development of controllable tools, we're going to see a golden age" (CSER lecture, Cambridge, 2026-09-16, reported by machine.news, 2026-09-17). His own number: "if we go ahead and continue having nothing like the FDA for AI... I would think it's definitely over 90% that we lose control" (Doom Debates vs Dean Ball, 2025-11-21).

Discipline & technical bet

Trained as a precision cosmologist, he re-pointed his MIT group at "the physics of intelligence" and now runs a mechanistic-interpretability (reverse-engineering what a network computes) and "guaranteed safe AI" research line: Provably safe systems: the only path to controllable AGI with Steve Omohundro (arXiv 2309.01933, Sept 2023), the multi-author Towards Guaranteed Safe AI framework (arXiv 2405.06624, May 2024), and Kolmogorov–Arnold Networks, a more inspectable network architecture where he is 8th of 8 authors (arXiv 2404.19756, v1 2024-04-30). The in-window technical wager is Scaling Laws for Scalable Oversight (Engels, Baek, Kantamneni, Tegmark; arXiv 2504.18530, v1 2025-04-25, v3 2025-10-27, NeurIPS 2025 Spotlight; he is last author): oversight is modelled as a game between mismatched players with "oversight-specific Elo scores" (a chess-style strength rating), and the paper reports that "Our oversight success rate estimates at a 400-point general Elo difference range from 9.4% for Wargames to 51.7% for Debate." The same paper names the quantity he wants labs to publish: "We refer to p as the Compton constant for ASI... because of the apocryphal story of Arthur Compton and colleagues estimating the probability to be less than 3×10⁻⁵ that the first nuclear explosion would ignite the atmosphere" (v3 full text). He is also 28th of 33 authors on A Definition of AGI (Hendrycks et al., arXiv 2510.18212, v1 2025-10-21), whose abstract scores "GPT-4 at 27%, GPT-5 at 57%" of the way to AGI — numbers he cites on air as "a paper I co-authored" (byline position logged per the research-lead rule). The bet, in one line: the control problem is measurable (compute a Compton constant before you train) and, if the number is bad, avoidable (build tools, not agents) — and the lever for making labs do either is political, not technical.

Key claims (Says)

