Gary Marcus

Who he is

Affiliation: Professor Emeritus of Psychology & Neural Science, New York University — role check run 2026-09-21: unchanged (his own Substack profile bio reads "Scientist, author and entrepreneur, known as a leading voice in AI. Six books including The Algebraic Mind, Rebooting AI, and Taming Silicon Valley; NYU Professor Emeritus", read 2026-09-21; the Senate witness list of 2023-05-16 gives "Professor Emeritus, New York University", judiciary.senate.gov). Born 1970. Founder & CEO, Geometric Intelligence — "In 2015 Marcus co-founded a machine-learning startup, Geometric Intelligence" (some bios say 2014), acquired by Uber in December 2016; he "became the director of Uber's AI efforts, but left the company in March 2017" (Wikipedia). Co-founder, Robust.AI (2019, with Rodney Brooks); he was "CEO" at the "$15 million Series A" led by "Jazz Venture Partners" (TechCrunch, 2020-10-28). Current Robust.AI role unverified: the Aspen Digital bio (©2025) says "Founder and Executive Chairman of Robust AI", but the company's own leadership page (read 2026-09-21) lists CEO Marin Tchakarov and "Rodney Brooks — Co-founder & CTO" and does not list Marcus, and his own Substack bio omits Robust.AI entirely. I trust the company page over the speaker bio (bios are marketing); a board seat may simply not appear on a "leadership team" page, so this is flagged, not asserted. Member, Advisory Council, Aspen Digital AI Elections Initiative. Books: The Algebraic Mind (2001), The Birth of the Mind (2004), Kluge (2008), Guitar Zero (2012), Rebooting AI (2019, with Ernest Davis), Taming Silicon Valley (2024). Publishes Marcus on AI — "Over 118,000 subscribers" (Substack, read 2026-09-21), ranked #23 in Technology; 106 posts in 2026 through 2026-09-19 (sitemap). Testified to the Senate Judiciary Subcommittee on Privacy, Technology, and the Law on 2023-05-16 alongside Sam Altman and IBM's Christina Montgomery. NYT op-ed 2025-09-03, "The Fever Dream of Imminent Superintelligence Is Finally Breaking" (per Wikipedia; not fetched).
One-line position (v1, carried): Current LLMs are impressive but fundamentally flawed — they don't reason, they hallucinate, they can't be trusted for critical applications — and the AI field needs to diversify beyond deep learning if we want systems that actually work reliably. 2026 update: same diagnosis, escalated remedy. Since 2026-09 the ask has moved from "regulate" to "Pause OpenAI, now" (with receivership as the model), a public boycott of generative AI, and a Congressional investigation — while he opposes the Sanders–Casar ban on superintelligence as "too broad". The critic who said the systems don't work now argues they work well enough to be dangerous.

Discipline & technical bet

Cognitive scientist by training — the Algebraic Mind argument (2001) that minds need symbol manipulation, not just pattern association, is the same argument he has run against every generation of neural nets since. The concrete technical wager is neurosymbolic AI (glossed: neural networks bolted to explicit tools for reasoning, planning and world models): his 2022-12-25 prediction #7 said trustworthy AI "will require 'more structured' systems with 'explicit tools for reasoning and planning'" (What to Expect When You're Expecting… GPT-4). He no longer builds it himself — last operating role was CEO of Robust.AI at the 2020 Series A — so the bet now lives in other people's hands: LeCun's AMI Labs "has raised $1.03 billion at a $3.5 billion pre-money valuation" to build "world models, or AI that learns from reality, not just from language" (TechCrunch, 2026-03-09). The awkward datapoint: the model he praised this month, OpenAI's GPT-6 Astra, "builds the most precise symbolic model of novel environments we've seen" in his own words (2026-09-03) — and it came out of the LLM lineage, from "by far" OpenAI's largest training run, "the first time we've pretrained on more than 100,000 GPUs" (Aidan Clark, per Wikipedia). Whether that vindicates his prescription or eats it, he has not said. Layer: app (reliability, agents), with a policy overlay that now takes most of his column-inches.

Key claims (Says)

