AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › People & Organizations

Gary Marcus

high confidence · updated 2026-09-14

Cognitive scientist; emeritus professor at NYU; the most persistent public skeptic of LLMs and scaling as a path to general intelligence.

Contents

Type: Individual (cognitive scientist, entrepreneur) Affiliations: New York University (Professor Emeritus of Psychology and Neural Science); formerly founder and CEO of Geometric Intelligence (acquired by Uber in 2016 to form Uber AI Labs) Notable recognition: Congressional witness on AI (alongside Sam Altman, May 2023); widely read Substack Marcus on AI

Bio Summary

American cognitive scientist whose research career at NYU focused on language acquisition, innate cognitive structure, and the limits of learning-only accounts of mind. Since the mid-2010s, has been the most prominent public critic of the "scaling solves everything" consensus in deep learning, arguing that LLMs are fundamentally limited in reasoning, reliability, and grounding — and that hybrid neurosymbolic approaches are required.

Founded Geometric Intelligence in 2014, sold to Uber in 2016 to seed Uber AI Labs, then returned to public-facing writing and policy engagement.

Current Role

Professor Emeritus at NYU; full-time writer and public commentator on AI via Marcus on AI; frequent congressional testimony witness; active in AI policy circles (particularly the Montreal Declaration for Responsible AI and similar initiatives).

Key Contributions

  • Author of The Algebraic Mind (2001), Kluge (2008), Rebooting AI (2019, with Ernest Davis), and Taming Silicon Valley (2024).
  • Founded Geometric Intelligence (2014); sold to Uber (2016).
  • Testified before the US Senate Judiciary Committee in May 2023 alongside Sam Altman — a defining moment in the popular framing of AI regulation.
  • Early and vocal critic of GPT-class models' reliability, grounding, and reasoning; documented failure modes and hallucinations years before they became mainstream policy concerns (see Sycophancy and Hallucination).

AI Policy Positions

Marcus occupies a distinctive position — not a safety-doomer, not a capability-maximalist — that has made him a recurring outlier on most debates.

  • LLM skeptic. Argues that current LLMs do not and cannot reliably reason, that hallucination and brittleness are structural rather than temporary, and that AGI via pure scaling is highly unlikely. Frequent public sparring with scaling advocates.
  • Pro-regulation. Supports strong, enforceable AI regulation — including licensing regimes, transparency requirements, and liability rules — and has criticised voluntary-commitments frameworks as insufficient. On this he is closer to the "pause" camp than to the "surgical intervention" camp.
  • Deep skeptic of industry self-regulation. Taming Silicon Valley is a sustained argument that AI companies will not internalise social cost without binding rules.
  • Opposed to hype-driven policy. Argues much of the AI-risk discourse (both utopian and catastrophic) is marketing rather than evidence, and has criticised both OpenAI and the "AGI-soon" camp on these grounds.
  • Diverges from LeCun on regulation. Agrees with LeCun that LLMs are not a path to AGI; disagrees strongly on regulation, where LeCun is permissive and Marcus is restrictive.

Responding to the OpenAI–Hugging Face evaluation-security incident, Marcus argued on July 22, 2026 for slowing down or pausing until security practices catch up (Source: garymarcus.substack.com). See AI Autonomy Risk.

On July 31, 2026 he published a response to Anthropic's report on three incidents in its cybersecurity evaluations, writing that "the supposed leaders in AI and AI safety are clearly in over their heads" and locating the cause in human error at Anthropic — allowing the models to reach the internet at all — rather than in model capability. The post collected two other reactions: investor Bill Gurley contrasting "mistakes were made" with "Claude did illegal things," and Zack Korman of Embroidery.io, who wrote that "a bad security culture is saying 'if we didn't catch it, who could have? This shows how hard this is.'" (Source: garymarcus.substack.com). See Investigating Three Real-World Incidents in Our Cybersecurity Evaluations (Anthropic Frontier Red Team, July 2026).

The Huang exchange and the definitional argument (September 2026)

When Nvidia chief executive Jensen Huang declared the race to AGI over on September 6, 2026, Marcus replied within hours that the declaration came with "no evidence and no definitions" and that declaring victory without a definition "simply muddies the waters." He pointed Huang to "A Definition of AGI", the 2025 paper he co-authored with Dan Hendrycks, Yoshua Bengio and 30 others, which operationalises AGI against psychometric theory and scores GPT-5 at 57% (Source: garymarcus.substack.com).