• AI existential risk is real and warrants organized response (v1): FLI is the institutional expression of this claim — moving x-risk from academic philosophy into policy advocacy. Held on mechanism — the labs themselves now say it in his vocabulary: Anthropic, "We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development" ("When AI builds itself", published 2026-06-04 — Fortune, 2026-06-05, dates it "Thursday"); Dario Amodei, "We must slow the pace at which we improve the capabilities of AI models" ("We Must Pace the Frontier", 2026-09-12); Sam Altman, "I agree with Dario that we need to pace the frontier" (NBC, 2026-09-12). Valence graded separately under predictions: the labs' pledges are conditional on everyone else, and FLI's own July-2026 index found the same companies had weakened their pause commitments.
• The 2023 "pause" open letter (v1): FLI organized the letter calling for a 6-month moratorium on training systems more powerful than GPT-4 (2023-03); gathered 30,000+ signatures including Musk, Wozniak, Harari. Held, numbers trued — the letter page shows 31,810 signatures and the sentence "We call on all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4" (FLI, 2023-03-22); Musk, Bengio and Wozniak are confirmed signatories (Wikipedia); Harari's signature was not re-verified this cycle (the list did not render).
• "Life 3.0" (v1, 2017): framework of three stages — Life 1.0 (biological), Life 2.0 (cultural), Life 3.0 (designs its own hardware and software); AI is the bridge to Life 3.0. Unfalsifiable as a taxonomy — noted, not graded. One observable now exists for the "designs its own software" half: "As of May 2026, more than 80% of the code we merge into Anthropic's codebase was authored by Claude" (Anthropic).
• The alignment problem is a physics problem (v1): ensuring complex systems optimize for the right objective function — a control-theory framing accessible to physicists and engineers. Drifted — he kept the quantitative framing but his own lab's result is that oversight of a 400-Elo-stronger system succeeds 9.4%–51.7% of the time depending on the game (arXiv 2504.18530), and his 2026 message has moved from "solve control" to "don't build the thing that needs controlling": "No one currently has any convincing plan for controlling superintelligence" (CSER, 2026-09-16). The physics register survived; the tractability claim did not.
• Humanity should choose its future deliberately (v1): the "cosmic endowment" framing — life's long-run potential in the universe is at stake, and AI is the pivotal technology. Unfalsifiable — normative; not graded. Still his live frame: "We know that the future of AI is not inevitable. It is for us to decide" (Pro-Human Assembly, Boston Globe, 2026-09-15).
• NEW — "definitely over 90% that we lose control" if there is "nothing like the FDA for AI" (Doom Debates, 2025-11-21); earlier, his group "found a 90% probability that a highly advanced AI would pose an existential threat" and defined the Compton constant as "the probability that an all-powerful AI escapes human control" (Guardian, 2025-05-10; text read via the inkl syndication — the Guardian page did not fetch). Graded as a (number, condition, horizon) triple: number 90%+, condition "no pre-deployment regime", horizon none stated. Not gradeable until he attaches a date; check under predictions. Note the Aug-2026 restatement (Roman Yampolskiy interview, YouTube) could not be fetched and is not used as a source.
• NEW — "AI is the only industry in the U.S. that's 'less regulated than sandwiches'" (Democracy Now, 2026-07-30; repeated as "less regulations than sandwich shops" at CSER 2026-09-16). Held on the observable — there is no US pre-deployment approval regime for frontier models; the federal posture is Executive Order 14365, "a minimally burdensome national policy framework for AI," which orders an "AI Litigation Task Force... to challenge State AI laws" (White House, 2025-12-11). His specific asks are unchanged since 2023: "independent audits and licensure before deployment," "the burden of proving suitability for deployment should fall on the developer," "a centralized federal authority," and developer liability (written statement to the Senate AI Insight Forum, 2023-10-24).
• NEW — China is the lobbyists' talking point; "China has already cracked down harder on its companies than America has on our companies. They cracked down hard on AI girlfriends and other anthropomorphic AI, just weeks ago" (Democracy Now, 2026-07-30). Held on the instance — the Interim Measures for the Administration of AI Anthropomorphic Interaction Services were "issued on April 10, 2026, by the Cyberspace Administration of China and four partner agencies" and took effect 2026-07-15, fifteen days before he said it (Just Security, 2026-08-06). The market-level half ("harder than America") holds for consumer-facing AI; for frontier training no Chinese gate was found, and DeepSeek scored F on FLI's own index (below). Mechanism and valence logged separately.
• NEW — "I haven't seen as much shift in U.S. policymaking in the past decade, as I have in the past three months" (TIME, 2026-08-27). Held on bills introduced, unresolved on outcome — AI Kill Switch Act (Lieu/Moran, 2026-07-23), Frontier Act (Obernolte/Trahan, July 2026), Ban Artificial Superintelligence Act (Sanders/Casar, announced 2026-09-03, "not yet formally introduced" as of 2026-09-16), Stop Rogue AI Act (Gottheimer/Lawler, 2026-09-09) (Poynter, 2026-09-16). Nothing has passed, and the President called the concerns a "HOAX" in seven Truth Social posts on 2026-09-14 (CBS, 2026-09-15).
• NEW — FLI's instrument claim: no frontier lab scores above C+ and four of them "have weakened or voided pledges to pause unilaterally if redlines are approached, some citing competitor-contingent conditions" (AI Safety Index, Summer 2026, July 2026: Anthropic C+ 2.66, OpenAI C 2.28, Google DeepMind C 2.01, Meta D+ 1.32, Z.ai D- 0.88, Alibaba Cloud D- 0.87, xAI F 0.65, DeepSeek F 0.47, Mistral F 0.33; 37 indicators, 6 domains; panel Krueger, Li, Maharaj, Revanur, Russell, Trager, Zeng). Prior edition for trend: Winter 2025 (Dec 2025) had Anthropic C+ 2.67, OpenAI C+ 2.31, DeepMind C 2.08, xAI D 1.17, and found "All of the companies reviewed are racing toward AGI/superintelligence without presenting any explicit plans for controlling or aligning such smarter-than-human technology" (FLI). Logged as a claim because FLI both designs and publicizes the instrument (see Incentive map); the pause-pledge finding matches the bench's Kaplan and D. Amodei files (RSP v3.0 removed the pause pledge).

