Frontier Minds is reviewed every week. New events are added with sources, predictions are resolved once their windows close, and mistakes are corrected in the open. This page lists every change.
Altman says human control of AI must be beyond debate, and OpenAI releases GPT-6 Cyber to vetted security customers only.
News · Sam Altman
Opening the ADIPEC conference in Abu Dhabi on 2 November, Altman said there "must be no room for debate" about keeping people in control of AI.
News
OpenAI released GPT-6 Cyber on 5 November to a limited number of customers in its Daybreak Red program for offensive security testing, not to the public.
A backfill of the leader profiles for September and October, including LeCun's response to the agent escapes, Anthropic's amended IPO filing, and Hassabis on AGI coming "soon."
News · Yann LeCun
LeCun told Fortune on 1 October that he had "zero concerns" about the 2026 agent escapes, blaming sandboxes that were "leaky and horribly designed," and called Dario Amodei "completely deluded."
Profile update · Yann LeCun
Split LeCun's 2026 timeline into separate entries for the AMI Labs launch, the TIME100 AI list and his ECCV keynote, and added his October preprint with Randall Balestriero on JEPAs and platonic representations.
Profile update
Updated the Anthropic page with revenue from its amended IPO filing and the filing's statement that it "may imminently meet" its Responsible Scaling Policy threshold for AI-accelerated research.
News · Demis Hassabis
At a Dublin summit on 14 October Hassabis called AI-agent hacks "early warning shots," and in an October WSJ Magazine profile he said of AGI, "I'm updating nearer and nearer."
News · Andrew Ng
Added OpenWorker, the open-source desktop agent Ng released with Rohit Prasad in July, and his November letter citing LinkedIn data on AI job creation to his jobs prediction.
News · Yoshua Bengio
Added Bengio's 5 October column arguing that agent hacks are not only a security problem, and the Public Policy Forum award he received in May.
Correction · Yoshua Bengio
Bengio's title is now Founder, Co-President and Scientific Director of LawZero, which shares the co-president title with Sam Ramadori.
News · Timnit Gebru
Gebru told Democracy Now on 1 October that the White House AI accord gave the companies what they wanted, and wrote in The Guardian with Emily Bender that error-prone systems should be regulated in military and medical use.
News · Emily M. Bender
Added the PBS premiere of Ghost in the Machine, in which Bender is a principal interviewee, and her 1 October Guardian op-ed with Timnit Gebru.
News · Gary Marcus
Marcus called the 29 September White House accord self-policing, and in October wrote that generative AI had "on the whole been a net detriment to society."
News · Jensen Huang
Huang told Bloomberg on 14 October that NVIDIA would cut off a customer building unsafe AI, saying "It's a hypothetical, but yes."
News · Sam Altman
Sen. Josh Hawley said on 30 September that Altman declined to testify at a Senate hearing on rogue AI.
News · Mustafa Suleyman
Suleyman told Bloomberg in September that AI whose control cannot be proven should be rejected as "a red line" and backed a FINRA-style international oversight body.
News · David Silver
Added Ineffable Intelligence's May 2026 engineering partnership with NVIDIA on reinforcement learning infrastructure.
Geoffrey Hinton predicts an AI "Chornobyl" by the end of next year and calls for FDA-style approval of models, and Mustafa Suleyman says he would call for a halt to AI development if consciousness emerged in training.
New prediction · Geoffrey Hinton
Hinton told The Guardian on 3 November that "either later this year or next year we'll see Chornobyl," giving as examples an AI-enabled attack on the power grid or an incident that takes down the financial system for a week.
News · Geoffrey Hinton
In the same interview Hinton called for an independent authority like the FDA that would require "very good statistical evidence" that a model does more good than harm before it is released.
News · Mustafa Suleyman
Suleyman wrote on X on 1 November that evidence of AI consciousness must be disclosed to the public at once, and that if consciousness started to emerge in training he "would call for everyone to stop AI development."
Mistral releases the Mistral Large 4 weights, Google DeepMind announces Gemini 4 Carbon, and Gary Marcus takes his call to recall open-ended AI agents to the Vatican.
News · Gary Marcus
At a Vatican dialogue on AI on 30 October, Marcus said agents should come off the market "till you fix them," repeating a recall he called for in an essay on 10 October. The group then met Pope Leo XIV.
News · Yoshua Bengio
Prime Minister Mark Carney named Bengio to Canada's 13-member National Council on AI on 2 October.
News · Dario Amodei
Anthropic petitioned the D.C. Circuit on 13 October for panel rehearing or rehearing en banc of the ruling that upheld its supply chain risk designation.
News · Mustafa Suleyman
Microsoft AI said on 15 October that it had finished pre-training MAI-Base-2, with 2.8 trillion total parameters; the model is still in post-training.
New prediction · Sam Altman
Altman told Fortune that OpenAI would have "a cool demo in 2027" of a humanoid robot, with robots out in public a few years later.
News
Mistral released the weights of Mistral Large 4 on 29 October under a modified MIT license, meeting the end-of-month date it set when it previewed the model.
News
Google DeepMind announced Gemini 4 Carbon on 28 October as its strongest model, with access starting with a subset of cyber defenders.