The reply also gave a specific rather than categorical assessment of Astra: of the ten items in his standing bet with Miles Brundage, Marcus wrote that he expects Astra may have reached autoformalization and possibly reliable coding, while doubting it has met the other eight. He recorded that François Chollet of ARC Prize had written on September 3, 2026 that saturating ARC 3 is not a claim of AGI, and that Huang had declared victory once before, roughly six months earlier, with respect to an earlier model. See Artificial General Intelligence (AGI).

The botnet forecast (September 2026)

With Nathan Hamiel of Kudelski Security and Zack Korman of Embroidery, Marcus published a joint response on September 12, 2026 to Dario Amodei's pacing essay, disputing its estimate that a misaligned agent swarm could take over the internet within 6–12 months (Could rogue agent swarms take over the entire internet in the next six months? (Marcus, Hamiel and Korman, September 2026)). The authors note first that the six-month framing carried in news coverage is not precisely what the essay says. They argue the claim as stated is too vague to assess and implausible on that horizon; that Cloudflare, Google and Amazon Web Services are unlikely to be compromised simultaneously and are better defended than typical sites; that attackers lack a motive or budget for an attack at that scale, and that how the agents would coordinate is unclear; and that since only Anthropic and OpenAI hold frontier AI, the warning is in effect about their own potential negligence. They cite a UK AI Security Institute finding that Anthropic's Mythos model could autonomously compromise only small, weakly defended systems. They do not dismiss the concern, noting that many individual sites remain poorly defended, that cost is a present barrier rather than a permanent one, and that local hardware improvements could make such attacks feasible outside large data centers.

The two co-authors are security practitioners rather than Marcus's usual co-signatories, and the essay's technical objections are theirs: Hamiel leads the AI, machine learning and data science track at Black Hat, and Korman runs an agent monitoring and detection company.

Notable Publications / Works

In a July 20, 2026 essay, Marcus argued that the US "is not going to win" the AI race, contending China has all but caught up; weighing seven possible policy responses, he preferred an international "CERN for AI" over bans on Chinese models or a regulatory moat (China Has All But Caught Up (Gary Marcus, July 2026)). He extended the argument in an open letter to presidential AI adviser David Sacks published July 24, 2026, writing that "China is closing the gap, but it's not because we have too much regulation on AI," and that the United States should consider cooperating with China on AI for global benefit rather than creating a new cold war (Source: garymarcus.substack.com). See US-China AI Competition: Different Races, Different Metrics, Open-Weight Frontier Models.

  • Rebooting AI (Marcus and Davis, 2019)
  • Taming Silicon Valley (2024)
  • Marcus on AI Substack (ongoing)
  • US Senate Judiciary testimony (May 2023)
  • "Have LLMs improved patient outcomes?" (Marcus on AI, May 3, 2026) — short synthesis essay echoing Eric Topol's clinical-outcomes review; folded into the wiki via Topol/Marcus: LLMs and Patient Outcomes — three-document cluster.
  • "No need to panic about Anthropic's new blog" (Marcus on AI, June 4, 2026) — Marcus's same-day pushback on alarmed reactions to Anthropic's "When AI Builds Itself", arguing the recursive-self-improvement framing is overstated. Characteristic of his "AI-risk discourse is often marketing" position applied to a safety-flavored lab post. (Source: garymarcus.substack.com)
  • "Sir Demis Hassabis vs Sir Demis Hassabis" (Marcus on AI, June 5, 2026) — flags Hassabis's two divergent AGI timelines (Stanford "2030 ±1" vs. Davos "5–10 years"), siding with the slower one and reading the gap as AGI-definition slippage. (Source: garymarcus.substack.com) See AGI Timelines.
  • "AI's Black Friday" (Marcus on AI, June 6, 2026) — reads the June 5, 2026 AI-stock selloff as evidence the generative-AI industry "may be in worse shape than people realize," argues SpaceX's GPU-leasing to competitors signals a weakening of the frontier-scaling thesis, and labels the prospective U.S.-government equity stake in OpenAI "crony socialism." (Source: garymarcus.substack.com) See AI Bubble Debate, AI Dividends (Universal Basic Capital, Digital Dividend, Global Dividend).
  • Response to the Hassabis frontier-AI framework (Marcus on AI, July 14, 2026) — Marcus, who advocated FDA-style preflight testing of AI models in his 2023 Senate testimony, welcomed Demis Hassabis's July 14, 2026 proposal for a FINRA-style U.S. frontier-AI standards body with pre-release safety testing, while cautioning that transparency and independence remain essential given U.S. government interest in taking stakes in AI companies. (Source: garymarcus.substack.com) See AI Pre-Release Vetting.

Relationships

Categories: People & Organizations | academic | opinion | us | frontier-models
What links here (59)