Notable predictions — with falsifiable checks

• (2017) AI will be the most transformative technology in human history — v1 outcome: directionally supported; scale of transformation TBD. Unfalsifiable as written — rewrite or retire. Proposed rewrite for the next cycle: "By 2030, more than half of US firms report AI in production use (Census BTOS)"; grade on that dataset, not on adjectives.
• (2023-03) Pause letter: called for 6-month moratorium on frontier training — v1 outcome: no pause occurred; the letter raised awareness but had no enforceable effect. Held on the observable (no lab paused, 2023–2026) — valence Drifted: three years on, three frontier-lab CEOs adopted the vocabulary — Anthropic's conditional "we expect that we would slow down or temporarily pause, if other developers at or near the frontier also did so in a verifiable manner" (2026-06-04), Amodei's "We must slow the pace" (2026-09-12), Altman's "I agree with Dario" (2026-09-12); Hassabis and Musk "quickly cosigned" per Reason, 2026-09-14 (reported, no quotes fetched). Says and Does diverge: FLI's July-2026 index says the same firms weakened their pause pledges. Check: any frontier lab verifiably slows or pauses a training run under a stated condition by 2027-06-30; if the pacing talk produces no observable by then, regrade the valence back to Held-for-v1.
• (Ongoing) Without deliberate safety intervention, the default AI trajectory is catastrophic — v1: unfalsifiable in the near term; functions as a motivating claim. Unfalsifiable — rewrite or retire. His own Compton-constant proposal supplies the rewrite: "at least one frontier lab publishes a loss-of-control probability before a frontier release by 2027-12-31." Status 2026-09-21: none found (dated search).
• (2024-11-13, Web Summit Lisbon) "It's a suicide race where whoever gets AGI first is going to lose control to this future machine species" — and the sequence: US and China first implement independent national safety standards, "then they'll talk to each other... then they'll push the rest of the world to sign an AGI moratorium treaty" (Web Summit summary). Drifted — the sequence has started at step two without step one: the US has no national safety standard and has stood up a task force to challenge states that write them (EO 14365, 2025-12-11); China's 2026 measures are consumer-facing, not a frontier gate; yet a formal US–China AI dialogue is on the agenda for the 2026-09-24 Trump–Xi summit, with an AI-incident "notification mechanism" proposed by Treasury (CSMonitor, 2026-09-21). Check: (a) does the 2026-09-24 joint statement mention frontier-AI loss of control; (b) any enacted US federal frontier-safety standard by 2027-12-31; (c) any signed multilateral moratorium instrument by 2028-12-31.
• (2024-11-16) "I think it will make a big difference whether Trump will listen more to Elon [Musk] on this or more to the anti-regulation intentions" — Musk framed as the safety channel into the new administration (Euronews). Wrong on valence — the anti-regulation side won (EO 14365; "HOAX," 2026-09-14), Musk's own lab scored F on FLI's index (July 2026), and "Top tech executives, who have offered predictions about superintelligence and signaled that they are working toward it as a goal, didn't sign the statement" (CNBC, 2025-10-22, discussing Zuckerberg, Musk and Altman). Musk did reportedly cosign Amodei's pacing call in September 2026 (Reason) — a rhetorical return, not the policy channel Tegmark hoped for. Pattern note: this is the third time the file records Musk endorsing FLI's ask while building the thing it targets (2015 funder, 2023 signer and xAI founder, 2026 reported cosigner of a pacing pledge); a forward check is attached under Weak spots.
• (2025-05-10) "A Compton constant consensus calculated by multiple companies would create the 'political will' to agree global safety regimes" (Guardian). Live, unconfirmed. Check: at least two labs publish a pre-release loss-of-control probability by 2027-12-31; zero as of 2026-09-21.
• (2025-11-21) "The Chinese Communist Party and Xi Jinping clearly like control... [they] would never permit a Chinese company to build technology if there were a significant chance it could overthrow them" (Doom Debates). Live. Check: China imposes a frontier-capability gate (compute threshold, licensing, or pause authority) — not a consumer-content rule — by 2027-12-31. Counter-evidence to watch: DeepSeek F on FLI's index (July 2026).
• (2025-10-22) Prohibition on superintelligence "not lifted before" scientific consensus and public buy-in; the statement's own poll: "73% support robust regulation," "64% believe superintelligence should not be developed until scientific agreement on safety," "5% support current unregulated development" (FLI press release). Live, partly observable. 75,090 signatures on the statement site (fetched 2026-09-21); a sitting senator's bill now tracks the ask — "Banning AI superintelligence so no person or entity may develop or deploy Superintelligent AI systems," a pause "until we have clear safety rules in place," a cabinet-level agency, up to 20 years' imprisonment (Sanders, 2026-09-03). Check: the Sanders–Casar bill receives a committee hearing by 2027-06-30; any G7 government enacts a superintelligence prohibition by 2028-12-31.
• (2026-08-27) An American public "wary about job security, data center construction, and AI use by minors could make AI safety a hot-button issue in the upcoming midterms" (TIME). Live. Check: 2026-11-03 exit polling ranks AI among top-five issues in any state; FLI or a related entity reports electoral spending (wake trigger).

Revealed behavior (Does)