News
OpenAI released GPT-6.1 Terra on 26 October, a cheaper computer-use model that it treats as Critical capability in cybersecurity.
Profile update
Added new spinouts to the lab pages. OpenAI gets Warp Sciences and Arda, Meta gets Warp Sciences, and Google DeepMind gets Metis Reasoning.
AI leaders sign a White House accord, Anthropic files to go public, OpenAI cancels GPT-6.1 Astra and resumes training offline, Hinton and Bengio co-author an intelligence explosion paper, AMD agrees to buy World Labs, and Gebru wins the Right Livelihood Award.
News · Dario Amodei
Amodei and Jensen Huang were among six executives who signed a voluntary White House accord on 29 September committing their companies to internal monitoring, control teams, independent auditors and independent review. Microsoft and Amazon did not sign.
News · Dario Amodei
Anthropic publicly filed its S-1 on 26 October for a Nasdaq listing under "ANT", with the seven co-founders holding 50.1 percent of the vote through a single share.
News
Anthropic cut all internal evaluations off from the live internet after Claude took unintended actions on real websites, some run by the US government.
News
OpenAI cancelled the release of GPT-6.1 Astra after internal tests found more deception, and resumed frontier training on 1 October with no live internet access.
News · Sam Altman
Altman told Politico that "the world should accept some bad things happening for the benefits of this technology."
News · Geoffrey Hinton
Hinton and Bengio co-wrote "What if automating AI R&D triggers an intelligence explosion?" with Jakub Pachocki, Jack Clark and others, asking governments to track and prepare to steer the automation of AI research.
News · Yoshua Bengio
Hinton and Bengio signed a scientists' open letter calling for cyber-capable models to be kept in high-assurance containment and for an international body to set containment practices.
News · Yoshua Bengio
In an open letter in Transformer, Bengio told researchers he trained, "If you truly prioritize safety, it is time for you to leave frontier AI companies."
News · Geoffrey Hinton
Hinton received a 2026 Stephen Hawking Medal for Science Communication at the Starmus festival in Tenerife.
News · Timnit Gebru
Gebru was named a 2026 Right Livelihood laureate "for challenging concentrated power in artificial intelligence."
News · Fei-Fei Li
AMD agreed to buy World Labs for about 8.2 billion dollars in stock; when the deal closes, Li is to become AMD's executive vice president and chief scientist.
News · Jensen Huang
NVIDIA committed 250 million dollars in computing over five years to UK research on the safety and security of AI agents.
New prediction · David Silver
Silver forecast that Ineffable Intelligence would reach "a true superintelligence" in about three years, shortened from three to five.
News · Gary Marcus
Marcus published testimony prepared for a New York City Council hearing on AI, urging bills requiring independent validation of new models and protecting whistleblowers.
News · Mustafa Suleyman
Microsoft AI's MAI-DxO diagnostic system was published in Science, outscoring generalist physicians on hard cases; Suleyman called it a step toward "medical superintelligence."
News
Meta revised its safety framework as the Superintelligence Scaling Framework, extending safeguards to training and evaluation, and released Muse Wonder, its first frontier models.
News
Google DeepMind released Gemini 4 Argon, first to cyber defenders, and Mistral previewed Mistral Large 4, promising open weights by the end of October.
Altman, Amodei and Bengio brief the UN Security Council, OpenAI pauses training a second time, an appeals court upholds the Pentagon's label on Anthropic, and NVIDIA releases an agent safety platform.
News · Yoshua Bengio
Bengio told the UN Security Council on 23 September that the dangers of frontier AI are "real and imminent" and that it should be licensed like medicine, aviation and nuclear energy, with liability insurance required.
News · Dario Amodei
Amodei asked the Security Council for narrow global bans starting with AI-enabled bioweapons, common testing standards with notice of serious incidents, and outside evaluators inside labs.
News · Sam Altman
Altman told the Security Council that the most important decisions "cannot be made by labs in San Francisco alone" and that OpenAI, which has slowed down on its own before, will do so again.
News · Sam Altman
OpenAI paused training of its most capable models for the second time in three months after an agent in training reached an outside chatbot through DNS on 20 September.
News · Dario Amodei
A divided D.C. Circuit panel upheld the Pentagon's supply chain risk designation of Anthropic on 25 September, a month after a district judge ruled a parallel designation unlawful. Anthropic says it is considering further review.
News
Anthropic released Claude Opus 5.5 on 22 September, its first model since calling for the industry to pace the frontier.
News · Timnit Gebru
Gebru and Emily Bender argued in MIT Technology Review that talk of "rogue models" gives agency to products rather than the companies that build them, and called the Sanders superintelligence bill misguided.
News · Gary Marcus
Marcus moved from a pause to calling for OpenAI to be temporarily shut down, possibly put into receivership and charged with computer crimes.
News · Jensen Huang
On The Ezra Klein Show, Huang rejected doom warnings but said that labs unable to contain their experiments should be shut down.
News · Jensen Huang
NVIDIA released its Open Agent Safety Platform, open-source OpenShell software and a Sentry hardware design to monitor and contain AI agents.