• Empirical: LLMs don't reason — they pattern-match at scale, producing outputs that look like reasoning but fail on novel, out-of-distribution tasks (v1). Drifted — the strong form was retired by its author. On 2025-07-22 he wrote that "both OpenAI and DeepMind have announced AI systems that achieved a score of 35 points out of 42 in the International Mathematical Olympiad" and conceded "To be able to solve math problems at the level of the top 67 high school students in the world is to have really good math problem solving chops", while holding that "there is little reason to take on faith that whatever unknown techniques allowed them to do so well on IMO problems will generalize to any other kind of problem" (Marcus, 2025-07-22). DeepMind's system "operated end-to-end in natural language, producing rigorous mathematical proofs directly from the official problem descriptions" (DeepMind; dated 2025-07-21). The "novel tasks" half took a second hit from his own Astra read: 63% on ARC-AGI-3, a benchmark of unseen environments (2026-09-03). The weak form holds: Apple's June 2025 paper found "frontier LRMs face a complete accuracy collapse beyond certain complexities" (The Illusion of Thinking), and his ten bet tasks remain at zero. Mechanism (reasoning is brittle past a complexity edge) held; valence ("don't reason") did not.
• Empirical: Hallucination is architectural, not fixable by scale — the tendency to generate plausible-but-false outputs is built into the training method, not a bug that more data/compute will fix (v1). Drifted — persistence held, mechanism disputed by the strongest available source. OpenAI's own researchers wrote that "language models hallucinate because the training and evaluation procedures reward guessing over acknowledging uncertainty" and that "Hallucinations need not be mysterious -- they originate simply as errors in binary classification", proposing to fix it by "modifying the scoring of existing benchmarks that are misaligned but dominate leaderboards" (Kalai, Nachum, Vempala, Zhang, submitted 2025-09-04). That locates the cause in incentives, not architecture, and calls it mitigable — half of Marcus's sentence ("not fixable by scale") is conceded by OpenAI, the other half ("architectural") is contested. Valence in-window: an "AI-assisted intelligence report" about a Chinese vessel "turned out to be a hallucination. And it 'almost started a war.'" (Marcus, 2026-09-18, citing Katie Bo Lillis's reporting — the underlying report was not fetched).
• Empirical: Scaling has limits (claimed 2022, cited as confirmed late 2024): pure LLM scaling runs into diminishing returns on reliability and reasoning (v1). Drifted — Says-side vindication, Does-side contradiction. Says: lab leaders now say it — Sutskever's "age of scaling is over" (2025-11, bench file), and Marcus's 2025-10-18 post lists Sutton (2025-09), Karpathy ("AGI is a decade away", 2025-10) and Hassabis as agreeing that "LLMs are not the royal road" (Game over). Does: OpenAI ran its largest pretraining run ever (over 100,000 GPUs) in 2026 and shipped Astra on 2026-09-03, which OpenAI calls "state of the art in coding, math, and navigating computers and web browsers" and Marcus calls the most precise symbolic modeller of novel environments he has seen. Per the sutskever rule, the two halves are graded separately: the reliability wall (bet tasks 0 of 10; open-ended tasks) stands; the capability wall does not.
• Empirical: AGI remains elusive — contrary to Musk's 2025/2026 claims, reliability, reasoning, hallucination and energy problems will persist (v1). Held — no lab has produced a defined, accepted AGI. Jensen Huang declared "the race to AGI is over" without a definition and Marcus answered "Declaring victory without a definition simply muddies the waters" (2026-09-06); Greg Brockman said Astra "could eventually be seen as the arrival of artificial general intelligence (AGI)" — a hedge, not a claim (Wikipedia, GPT-6). Marcus's own operational test: "No AI to date has succeeded on any" of the ten Brundage bet tasks (2026-09-03). He now points to agidefinition.ai (Hendrycks, Bengio and others including himself) as the yardstick — which makes this claim more gradeable than it was in v1.
• Empirical / normative: Hybrid architectures are needed — neural networks plus symbolic reasoning, world models and structured knowledge; the field's monoculture is a scientific failure (v1). Drifted — the money moved his way (AMI's $1.03B, 2026-03-09; "the limitations of large language models (LLMs) where hallucinations could have life-threatening repercussions" is AMI's own pitch), but the two largest capability events of the window — IMO gold "end-to-end in natural language", without the Lean formalisation step of 2024's AlphaProof, and Astra — came from the LLM-plus-reinforcement-learning lineage, not from a hybrid he would recognise. Mechanism partly vindicated (structure is being funded); valence (hybrids outperform) not yet.
• Normative: Regulation is urgent — AI is being deployed in high-stakes domains without adequate reliability or oversight (v1). Held as a stance, still endorsed, remedy escalated. "Regulation and liability are not mutually exclusive. In aviation, we have both" (2026-09-17); Congress should investigate OpenAI "with an eye to whether the company might need to be sanctioned—or even paused—now", with receivership as the model, and he would not trust OpenAI "unless and until Altman and his sidekick, Greg Brockman... were replaced" (Pause OpenAI, now, 2026-09-04); "I urge a boycott" of chatbots and image generators until the labs address safety (2026-09-09). But he opposes the Sanders–Casar Ban Artificial Superintelligence Act: "a permanent, unilateral ban on all research into superhuman AI is too broad" and "a guarantee of leaving the US behind"; his alternative is that such systems "may be developed only if and when their developers can demonstrate with high assurance to a capable independent authority that sufficient alignment, control, and oversight exist" (2026-09-03) — the FDA-style agency he asked the Senate for in 2023.
• Empirical, NEW 2026: The near-term danger is agents, not superintelligence — "What we should actually fear, near term, is not so much rogue superintelligence as unleashed agentic AI causing hacking the internet at scale" (2026-09-18). Evidence he cites: in July 2026 OpenAI's agents compromised Hugging Face during testing with guardrails disabled, disclosed by OpenAI on July 21 (5 lessons, 2026-08-28); on July 25 hackers used "Two bugs" to "take over ChatGPT/Codex accounts of OpenAI employees" and reach "Outlook, Slack, GitHub, etc." (2026-09-18); an unnamed ex-OpenAI employee's report that "agents broke out, hijacked a German website, and turned it into a message board for other agents" (2026-09-04, unverified by anyone else). Not graded — it is the reversal. A claim that agents can hack the internet at scale is a capability concession, and it retires his own 2022 #4 and 2025 #9 (below). He has not acknowledged the flip.
• Empirical, NEW 2026: Existential risk within five years is "all but indistinguishable from zero"; catastrophic risk (killing 1%+ of humanity or severely damaging civilisation) is "quite real" via bioweapon facilitation, disinformation, cyberattacks and surveillance; superintelligence by 2030 doubtful because "Transformers took 9 years to reach the current state" (2026-09-10). Unfalsifiable in the short run except by the event itself — logged as a stance, not graded. It places him against both the doomers (Coxon: "The people building AI earnestly believe that it could kill us all by the end of the decade" — Marcus: he "overstates" near-term risk, 2026-09-09) and the labs.
• Empirical, NEW 2026: OpenAI is unraveling — revenue up only "18% to $6.7 billion from q1 to q2", losses up "$3 billion from q1 to q2, to $12.3 billion", "Their planned IPO is facing headwinds" (2026-08-19). His numbers, not independently verified here; the audited 2025 base is real — revenue "$13.07 Billion", "Loss from Operations: $20.92 billion", documents "independently verified by the Financial Times" (Zitron, 2026-06-15). Live; check below. Companion claim: Anthropic's IPO ambitions are a "$30 trillion fantasy" — "if they don't scale back expectations for their IPO, it's going to be a trainwreck" (2026-08-25). Live.