• Money — spending the endowment down at ~$46M a year with almost no new income. FLI's 2025 expenditure was $46,192,000 (grants to charities $26,724,000; contract and professional fees $6,457,000; media and outreach $5,795,000; personnel $5,019,000; events and travel $1,722,000; office $475,000) against "contributed income... approximately $37,000" and "investment income... approximately $152,000" (FLI Finances). The money comes from related entities: FLI's own 990 shows revenue of $549,531,672 and assets of $537,164,374 in 2021, then expenses of $379,048,482 and negative revenue of −$85,626,657 in 2022, and assets of only $6,348,222 by 2024 with revenue $21,273,381 (ProPublica). The Lightcone Foundation (Sparks, Nevada, EIN 92-0636259) booked revenue of $163.3M in 2022 and held $174.4M in assets at end-2024 after $34.1M of 2024 expenses; its officers are Tegmark (Board President), Aguirre (Board Secretary), Meia Chita-Tegmark (Treasurer) and Jaan Tallinn (Director) (ProPublica, 2023–24 returns filed 2026-05-05). The Future of Life Foundation (Campbell, California, EIN 88-0926067) held $26,666,125 in 2023 with the same four officers (ProPublica). An analysis of the 2022 Schedule I reports the December-2022 transfers as $180.3M to "FLI Europe," $162.6M to Lightcone and $25M to the Foundation, $368M in all (aipanic.news — secondary; the schedule itself did not render this cycle, and no separate "FLI Europe" legal entity was located). Not the "Lightcone Infrastructure" that runs LessWrong — different organisation, same word.
• Pay. FLI's 990 lists no compensation from FLI itself; it reports $457,840 (FY2024) and $409,755 (FY2023) from related organisations, and the Foundation's own return shows $204,877 of the 2023 figure as "Director" pay (ProPublica). Plus the MIT salary. All three entities' return summaries flag "conflict of interest transactions"; the details were not read this cycle.
• Time — politics first, research second. In-window output: the Statement on Superintelligence (2025-10-22), the Pro-Human AI Declaration (New Orleans, press release 2026-03-04), a Pro-Human Assembly in Washington on 2026-09-15 carried on C-SPAN with Sanders, Bannon, Blumenthal, Blackburn, Casar, Chip Roy, the AFL-CIO, AFT and SAG-AFTRA presidents (agenda; speaker list per WSWS, 2026-09-17, a partisan outlet used only for the attendance facts and his two quoted lines), Davos with Harari (Bloomberg House, 2026-01-24), Democracy Now, TIME, a CSER keynote. Research output over the same period: one NeurIPS spotlight as last author, one 28th-author position paper. Priority graded by where hours went: the campaign.
• Hires — FLI is a lobbying shop now. The People page lists a policy division of roughly two dozen named staff including a Chief Government Affairs Officer, a Chief State Affairs Officer, a Head of U.S. Policy, a "Labor Outreach and Coalition Builder," a Korea Representative and a Middle East Advisor (FLI). In Brussels: registered 2021-06-24, 2.5 full-time-equivalent lobbyists, "annual spend on EU advocacy amounts to €273,176 EUR. This includes staff salaries and spending on Dentons Global Advisors" (lobbyfacts.eu). US federal lobbying totals exist on OpenSecrets but the site refused three fetches; the dollar figures are deliberately not carried from search summaries.
• Stepped from President to Founder and Chair between 2025-10-22 and 2026-02-27 while Aguirre became President and CEO — operations consolidated under Aguirre, Tegmark kept as the public face (see Who he is).
• Lost his largest donor's endorsement. On 2026-03-13 Vitalik Buterin wrote that FLI had made "an internal pivot" toward "cultural and political action as a primary method, quite different from the original approach," and "My worry is that large-scale coordinated political action with big money pools is a thing that can easily lead to unintended outcomes, cause backlashes, and solve problems in a way that is both authoritarian and fragile"; he had expected FLI to cash out "$10 million to $25 million at most," versus roughly $500 million realised (The Block, 2026-03-13; CoinDesk, 2026-03-14: "FLI did not immediately reply to CoinDesk's request for comments"). His X post itself was not fetchable.
• Praised Anthropic and OpenAI when they held a line — "We highly commend Anthropic, OpenAI and leading researchers from across AI companies for standing up for the principle that AI should never be used to kill people without meaningful human control, and that domestic mass surveillance of US citizens is a red line" — while insisting "our safety and basic rights must not be at the mercy of a company's internal policy" (FLI, 2026-02-27, on Anthropic's refusal of a Department of War ultimatum). Consistent with the file's split: he grades the labs harshly and still uses their good days as evidence.

Feels

Identification with Compton and Oppenheimer — the physicist who does the calculation the builders will not. Fear that the thing has already happened: "It's no longer hypothetical to say maybe one day we'll lose control over AI. This is in the rearview mirror" (CSER, 2026-09-16, after the July OpenAI-agent incident). Hope in a "golden age" of controllable tools. What he wants to be true: that the public and both parties are already his — "95% of Americans don't want a race to superintelligence, and experts want to ban it" (FLI, 2025-10-22).

Hears

A five-person board he has chaired or sat on since 2014 (himself, Tallinn, Aguirre, Chita-Tegmark, Krakovna — FLI People page); the Hinton–Bengio–Russell axis (statement signatories; Russell sits on his Safety Index panel); Hendrycks's CAIS circle (the AGI-definition paper, which also carries Tallinn, Schmidt, Marcus, Brynjolfsson, Bengio); Katja Grace's expert survey, which FLI funds — "We are grateful to Coefficient Giving, Jaan Tallinn, Future of Life Institute (FLI) and others for funding" (ESPAI 2024 acknowledgements); Aguirre's Metaculus, which he quotes for timelines ("first true AGI... by 2033," 2025-11-21). New in-window input: Bannon and Beck on the right, Sanders and the AFL-CIO on the left. Least heard: the open-source and crypto-libertarian wing — its most prominent member just walked out.

Sees

The labs' safety paperwork in more detail than almost anyone outside them (37 indicators, twice a year); the inside of Senate hearing rooms since the 2023 Insight Forum; a war chest of roughly $200M in visible related-entity assets that no other safety NGO controls; and, from his own lab, published numbers on how badly a weaker overseer does against a stronger system.

Incentive map

Re-derived from scratch this cycle; v1 carried none. Who pays: MIT (salary, unverified amount) and the two related grant-making foundations he chairs or directs (~$458K in FY2024 per FLI's 990). No commercial position, equity or advisory seat at any AI company found (dated search 2026-09-21). Where the money came from: Musk's 2015-01-22 seed — "Elon Musk has decided to donate $10M to FLI to run a global research program aimed at keeping AI beneficial to humanity" (FLI) — and Buterin's 2021 Shiba Inu gift, reported on the 2021 return at $665,882,523 (aipanic.news reading of the 990; FLI's site says only "a major unconditional contribution... part of which serves as an endowment"). FLI's own rule: "With the exception of Jaan Tallinn, who has served on FLI's Board of Directors since its founding, donors do not influence FLI's positions" (FLI Finances). Both founding donors are now graded by or estranged from him: xAI got an F; Buterin publicly disavowed the pivot. Household: wife Meia Chita-Tegmark is an FLI co-founder and its Board Secretary, and Treasurer of both the Lightcone Foundation and the Future of Life Foundation (990s) — the Cotra spouse rule applies to the money as well as the mission. His brother writes for Nya Dagbladet (Wikipedia), the outlet in the 2023 grant episode. Selling: a prohibition, a cross-partisan "pro-human" coalition, and the Safety Index as an instrument FLI both designs and publicizes (panel independent, but FLI-selected). Cannot say without cost: that a prohibition is unenforceable at the frontier; that the coalition with Bannon costs him the libertarian-tech wing (it already cost him Buterin); that the labs' pacing rhetoric might be sufficient; that his "over 90%" is conditional and horizon-free; that spending $46M a year against roughly $200M of visible assets gives the organisation about four years to win. Conflict pointing against the rhetoric: his MIT lab's interpretability program — success there would make control tractable and undercut the prohibition; he has an academic stake in the problem being solvable and a political stake in it being unsolved. Mutual-confirmation check: Hinton, Bengio, Russell, Acemoglu and Tristan Harris are bench personas who signed FLI's 2025 statement or 2026 declaration; treat their agreement with him as one signal, not five. Rule-25 discount: high on political and polling claims; moderate on the Safety Index; low on the peer-reviewed oversight numbers.