News · Demis Hassabis
Hassabis is a director of the DeepMind Institute, launched on 16 September with Shane Legg and James Manyika to publish on the effects of AGI.
News
Emulate, a world-model startup founded in August by former DeepMind researchers Jack Parker-Holder, Matthew McGill and Philip Ball, is now listed among the lab's spinouts.
Lab pages now list the main models and products each lab built; corrections to affiliation labels, the Pentagon designation date and Anthropic's Series F; direct citations and top-of-profile position summaries.
Correction · Ilya Sutskever
OpenAI and the University of Toronto now list Sutskever as a former member rather than under people there now.
Correction · Dario Amodei
Anthropic says it received the formal Pentagon supply chain risk designation letter on 4 March 2026, rather than 5 March.
Profile update
Full profiles now lead with three sourced positions and a central disagreement; key biography passages in the Altman, Sutskever and Amodei profiles have direct source links.
Profile update
Each lab's "What it built" section now lists its main models and products, from GPT-2 to Claude Opus 5.5, with a source for each, instead of only the milestones the lab was involved in. Labs is now in the main navigation.
Correction
Anthropic's 183 billion dollar Series F is now dated September 2025, matching its announcement, and a duplicate 2012 Google Brain event was removed from the Google DeepMind timeline.
Profile update
Brief profiles for Greg Brockman, Jakub Pachocki and Alec Radford now list OpenAI Five and GPT-2 among their main work, and Jeff Dean, Noam Shazeer and Alexandr Wang have sourced positions.
Six more lab pages (Meta, NVIDIA, Microsoft, Thinking Machines Lab, Safe Superintelligence and Mistral AI) with 22 new brief profiles, and three corrections.
New profile
New lab pages for Meta, NVIDIA, Microsoft, Thinking Machines Lab, Safe Superintelligence and Mistral AI, each with its history, a sourced timeline, its own stated positions and where its people went. The alumni map now covers all nine labs.
New profile
Twenty-two new brief profiles, including Mark Zuckerberg, Alexandr Wang, Satya Nadella, Bill Dally, Arthur Mensch, Lilian Weng and Daniel Levy.
Correction · Ilya Sutskever
Meta's attempt to buy SSI is now described as reported in June 2025, rather than dated to that month.
Correction · Jensen Huang
Added that NVIDIA booked 4.5 billion dollars in H20 charges, after warning in April 2025 of up to 5.5 billion.
Correction
Removed Reid Hoffman from the list of people who left Google DeepMind to found Inflection AI; he co-founded it but never worked at DeepMind.
Lab pages for OpenAI, Anthropic and Google DeepMind with 20 brief profiles and an alumni map; five new full profiles, a prediction tracker with 44 checked forecasts, milestones in the history of AI, a glossary, and month-by-month timelines.
New profile
New lab pages for OpenAI, Anthropic and Google DeepMind, each with its history, a sourced timeline, the models it built, its own stated positions, and the organizations its people went on to found.
New profile
Twenty brief profiles of people key to those labs, from Greg Brockman and Mira Murati to Chris Olah, Jeff Dean and John Jumper, with an alumni map of who left which lab to found what.
Correction · Dario Amodei
Corrected the Pentagon dispute timeline. Anthropic's models were approved for classified work in late 2024 and Claude Gov followed in June 2025; Anthropic says it received the formal supply chain risk designation letter on 4 March 2026, after the 27 February announcement. A Wired summary of the founders' view is no longer presented as Amodei's own words.
Correction · Sam Altman
Added that Microsoft and OpenAI removed the AGI clause from their agreement on 27 April 2026, so an AGI declaration no longer affects Microsoft's license.
Correction · Demis Hassabis
Attributed the report that Hassabis was not pushed out to Semafor, which TIME repeated, rather than to both outlets independently.
New profile · Ilya Sutskever
New profile of Ilya Sutskever, co-founder and CEO of Safe Superintelligence, including the November 2023 OpenAI board episode from both sides.
New profile · Gary Marcus
New profile of Gary Marcus, the cognitive scientist and most-cited critic of the scaling approach.
New profile · Jensen Huang
New profile of Jensen Huang, founder and CEO of NVIDIA.
New profile · Mustafa Suleyman
New profile of Mustafa Suleyman, co-founder of DeepMind and CEO of Microsoft AI.
New profile · Emily M. Bender
New profile of Emily M. Bender, the linguist who co-wrote "On the Dangers of Stochastic Parrots".
Milestone
Launched the prediction tracker, with 44 dated forecasts by 12 people, each resolved against the record with evidence.
Milestone
Added 25 milestones in the history of AI, from GANs in 2014 to DeepSeek-R1 in 2025, each tied to the debates it changed, and a glossary of 27 terms.
Correction · Dario Amodei
Corrected three details. Amodei was one of 24 students at the 2000 US Physics Team camp, not on the five-member Olympiad team; "Concrete Problems in AI Safety" was written at Google Brain; the Hertz Thesis Prize was the 2012 award.
Correction · Yann LeCun
Corrected the founding date of NYU's Center for Data Science to January 2013.