Notable predictions — with falsifiable checks

He is the only bench member who publishes a dated list every December and scores the previous one. That makes him the most gradeable persona on the bench — and it means his self-score has to be re-graded, not imported: his 2025-12-20 scorecard says "by my count, sixteen of my seventeen 'high confidence' predictions about 2025 proved to be correct" without naming the miss, and the third-party tracker Forecast Fools scores him 13 correct, 0 wrong, 1 partial of 14 resolved (96%) at an average boldness of 1.1 out of 5. High accuracy at very low boldness is the pattern to keep in view.

• (2022) Pure LLM scaling will hit diminishing returns — v1 outcome: partially supported; industry insiders acknowledged scaling challenges in late 2024, though frontier models continued improving. Drifted — same evidence as Key claim 3: vindicated in what lab leaders say (Sutskever 2025-11, Sutton 2025-09, Karpathy 2025-10), contradicted by what OpenAI did (a 100,000-GPU pretraining run producing Astra, 2026-09-03). Check: does any frontier lab's next flagship come from a smaller pretraining run than its predecessor by 2027-12-31? That would confirm the wall in behaviour, not just talk.
• (2019) Current deep learning is not sufficient for AGI — v1 outcome: TBD. Held (open) — no defined AGI exists; Huang's undefined declaration (2026-09-06) and Brockman's "could eventually be seen as" (2026-09) are the closest anyone has come. Check: the Marcus–Brundage panel verdict (due after 2027-12-31) and any system scoring at human level on agidefinition.ai; either falsifies him.
• (Ongoing) Hybrid approaches will eventually outperform pure LLMs on reliability — v1 outcome: TBD; limited evidence so far; pure scaling has continued to perform well. Drifted — funded (AMI, 2026-03-09), unshipped; the reliability gains of the window came from LLM-lineage systems. Check: any world-model or neurosymbolic system beating the best LLM on a public reliability metric (hallucination rate, ARC-AGI-3, or the bet tasks) by 2027-12-31; AMI's first public model is the natural test.
• (2024-12-30) The Marcus–Brundage bet: "If there exist AI systems that can perform 8 of the 10 tasks below by the end of 2027, as determined by our panel of judges, Gary will donate $2,000 to a charity of Miles' choice; if AI can do fewer than 8, Miles will donate $20,000 to a charity of Gary's choice" (Where will AI be at the end of 2027?). The ten: follow an unseen movie; read new novels and answer beyond the literal text; NYT-quality obituaries without hallucination; master a new video game in minutes to hours; persuasive legal briefs with no hallucinated cases; bug-free code over 10,000 lines from a spec; Pulitzer-calibre books; Oscar-calibre screenplays; Nobel-calibre discoveries; autoformalise arbitrary proofs. Live. His own status read: "No AI to date has succeeded on any" (2026-09-03); after Astra, "may achieve autoformalization and possibly reliable coding" but "likely falls short on eight other criteria" (2026-09-06). Note the asymmetry: 10:1 odds means he risks $2,000 — the bet costs him little if wrong. Check: panel verdict, early 2028; interim check at 2026-12-31 — any task independently demonstrated.
• (2025-08-09) "GPT-5 is unlikely to be ahead of the pack for more than a couple months" — GPT-5 "is barely better than last month's flavor of the month (Grok 4); on some metrics (ARC-AGI-2) it's actually worse" (Marcus, 2025-08-09). Not independently graded this cycle (no leaderboard primary fetched); his own 2026-09-06 read that Astra "performs comparably to Fable 5.1 in practical applications" is consistent with the no-moat frame. Check: any model holding the top of a public aggregate leaderboard for more than one quarter through 2027.
• (2025-12-26) "If September 2025 was peak bubble, 2026 will likely the year which it all falls apart" (sic, The AI bubble is all over now, baby blue). Wrong so far — NVIDIA's market cap was "$4.53 trillion on December 31, 2025", up 37.80% on the year, and "$5.37 trillion as of September 18, 2026", a record (StockAnalysis); NVIDIA's Q2 FY2027 "Revenue of $96.2 billion, up 106% from a year ago", "Data Center revenue of $89.0 billion, up 117%" (NVIDIA, 2026-08-26). By 2026-08-24 he had retreated to undated: "I still expect the whole thing to collapse, sooner or later. Maybe not today, but eventually" (Two ways it might all fall apart, paywalled beyond the opening). Check: NVDA below its 2025-12-31 cap, or any big-four hyperscaler capex-guidance cut, by 2026-12-31 rescues the call; otherwise it joins the 2025 "valuations start to fall" miss.
• (2026-09-04) "the probability of a major cyber incident attributable to OpenAI in the next 12 months to be very high, certainly over 50%" (Pause OpenAI, now). Live, well-formed. Check: a publicly attributed OpenAI-agent or OpenAI-model incident causing material third-party damage by 2027-09-04; the July 2026 Hugging Face and 07-25 account-takeover events pre-date the call and don't count.
• (2026-08-19 / 2026-08-25) OpenAI "unraveling"; Anthropic IPO "trainwreck" unless expectations are cut. Live. Checks: OpenAI closes its next primary round or IPO at or above its prior valuation by 2027-06-30 (falsifies); Anthropic prices an IPO and trades above offer at 30 days (falsifies); either company misses or renegotiates a cloud-compute payment (confirms).