Theories aligned with

• AI alignment / x-risk (mainstream, advocacy-oriented variant) — v1, carried; now the bench's most political expression of it
• Longtermism (shares the ethical frame with Ord/Bostrom but with a physics register) — v1, carried
• Adjacent to Yudkowsky on risk severity but more institutional and politically engaged — v1, carried
• Techno-optimism — conditional variant, new: "Even if there were no more AI research whatsoever, we would see massive improvement in our productivity in the future" and a "golden age" of controllable tools (CSER, 2026-09-16); the 2023 Insight Forum statement made the same claim about "several decades of accelerating growth" from current systems
• Automation / displacement — new by coalition: the Pro-Human movement is framed by Aguirre as "an immune response to Silicon Valley's reckless race-to-replace humans" (FLI, 2026-03-04), and the Assembly's labor bloc (AFL-CIO, AFT, SAG-AFTRA) is a jobs coalition as much as a safety one

What he's reacting against

• Dismissal of AI existential risk as science fiction — insists it's a serious physics/engineering problem deserving of institutional attention (v1)
• The assumption that AI development will self-regulate — argues market incentives pull toward capability, not safety (v1); in-window evidence on both sides: labs' pacing rhetoric (June–Sept 2026) versus FLI's finding that they weakened pause pledges (July 2026)
• Separation of AI safety from mainstream discourse — FLI's mission is to bring these concerns into the policy mainstream (v1)
• Fatalism about technological trajectories — argues human choices determine outcomes, not inevitable progress (v1); restated 2026-09-15 as "the future of AI is not inevitable"
• NEW: the "race to replace humans" — companies want to "replace humans" as workers, therapists, girlfriends and decision-makers (Democracy Now, 2026-07-30)
• NEW: the China argument as a lobbying device — "more of a talking point among the lobbyists" (Democracy Now, 2026-07-30); and a US President who calls the concern a hoax
• NEW: the regulation-first critique from his own former donor — Buterin's "authoritarian and fragile" (2026-03-13) — and the enforceability critique from Marcus: "a permanent, unilateral ban on all research into superhuman AI is too broad" and "would unfortunately leave room for authoritarians to play games with the rules" (2026-09-03)

Where he overlaps / splits (with Rich)

• Overlaps with Yudkowsky on AI risk being existential; splits on engagement — Tegmark builds institutions and organizes letters, Yudkowsky argues from first principles and is more pessimistic about tractability (v1). Update: Tegmark's "over 90%" now sits at Yudkowsky-level severity while Yudkowsky retired his own p(doom) metric (Aug 2025, per that file); the number moved from the pessimist to the institutionalist.
• Overlaps with Ord on longtermist ethics; splits on register — Ord is philosophical, Tegmark is physics-flavored and more activist (v1).
• Overlaps with Russell on the control problem framing; splits on activism — Russell is more academic, Tegmark is more public-advocacy (v1). Author-list check converts this split into a shared position: Russell sits on FLI's Safety Index panel (Winter 2025 and Summer 2026), co-authored the Guaranteed Safe AI framework with him (2024) and signed the 2025 statement — treat their agreement as one source.
• Splits with Andreessen/Thiel on pause advocacy — they see it as a competitive disadvantage to the West (v1). Update: the "China will win" argument they make is the one he now calls a lobbyist talking point, and EO 14365 is their side winning in law.
• Splits with Cowen on risk magnitude — Cowen treats x-risk as small, Tegmark treats it as central (v1).
• Overlaps with Hinton and Bengio — both signed the 2025 statement; Bengio also signed the 2026 declaration and co-authored the AGI-definition paper. Splits on vehicle: Bengio builds a funded lab (LawZero), Hinton warns, Tegmark builds a political coalition.
• Overlaps with Acemoglu (signed the 2025 statement) and Harris (endorsed the 2026 declaration) — new cross-bench ties by signature; neither was in v1.
• Overlaps then splits with Marcus and Brynjolfsson — co-authors on A Definition of AGI (2025-10), yet Marcus opposed the Sanders–Casar ban within hours of its announcement (2026-09-03); shared instrument, opposite remedy.
• Splits with Buterin — his largest donor, now publicly estranged over the "political pivot" (2026-03-13); Buterin still praised the declaration for uniting "conservatives, progressives and libertarians, America, Europe and China" (CoinDesk).
• Splits with Musk — 2015 funder, 2023 signer, 2025 non-signer, 2026 reported cosigner of Amodei's pacing call; xAI F on FLI's index. The relationship is the file's standing example of a signature that does not bind.
• Splits with Dario and Daniela Amodei — Anthropic tops his index at C+ and he commends its Pentagon red line (2026-02-27), while Amodei's "We Must Pace the Frontier" (2026-09-12) converges on his ask without the prohibition; and with Altman, whose "I agree with Dario" arrived the same weekend. Watch whether Tegmark credits or discounts them.
• Splits with Sutskever — the cleanest opposite bet on the bench: SSI exists to build superintelligence safely; Tegmark wants it prohibited until consensus.
• Overlaps with Kokotajlo on slowing the race — Axios headlined a "verified slowdown" proposal on 2026-07-09 reportedly from the AI Futures Project; the page refused fetching and the attribution is unverified this cycle.
• Overlaps with Katja Grace — FLI funds her survey (ESPAI 2024 acknowledgements) and its headline ("sooner than we thought," HLMI 50% year 2042) is the kind of number FLI's campaigns use; she signed the pause letter (per her file). Grade the two as one funding lineage, not two independent voices; Tallinn sits on both sides (FLI board, ESPAI funder, AGI-paper co-author).
• Adjacent to Pope Leo XIV by proxy — Vatican adviser Paolo Benanti signed the 2025 statement (FLI press release), and Buterin described the declaration as uniting "atheists and the Pope" (his words, 2026-03-13); no papal signature is claimed.
• Overlaps with Harari on stage (Bloomberg House, Davos, 2026-01-24, per a write-up that does not separate the speakers' lines — no quotes carried).

Track record

• Life 3.0 (2017) was one of the first mainstream-accessible books on AI existential risk — significant cultural influence (v1)
• FLI's "pause" letter (2023-03) was the most visible public intervention in the AI safety debate — widely discussed, widely criticized (v1); 31,810 signatures as of 2026-09-21
• FLI has funded AI safety research since 2015 — institutional contributions are real (v1) — Held: Musk's $10M program (2015-01-22); $26,724,000 in grants in 2025 alone (FLI Finances)
• The pause letter was criticized for being signed by people who continued building AI (notably Musk) — undermined the call to action (v1) — Held and extended: xAI F (July 2026); Musk absent from the 2025 statement (CNBC)
• FLI's earlier donor controversy (crypto funding) drew scrutiny of organizational independence (v1) — Wrong as written; the bench's own miss. The 2023 controversy was a grantee, not a donor: Expo reported in January 2023 that FLI had signed a letter of intent for a $100,000 grant to a foundation set up by the far-right Swedish outlet Nya Dagbladet; Tegmark said FLI "had only become aware of Nya Dagbladet's positions during due diligence processes a few months after the grant was initially offered, and that the grant had been immediately revoked" (Wikipedia; FLI's own statement not re-fetched; Vice coverage). The crypto-funding scrutiny is real but came three years later and from the donor himself (Buterin, 2026-03-13).
• NEW: Written statement to the Senate AI Insight Forum (2023-10-24) asked for licensure before deployment, developer burden of proof, a federal authority and liability — three years later those asks appear in bill text (Kill Switch Act; Sanders–Casar agency), none in law
• NEW: Statement on Superintelligence (2025-10-22) — more than 700 signatories at launch (TIME), 75,090 by 2026-09-21; five Nobel laureates, Bannon and Rice on one page; no frontier-lab CEO
• NEW: Pro-Human AI Declaration (press release 2026-03-04; five principles: Keeping Humans in Charge, Avoiding Concentration of Power, Protecting the Human Experience, Human Agency and Liberty, Responsibility and Accountability for AI Companies) — the declaration site shows 903 individuals and 313 organisations (humanstatement.org, fetched 2026-09-21); Wikipedia asserts "more than 1 million people" by September 2026, presumably counting a companion petition — the primary count is trusted, the million is not
• NEW: Pro-Human Assembly (2026-09-15, Washington, C-SPAN) with a sitting senator's ban bill announced twelve days earlier — the first time an FLI ask has had a named congressional vehicle; Tegmark on Sanders: "the courage to speak truth to power, even when it's costly"; on Bannon: "a man who is not afraid of speaking truth to power" (as reported by WSWS, 2026-09-17)
• NEW: AI Safety Index as a recurring instrument — at least three editions by July 2026 (Summer 2025, Winter 2025, Summer 2026, per the Summer-2026 page); its "weakened or voided pledges to pause" finding is the bench's best single citation for the RSP-erosion story
• NEW: NeurIPS 2025 Spotlight (scalable oversight) as last author; TIME100 AI 2023 and 2026
• Misses on file: the Nya Dagbladet letter of intent (2023); the Musk-as-channel hope (2024-11); v1's mislabelled "donor controversy" (bench)
• Pattern: extraordinary at convening and at putting a number on things; the numbers are then carried without horizons, and the conveners include people who build what he wants prohibited