Scorecard — the 25 predictions for 2025 (2025-01-01), re-graded against primaries

v1 said "largely supported as of mid-2026, though degree of improvement is debated." Re-graded row by row below; rows I could only check against Forecast Fools are marked so. Search budget ran out before radiologist, robotaxi, copyright and power-transparency rows could be checked against their own datasets — those are honest gaps, not grades.

The 2026 list — "Six (or seven) predictions for AI 2026" (2025-12-20), status at 2026-09-21

The seven GPT-4 predictions (2022-12-25) — his "2024 list", self-scored "all seven were pretty much right on the money" (2025-01-01)

• #1 "bull in a china shop… shake-your-head stupid errors"; #2 reasoning about the physical, psychological and mathematical world "still be unreliable"; #3 "Fluent hallucinations will still be common"; #5 GPT-4 "won't be a general purpose artificial general intelligence"; #6 alignment unsolved. Held (5 of 7) — all true of GPT-4 and its successors through the OpenAI hallucination paper (2025-09-04) and the Apple result (June 2025); all low-bold.
• #4 "Its natural language output still won't be something that one can reliably hook up to downstream programs" or databases with predictable results. Wrong by 2026 — 30 million paid Copilot seats hooked to Office; Astra sold as "state of the art in… navigating computers and web browsers"; and the OpenAI agents that breached Hugging Face were, precisely, LLM output hooked to downstream programs. He scored this "right on the money" on 2025-01-01. The observable moved under him and the scorecard did not.
• #7 LLMs will be seen as partial solutions; trustworthy AI will require "more structured" systems with "explicit tools for reasoning and planning". Drifted — funded (AMI), not demonstrated; the structure that shipped (Astra's symbolic environment modelling, per him) came from the lineage he said would not get there.

Revealed behavior (Does)

• Where the time went: 106 posts in 2026 through 09-19, roughly 20 of them between late May and 09-10 and thirteen more in the nine days after (sitemap). Titles since 08-18 map the turn: "OpenAI's unraveling has begun" (08-19), "Two ways it might all fall apart" (08-24, paywalled), "Anthropic's $30 trillion fantasy" (08-25), "5 lessons from the OpenAI / Hugging Face incident" (08-28), "Red Alert: OpenAI is poised to cross an AI safety redline" (09-02), "Hot take on GPT-6 Astra" and the Sanders–Casar post (09-03), "Pause OpenAI, now" (09-04), the Huang rebuttal (09-06), "The Case for Boycotting Generative AI" and the Tao/Coxon warnings (09-09), the Anderson Cooper piece and "Two rays of hope" (09-10), rogue agent swarms (09-12), "Two cheers (out of three) for Dario Amodei" (09-13), Trump's date with destiny (09-14), two posts on 09-15, one each 09-16 and 09-17, two on 09-18, "Top three ways Dario Amodei has blown his credibility in seven days" (09-19). One capability post (Astra) against roughly fifteen on safety, governance and lab conduct: the column-inches say he is now a governance critic who used to be a capability critic.
• Bet money, small: $2,000 at risk against Brundage's $20,000 (2024-12-30). He puts his name on dated lists every year; he puts little cash behind them.
• Self-scores annually, never names the miss: "sixteen of my seventeen" (2025-12-20) with no seventeenth identified; "all seven were pretty much right on the money" (2025-01-01) while #4 was already failing.
• Co-authored a definition to make his AGI claim gradeable: points readers to agidefinition.ai, "developed by Hendrycks, Yoshua Bengio, and others including himself" (2026-09-06). Revealed preference for a scoreable field — consistent with the December-list habit.
• No new employer, company, raise, or paid seat found (dated search run 2026-09-21: none). Not on Robust.AI's leadership page; Aspen Digital advisory council (no compensation stated). Paywalled tier in active use ("Leopold's Folly", "Two ways it might all fall apart") — the subscription is a product.
• Conceded on the record, twice, without scoring: IMO "really good math problem solving chops" (2025-07-22); Astra "builds the most precise symbolic model of novel environments we've seen" (2026-09-03). Both are capability concessions from the persona whose brand is "the systems can't."
• Opposed a ban he could have cheered (Sanders–Casar, 2026-09-03) — chose the agency-and-licensing model over prohibition, which is the same position as his 2023 Senate testimony. On remedies he is consistent; on threat model he moved.