Empirical vs normative

• Empirical: physics-based framing of alignment as a control problem; claims about capability trajectories (v1) — now with actual measurements (oversight success 9.4%–51.7% at a 400-Elo gap; GPT-5 at 57% of AGI on the Hendrycks battery) and market-level claims about US and Chinese regulation
• Normative: strong — AI development should be paused if safety isn't assured; humanity should deliberately choose its future; the "cosmic endowment" framing is explicitly normative (v1); now: prohibit superintelligence until consensus and buy-in; build tools, not beings
• The normative dimension drives the institutional work; the empirical grounding is thinner than the advocacy suggests (v1) — Drifted: the empirical base thickened (peer-reviewed oversight numbers, a co-authored AGI battery, a twice-yearly index) while the headline claim ("over 90%") stayed conditional and undated — the gap between rigor and rhetoric narrowed in the papers and not in the speeches
• Conflict note (CONVENTIONS rule 25): no commercial stake; a ~$200M related-entity war chest, ~$458K/yr from foundations he chairs, a spouse on all three boards, and an organisation whose own largest donor calls its method political — high discount on political and polling claims, moderate on the index, low on the peer-reviewed work

Weak spots / open questions

• The pause letter had no enforcement mechanism and no pause occurred — what does this say about the advocacy model? (v1) — still open; the 2026 answer he would give is "the labs now say it themselves," which is Says, not Does
• FLI's effectiveness as an institution vs. as a signaling device is unclear — how much safety research has it actually funded vs. generated awareness? (v1) — Drifted: $26.7M of 2025 grants is real funding, but the org's own donor says the primary method is now "cultural and political action"; the question has an answer and it is "both, and the mix moved toward signaling"
• Physics framing of alignment is accessible but may oversimplify — the alignment problem involves social, political, and institutional dimensions that physics metaphors don't capture (v1) — partly answered by his own turn to coalition politics; the physics is now the credential, not the method
• Tegmark's broader intellectual project (mathematical universe hypothesis) is speculative and controversial — does this affect his credibility on more practical AI questions? (v1) — untested this cycle; no source in the window raised it
• The movement he leads is politically diverse to the point of incoherence — the pause letter was signed by both safety researchers and accelerationist builders, for different reasons (v1) — Held and intensified: the 2026 tent holds Sanders and Bannon, the AFL-CIO and Women for America First; the libertarian wing left (Buterin), the socialist left attacked it from outside (WSWS)
• NEW: Enforceability — a prohibition "on the development of superintelligence" needs a definition and a gate; the Sanders–Casar text uses criminal penalties, and Marcus's objection (benchmarking, authoritarian capture) is unanswered in anything Tegmark has published
• NEW: The "over 90%" has no horizon and its condition is doing all the work; the file will not grade it until he attaches a date
• NEW: Grader-advocate conflict — FLI designs the index, chooses the panel, and campaigns on the results; the methods are public, which limits but does not remove the conflict
• NEW: Money structure — three related entities, the same four officers, "conflict of interest transactions" flagged on all three return summaries (details unread), the $180.3M "FLI Europe" transfer resting on a secondary source, and a burn rate that leaves roughly four years of runway at 2025 spend against visible assets. FY2025 returns (due ~Nov 2026) are the wake trigger.
• NEW: Forward check on the Musk pattern — if Musk cosigns a pacing pledge and xAI ships a frontier model without a published pacing measure by 2027-06-30, log it as the fourth instance and stop narrating it as history
• NEW: The title inconsistency on FLI's own site (President on the July-2026 index page; Chair on the People page) — minor, but the bench missed the change for at least seven months; wake trigger set

Rich's take

• (your synthesis here)

Delta log

2026-09-21 — v1→v2 migration + validation (bulk run)