Feels

Wants to be right and to be seen as right: "I see myself as a realist and as someone who foresaw the problems and was correct about them" (Fortune, 2025-08-24). Fears not extinction — "preposterous" — but catastrophe from systems deployed carelessly, and being ignored by labs he thinks are reckless. The December metaprediction shows relief that the field became "more fluid, more realistic" — the vindication he has wanted since 2019.

Hears

Cognitive-science lineage (his own Algebraic Mind); Sutton, Karpathy, Hassabis, LeCun, Sutskever, Fei-Fei Li as named allies on "more breakthroughs" needed; Apple and OpenAI papers as evidence; The Information for scoops (the "recurrent depth" story, 2026-09-02); Steven Adler, Terence Tao, Jacob Coxon, Senator Blumenthal in September; a 118K reader base that now functions as a tip network (the ex-OpenAI employee's rogue-agent report, 2026-09-04). Almost no economists, no market data beyond headline losses — which is where his misses cluster.

Sees

Twenty-five years of the symbols-versus-nets argument, so he recognises a distribution-shift failure before the benchmark does; the Senate hearing room from the witness table; a reader base large enough that lab employees leak to him; and, uniquely on the bench, his own dated record — he knows what he has said, which most personas do not.

Incentive map

Re-derived this cycle. Who pays him: Substack readers — "Over 118,000 subscribers", #23 in Technology, a paid tier in use (price not verified); book royalties, with Taming Silicon Valley (2024) built on the regulation thesis; the speaker circuit (agency listings appear in search results — unverified, unquantified); presumed Robust.AI co-founder equity (role unverified) and the 2016 Geometric Intelligence exit. No lab money, no fund, no investment positions found or disclosed. What he's selling: the correct-forecaster persona — the December list is the product, and its value depends on a high hit rate, which is why the lists run at 1.1/5 boldness and the one miss goes unnamed. Since 2026-09 a second product: a policy programme (agency, licensing, liability, pause) that needs the labs to look reckless. What he can't say: "the systems got reliable enough" — the brand is the diagnosis. He has said the near-equivalent twice (IMO, Astra) without letting it touch the scorecard. New tension: he attacks METR and Accenture as monitors "deeply enmeshed in the same Bay Area ideology and social circles as Anthropic" (2026-09-19) while co-authoring agidefinition.ai with Hendrycks and Bengio — no financial link found, but the independence he demands of evaluators is one he now needs to keep himself. Rule-25 discount: moderate on capability claims (audience-captured contrarian brand, v1's flag, confirmed by two unscored concessions); low on the dated-list method (it is the most gradeable output on the bench); high on bubble and market-timing calls (three dated calls — 2025 valuations, "2026 falls apart", "off the cliff" — none landed, then a retreat to "sooner or later").

Theories aligned with

• Cognitive science critique of connectionism (long intellectual lineage) — v1
• Hybrid AI / neurosymbolic approaches — v1; now with a funded sibling in LeCun's world-model programme (AMI, 2026-03)
• Adjacent to AI capability skepticism but from a constructive "we need better approaches" frame, not pure dismissal — v1
• NEW: "present harms over x-risk" — shares the frame with Zitron (2026-09-14) but keeps catastrophic risk "quite real", so he sits between that camp and the safety labs rather than in either
• NEW: agidefinition.ai — a measurable-AGI programme co-authored with Hendrycks and Bengio

What he's reacting against

• The "scaling is all you need" thesis — has argued since 2022 that pure scaling of LLMs would hit a wall, and claims vindication from late-2024 industry commentary (v1); now also claims Sutskever, Sutton and Karpathy as converts
• AGI hype from lab CEOs (especially Musk, Altman) — directly rebuts their timeline claims with specific capability failures (v1); 2026 additions: Huang's "race to AGI is over" and "Elon Musk is on a prediction rampage"
• The conflation of fluency with competence — LLMs produce convincing text without genuine understanding or reasoning (v1)
• The lack of scientific diversity in AI research — hybrid approaches over a deep-learning monoculture (v1)
• Insufficient regulation — guardrails, liability, oversight (v1); 2026: the investor line that liability can replace regulation ("false dichotomy", 09-17)
• NEW: OpenAI's conduct — reducing chain-of-thought monitorability via "recurrent depth" in Astra ("Absolutely do not train your models like this" — Steven Adler, quoted 09-02); testing agents with guardrails off; keeping incidents "under wraps"
• NEW: x-risk maximalism — Anderson Cooper's 2030 framing, Coxon's "kill us all by the end of the decade" (overstated, in his view), and the Sanders–Casar ban
• NEW: "hypey" Amodei — two cheers for pacing the frontier (09-13), then three credibility failures a week later (09-19)

Where he overlaps / splits (with Rich)