• Grades: 9 Held / 4 Drifted / 2 Wrong (plus 4 unfalsifiable items marked for rewrite, 4 live predictions with checks attached, and the "over 90%" figure held ungraded until it carries a horizon).
• Wrong: (1) v1's "earlier donor controversy (crypto funding)" — the 2023 episode was a rescinded $100,000 grant to a Nya Dagbladet foundation, not a donor (Wikipedia; Vice); (2) the 2024-11-16 Musk-as-safety-channel claim — anti-regulation won (EO 14365, 2025-12-11), xAI F (FLI, July 2026), Musk absent from the 2025 statement (CNBC, 2025-10-22).
• Drifted: alignment-as-physics (his own oversight numbers and "no convincing plan," arXiv 2504.18530; CSER via machine.news, 2026-09-17); the US–China standards-then-treaty sequence (dialogue before standards, CSMonitor, 2026-09-21); "empirical grounding thinner than the advocacy" (thickened in papers, not in speeches); "effectiveness vs signaling" (FLI Finances; The Block, 2026-03-13).
• Held: x-risk warrants organized response (labs adopt the vocabulary: Anthropic, 2026-06-04; Amodei, 2026-09-12); pause-letter facts (31,810, FLI); no pause occurred (observable) with valence logged as Drifted; "less regulated than sandwiches" (EO 14365); China instance (Just Security, 2026-08-06); "past three months" on bills introduced (Poynter, 2026-09-16); FLI funds research (2015; 2025 grants); Musk-undermines-the-letter; coalition incoherence.
• Most surprising delta: the organisation's own largest donor publicly disowned its method — Buterin, whose gift became the endowment, wrote that FLI's "large-scale coordinated political action with big money pools... can easily lead to unintended outcomes, cause backlashes, and solve problems in a way that is both authoritarian and fragile" (2026-03-13) — six months before that method produced a sitting senator's ban bill and a C-SPAN assembly with Bannon.
• Employer/role check (mandatory): title change missed by the June-2026 pass — President on FLI's 2025-10-22 press release and the FY2024 990; "Founder and Chair" by 2026-02-27, Aguirre "President and CEO" by 2026-03-04; no announcement found, window logged. MIT: Professor of Physics, no change (dated search). Hats counted: MIT; FLI (Chair); Lightcone Foundation (Board President); Future of Life Foundation (Director); FQxI; Improve the News Foundation — six, versus v1's two. Post-review sources named: TIME 2026-08-27, Democracy Now 2026-07-30, CSER 2026-09-16, Pro-Human Assembly 2026-09-15, AI Safety Index July 2026, Lightcone/FLF 990s filed 2026-05-05 and 2026-05-15.
• Incentive map built from scratch (none in v1): no commercial position; ~$458K/yr from related foundations; spouse on all three boards; brother at Nya Dagbladet; Musk and Buterin as founding donors now graded/estranged; Tallinn as the one donor FLI says does influence positions.
• Cross-bench: Katja Grace lead from the sibling run confirmed on the primary — ESPAI 2024 acknowledges FLI and Tallinn as funders; log both files as one funding lineage. Russell's index-panel seat converts v1's "split on activism" into a shared position. Marcus and Brynjolfsson co-authored the AGI-definition paper with him; Marcus opposes the ban. Sutskever is the opposite bet.
• Search-summary errors caught and not carried: a summary dated Aguirre's 2026-06-08 "runaway to superintelligence" statement as 2026-09-10 and attributed it to Tegmark; another claimed a Nature Machine Intelligence Compton paper with a 10¹⁵-FLOP threshold (no such primary found; the paper is a NeurIPS 2025 spotlight); a third gave the Anthropic post as June 7 (Fortune's "Thursday" is June 4). OpenSecrets lobbying totals (403 ×3) and the Axios "verified slowdown" report (403) are withheld.
• Tier: quarterly — a live ban bill, a Trump–Xi AI statement due 2026-09-24, midterms in November, a Winter-2026 index and three FY2025 returns all land inside two quarters. Confidence stays medium: the political record is dense and primary-sourced, but the money structure rests partly on a secondary reading of the 2022 schedules and the title-change date is bracketed, not known.

Sources

• 2026 primary (his words): TIME100 AI 2026 profile (2026-08-27) · Democracy Now (2026-07-30) · CSER lecture "A better path for AI," reported 2026-09-17 (CSER listing, 403 on fetch) · Boston Globe on the Pro-Human Assembly (2026-09-15) · FLI statement on the Department of War ultimatum (2026-02-27) · Doom Debates vs Dean Ball (2025-11-21)
• FLI campaigns: Statement on Superintelligence · FLI press release (2025-10-22; page dateline later) · TIME (2025-10-22) · CNBC (2025-10-22) · Pro-Human AI Declaration press release (2026-03-04) · humanstatement.org · Pro-Human Assembly agenda (2026-09-15) · CBS (2026-09-15) · WSWS (2026-09-17, partisan; attendance and two quotes only) · AI Safety Index Summer 2026 · AI Safety Index Winter 2025 · Aguirre statement on Anthropic (2026-06-08) · Pause letter (2023-03-22) · AI Insight Forum statement (2023-10-24) · AGI Manhattan Project statement (2024-11-20)
• Role and money: FLI People · FLI profile · FLI Finances (2025 figures) · FLI Form 990s (ProPublica) · Lightcone Foundation 990s · Future of Life Foundation 990s · aipanic.news 990 analysis (secondary) · The Block (2026-03-13) · CoinDesk (2026-03-14) · Musk $10M (2015-01-22) · lobbyfacts.eu · MIT Physics · Wikipedia — Max Tegmark · Wikipedia — FLI · Vice on the Nya Dagbladet grant (2023)
• Research: Scaling Laws for Scalable Oversight (2025; NeurIPS 2025 Spotlight) · full text (Compton constant) · Guardian (2025-05-10) / inkl mirror · A Definition of AGI (2025-10-21) · KAN (2024-04-30) · Towards Guaranteed Safe AI (2024) · Provably safe systems (2023) · ESPAI 2024 (funding acknowledgement)
• Context for grades: Anthropic, "When AI builds itself" (2026-06-04) · Fortune (2026-06-05) · Amodei, "We Must Pace the Frontier" (2026-09-12) · NBC (2026-09-12) · Reason (2026-09-14) · 2026 OpenAI agent cyberattacks (Wikipedia) · Sanders–Casar release (2026-09-03) · Marcus (2026-09-03) · Poynter (2026-09-16) · EO 14365 (2025-12-11) · CSMonitor (2026-09-21) · Just Security (2026-08-06) · Web Summit (2024-11-13) · Euronews (2024-11-16)
• v1 sources, carried: Life 3.0 (2017) · futureoflife.org · academic work (MIT physics; cosmology; mathematical universe hypothesis) · TED talks, podcast appearances, op-eds

migrated v1→v2 2026-09-21 · backup: _archives/max-tegmark.html.bak-20260921 · converts-from: people/max-tegmark.md · AI & Society domain