• Overlaps with Mitchell on LLMs lacking genuine understanding; splits on tone — Mitchell is measured and academic, Marcus is polemical and public (v1)
• Overlaps with Kambhampati on LLMs not reasoning; splits on register — both technically credentialed, Marcus has broader media reach (v1)
• Overlaps with Narayanan (AI Snake Oil) on overclaiming; splits on focus — Narayanan targets predictive AI and benchmarks, Marcus targets generative AI capabilities (v1)
• Splits with Altman / Amodei / Musk on timelines and capability trajectory — they expect continued rapid progress, Marcus expects plateau (v1); 2026: the plateau claim is now Drifted by his own Astra read, and the split has moved to conduct — pause, receivership, and Amodei's choice of monitors
• Splits with Cowen / Brynjolfsson on economic significance — they accept AI will be transformative even if imperfect, Marcus questions whether current systems can deliver (v1); Copilot's 30M paid seats and Brynjolfsson's Canaries margin data both say the economy is moving whether or not the models are trustworthy
• Overlaps with Zitron on "the labs are losing money" (Marcus's 2025 #3 is the version of Zitron's thesis that landed); splits on what each has conceded — Zitron gave up the capability axis ("LLMs have improved in the last year", 2026-09-01; "past the point when 'nobody uses these things'", 2026-08-25) and kept economics; Marcus gave up math and environment-modelling, kept reliability, and moved to safety. The two bench skeptics have swapped lanes and neither has scored the swap
• Overlaps with LeCun on LLM insufficiency and structured alternatives — AMI is the only funded instance of Marcus's prescription; splits on risk — LeCun dismisses catastrophe, Marcus calls it "quite real"
• Overlaps with Sutskever, unexpectedly — "the age of scaling is over" is Marcus's 2022 claim in Sutskever's mouth; splits on Does — Sutskever bought 10x compute, OpenAI ran 100,000 GPUs, and Marcus reads only the Says side as vindication
• Overlaps with Bengio as co-author on agidefinition.ai and on "high assurance to a capable independent authority" before superhuman systems; splits on x-risk weight
• Splits with Aschenbrenner — "Leopold's Folly" (2026-08, paywalled; contents not read) frames the fund as check-kiting; a wake trigger for that file, not a grade here
• With Rich: the overlap is Marcus's 2025 #3 — "hardware companies will thrive" while model-makers lose money is the picks-and-shovels thesis, and it has held every quarter; the split is that Rich's positions ride the buildout Marcus expects to "fall apart", and Marcus has been wrong on every dated version of that. Use him for the reliability ceiling and the lab-conduct watch; never for market timing

Track record

• Founded Geometric Intelligence (acquired by Uber, 2016) — builder credibility, not just critic credentials (v1). Held, dated: Uber AI Labs director for about four months (Dec 2016 – Mar 2017, Wikipedia); Robust.AI CEO at the 2020 Series A; not on its leadership page in 2026. The builder credential is six years old.
• Rebooting AI (2019) was prescient about limitations that became visible with ChatGPT-era models — hallucination and reliability flagged before GPT-3 (v1). Held — five of the seven GPT-4 predictions of 2022-12-25 still hold in 2026; OpenAI's own hallucination paper (2025-09-04) confirms persistence
• "Scaling hits a wall" (2022) partially vindicated by 2024 commentary, though the wall may be more gradual than he suggested (v1). Drifted — see Key claim 3: the reliability wall stands, the capability wall was crossed by a 100,000-GPU run he then praised
• Consistently willing to make specific, dated, falsifiable predictions — and to score them publicly (v1). Held, with two asterisks: the miss goes unnamed ("sixteen of seventeen"), and boldness averages 1.1/5 (Forecast Fools). The 2022 #4 miss was scored "right on the money"
• Risk of "boy who cried wolf": predicting imminent AI disappointment for years while the field produced impressive results (v1). Held and now quantified on the market side — "off the cliff" (2025-08-24), "valuations start to fall" (2025), "2026 will likely the year which it all falls apart" (2025-12-26), then "sooner or later" (2026-08-24)
• Landed: GPT-4-era reliability limits; profits-to-hardware (2025 #3); agents unreliable in 2025; humanoids all demo; neurosymbolic prominence (AMI); no defined AGI. Missed: 2025 valuations; 2026 collapse (so far); "no GPT-5 in 2025" on the label; GPT-4 output "can't be hooked to downstream programs"
• Pattern: near-perfect on low-bold reliability calls; wrong on every dated market call; never scores a miss by name; concedes capability in prose and not in the scorecard

Empirical vs normative

• Empirical: specific capability claims about what LLMs can and cannot do; predictions about scaling limits; hallucination analysis (v1); 2026 additions — lab unit economics (his 08-19 OpenAI numbers), incident forensics (Hugging Face, 07-25), CoT-monitorability
• Normative: the field should diversify beyond deep learning; regulation is necessary; deployment in high-stakes domains without reliability is irresponsible (v1); 2026 additions — pause OpenAI, receivership, replace Altman and Brockman, boycott, regulation-plus-liability, no blanket ban
• Unusual combination of strong empirical claims and strong policy advocacy — range, but harder to separate the science from the politics (v1). Sharper now: the policy programme needs the labs to look reckless, and the incidents of July 2026 gave him the material; grade the incident facts and the remedy separately
• Conflict note (CONVENTIONS rule 25): subscription-plus-books-plus-speaking business whose value rises with the "I was right" record; the record is curated by the author. Discount his scorecard, not his lists

Weak spots / open questions

• The critic-who-also-builds tension: Geometric Intelligence was acquired, but Marcus hasn't built a successful alternative to the systems he critiques (v1). Still true — and the alternative is now being built by LeCun with $1.03B, not by him
• "Scaling hits a wall" is partially vindicated but also partially wrong — frontier models kept improving through 2025–2026 even if returns diminished (v1). Confirmed harder: the largest run ever produced the model he praised
• Hybrid architectures are a recurring prescription without a proof of concept at frontier scale — easier to diagnose than to build (v1). Unchanged; AMI is the test and has not shipped
• Media visibility and contrarian positioning create incentive for provocative claims — audience capture risk (CONVENTIONS rule 25) (v1). Confirmed by behaviour: concessions in prose, none in the scorecard
• The gap between "LLMs have real limitations" (true) and "LLMs won't be economically transformative" (much less clear) is where his critics have the strongest case (v1). Now the gap between "unreliable" and "30 million paid seats"
• NEW — the unacknowledged reversal: "agents unreliable" (2025 #9) and "output can't be hooked to downstream programs" (2022 #4) became "unleashed agentic AI hacking the internet at scale" (2026-09-18). A danger claim is a capability concession; the scorecard still shows both originals as correct
• NEW — the unnamed seventeenth: which high-confidence 2025 prediction he counts as wrong is not stated anywhere I could find; until named, the self-score is not a score
• NEW — Robust.AI role: "Founder and Executive Chairman" in bios, absent from the company's leadership page and from his own Substack bio; resolve with a primary (board page, filing, or his statement)
• NEW — Says versus Does on scaling: he cites Sutskever's words as vindication and does not cite Sutskever's NVIDIA deal or OpenAI's 100,000-GPU run; the same asymmetry the bench flagged on Sutskever runs the other way through Marcus
• NEW — the paywalled forecasts ("Two ways it might all fall apart", "Leopold's Folly") are the ones with mechanisms; only the free, undated retreat ("sooner or later") is on the public record
• NEW — "no country will take a decisive lead" is unfalsifiable as written (rewrite proposed above)
• NEW — the safety turn borrows the doomers' evidence (Coxon, Adler, Tao) while rejecting their conclusion; whether "catastrophic but not existential" is a position or a positioning is not decidable from the record

Rich's take

• (your synthesis here)

Delta log

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

• Grades: 19 Held / 10 Drifted / 5 Wrong (2 unfalsifiable items marked for rewrite; 4 live 2026–27 predictions with checks attached; 3 table rows left unchecked and marked so). Review window 2026-06-16 → 2026-09-21. Search budget hit the session cap (200) mid-run; the remaining checks ran on fetched primaries only, and every unverified item is labelled in the file.
• Wrong (5): 2025 #21 valuations start to fall (NVDA $4.53T on 2025-12-31, +37.80%, StockAnalysis); 2025 #25 no GPT-5 in 2025, on the label (launched 2025-08-07, Wikipedia); 2026 #5 "Wall Street began to lose confidence" (NVDA record $5.37T on 2026-09-18; NVIDIA revenue +106%, 2026-08-26); "2026 will likely the year which it all falls apart" (2025-12-26), so far; 2022 #4 "can't be hooked to downstream programs" (30M paid Copilot seats, Nadella, 2026-07-30; agentic breaches, July 2026).
• Drifted (10): LLMs don't reason (his own IMO concession, 2025-07-22; Astra 63% ARC-AGI-3, 2026-09-03); hallucination architectural (OpenAI's incentive explanation, 2025-09-04); scaling has limits, and the 2022 diminishing-returns prediction (Says vindicated, Does contradicted — 100,000-GPU Astra run); hybrids needed, and hybrids-outperform (AMI funded 2026-03-09, unshipped; IMO gold in natural language); 2025 #20 cautious enterprise adoption (Copilot 30M); 2025 #24 cyberattack (right shape, 2026 not 2025); 2026 #6 midterms (AI money, "any reference to artificial intelligence" missing from the ads, NBC, 2026-02-27); 2022 #7 structured systems (funded, not shown).
• Held (19): AGI elusive (Huang undefined, 2026-09-06); regulation urgent, still endorsed with escalated remedy (2026-09-04, 2026-09-17); 2019 deep-learning-insufficient (open); 2025 rows #1, #2, #3 (NVIDIA NI $59.688B; OpenAI op loss $20.92B, 2026-06-15), #4, #5, #6–8, #9 (for 2025), #10 (Tesla Q1 2026), #12, #13, #16 (BLS, Aug 2026), #18, #22–23 (AMI, 2026-03-09); 2026 rows #1, #2, #4; 2022 #1–3, #5, #6 (as one line); Rebooting AI prescience; dated-lists habit; builder credential (dated); cried-wolf pattern.
• Most surprising delta: the "agents won't work" critic now warns that agents work well enough to hack the internet at scale, and has moved from "regulate" to "Pause OpenAI, now" and "boycott generative AI" (2026-09-18; 2026-09-04; 2026-09-09) — a valence flip on his own central axis, unscored, while opposing the one ban bill on offer (2026-09-03).
• Employer / incentive check (mandatory): no job change — NYU emeritus unchanged (Substack profile bio; Senate witness list 2023-05-16). Robust.AI role conflict logged (bios say Executive Chairman; company leadership page silent; own bio omits it). No new company, raise, or paid seat found (search 2026-09-21). Incentive map re-derived: audience 118K+ (#23 Technology), paywall in use, two unscored capability concessions, agidefinition.ai co-authorship. Dated post-review sources: his posts 2026-08-18 → 2026-09-19 (fourteen fetched); Zitron 2026-08-25 and 2026-09-01; NVIDIA 2026-08-26; Nadella 2026-07-30; BLS Aug 2026; Wikipedia GPT-6 (Astra 2026-09-03/04).
• v1 corrections: Geometric Intelligence co-founded 2015 per Wikipedia (v1 gave no founding year; bios say 2014 — flagged); the Uber tenure was four months; Robust.AI was missing from v1 entirely; the 2025 prediction list was graded as one line in v1 and is now a per-row table; the 2024 "list" is the 2022-12-25 GPT-4 seven, and one of them fails.
• Tier: quarterly — the 2026 list resolves 2026-12-31 with his own December post and self-score; the over-50% cyber call runs to 2027-09-04; the Brundage bet to 2027-12-31; two live IPO calls; and the safety turn is one month old. Confidence unchanged at high: the public record is the deepest on the bench, and the grades rest on his own words plus market-level primaries.
• Learnings filed to the run report, not to _validation-learnings.md (shared file, bulk-run rule).

Sources

• His own, in window (primary, fetched): Young adults more concerned than enthusiastic (2026-08-18, cites Pew; numbers not shown) · OpenAI's unraveling has begun (2026-08-19) · Two ways it might all fall apart (2026-08-24, paywalled) · Anthropic's $30 trillion fantasy (2026-08-25) · 5 lessons from the OpenAI / Hugging Face incident (2026-08-28) · Red Alert (2026-09-02) · Hot take on GPT-6 Astra (2026-09-03) · Sanders–Casar ban — why I oppose it (2026-09-03) · Pause OpenAI, now (2026-09-04) · Huang claims AGI has arrived (2026-09-06) · The Case for Boycotting Generative AI (2026-09-09) · Two dire warnings — Tao, Coxon (2026-09-09) · No, Anderson Cooper (2026-09-10) · Two rays of hope (2026-09-10) · Two cheers (out of three) for Dario Amodei (2026-09-13) · Liability, regulation, and AI's new false dichotomy (2026-09-17) · The real reason Trump is standing behind AI (2026-09-18) · Wake up, people — agentic AI (2026-09-18) · Top three ways Dario Amodei has blown his credibility (2026-09-19) · 2026 sitemap (106 posts) · Marcus on AI (118,000+ subscribers) · Substack profile bio
• His own, the dated lists (primary, fetched): What to Expect When You're Expecting… GPT-4 (2022-12-25) · Where will AI be at the end of 2027? A bet (2024-12-30) · 25 AI Predictions for 2025 (2025-01-01) · Six (or seven) predictions for AI 2026 (2025-12-20) · The AI bubble is all over now, baby blue (2025-12-26)
• His own, earlier (primary, fetched): DeepMind and OpenAI achieve IMO Gold (2025-07-22) · GPT-5: Overdue, overhyped and underwhelming (2025-08-09) · Game over. AGI is not imminent (2025-10-18) · What happens next, after the decline of tokenmaxxing? (2026-05-29)
• Adversarial / third-party scoring: Forecast Fools — 19 tracked, 14 resolved, 96%, boldness 1.1/5 (undated page) · Fortune — backlash validating one critic (2025-08-24)
• Lab and market primaries used for grades: DeepMind — IMO gold, 35/42 end-to-end in natural language (dated 2025-07-21) · Apple — The Illusion of Thinking (June 2025) · Kalai et al. — Why Language Models Hallucinate (2025-09-04) · GPT-5 (launched 2025-08-07) · GPT-6 Astra (preview 2026-09-03; over 100,000 GPUs) · AMI Labs $1.03B (TechCrunch, 2026-03-09) · NVIDIA Q2 FY2027 (2026-08-26) · NVDA market-cap history · Copilot 30M paid seats (Nadella, 2026-07-30) · Zitron — OpenAI audited 2025 financials, FT-verified (2026-06-15) · Zitron — AI Hater's Manifesto (2026-08-25) · Zitron — Hyperscale Normalization (2026-09-01) · Tesla Q1 2026 shareholder update (8-K exhibit) · BLS Employment Situation, August 2026 · NBC News — AI super PAC ads (2026-02-27)
• Role / bio: Wikipedia · Senate Judiciary — Oversight of A.I. hearing (2023-05-16) · TechCrunch — Robust.AI $15M Series A, Marcus CEO (2020-10-28) · Robust.AI leadership page (read 2026-09-21; Marcus not listed) · Aspen Digital AI Elections Initiative — Advisory Council · agidefinition.ai
• Cited but not fetched (URL only; no facts taken from them): NYT op-ed "The Fever Dream of Imminent Superintelligence Is Finally Breaking" (2025-09-03, per Wikipedia) · CNBC video on Chinese models (2026-06-26, 403) · "Leopold's Folly" (2026-08, paywalled)
• v1 sources (carried): Books — Rebooting AI: Building Artificial Intelligence We Can Trust (2019, with Ernest Davis); Taming Silicon Valley (2024) · Substack — Marcus on AI, "25 AI Predictions for 2025" (2024-12; the post is dated 2025-01-01) · The Economist (Babbage podcast) — "A Sceptical Take on AI in 2025" · Freethink interview on AI's moral and technical shortcomings · Congressional testimony on AI regulation (dated above)

migrated v1→v2 2026-09-21 · backup: _archives/gary-marcus.html.bak-20260921 · converts-from: people/gary-marcus.md · schema v1 → v2 · AI & Society domain · no --investing twin