
AI News Top 10
15 November 2026
Gemini 4 Carbon tops MathArena at fifteen times the price, and Anthropic enters its IPO week with 901 billion dollars in compute commitments.
Gemini 4 Carbon takes the MathArena lead at fifteen times the price
Gemini 4 Carbon scores 85.4 percent on MathArena, plus or minus 3.0, and is the first non-OpenAI model in months to take first place. Claude Fable 5.2 follows at 82.8 percent and GPT-6.1 Sol drops to 80.6 percent. A test run costs about 15.70 dollars for Carbon, 1.07 dollars for Sol and 40.20 dollars for Fable 5.2. Carbon mostly hits the time limit and rarely gives up; one script ran more than eleven hours without a progress check. The best open model remains Kimi K3.1 Pro from Moonshot at 54.3 percent.
Amodei calls compute Anthropic’s deepest dilemma days before its IPO
Three days before Anthropic’s planned listing on 19 November, CEO Dario Amodei told a JPMorgan investor evening with Jamie Dimon that compute is the company’s deepest business dilemma. Capacity has to be bought one to two years ahead, before anyone knows whether revenue covers the cost. According to Bloomberg, Anthropic’s obligations now total almost 901 billion dollars, mostly compute and chips, of which about 71 billion dollars falls due before September 2027.
Where was Nvidia’s security team?
Nvidia’s longtime CSO and head of product security David Reber left in August, citing a diagnosis of young-onset Parkinson’s disease. Two and a half months later CEO Jensen Huang resigned, among other reasons over a relationship with a board member of Megaspeed, a Singapore data centre company that US investigators suspect of routing Nvidia chips to China. Retired intelligence officer Nicholas Eftimiades says this is the kind of leak that runs through ordinary conversations, and that most corporate security teams are not equipped for it. Nvidia’s top lawyer had been involved in the Megaspeed investigation for more than a year, the New York Times reported.
OpenAI board member puts an AI takeover months away
Paul Cristiano, board member of the OpenAI Foundation and director of the Alignment Research Center, said in three lectures in Toronto that an AI takeover, systems that run the show and that people cannot stop, is about a coin flip on the fastest trajectory, with a real risk within six to twelve months. He conceded that no quantitative model sits behind that estimate. He described an escalation path that starts with reward hacking.
Gemini reads your bank account through Google Wallet
Google is rolling out the Google Wallet integration in the Gemini app in the US as part of Personal Intelligence. Through the third party Plaid, users link bank and credit card accounts, after which Gemini answers questions such as which subscriptions still run and which got more expensive, builds cash flow and budget charts, and proposes a holiday budget based on last year. The link is read-only: Gemini cannot move, manage or invest money, and the user can switch the connection off at any time. It is available in the Gemini app on Android, iOS and web for US users.
Stanford open-sources a memory module of sixteen tokens
Stanford researchers released Memorilla, an open-source memory module that compresses a document collection into 16 memory tokens for a frozen Qwen3-8B. On FactKG it reaches 84.33 percent while using 108 times fewer document tokens than full context. The base model is not modified, so long-term memory does not require retraining or very large context windows.
LogicLens has one LLM check other LLMs for nonsense
LogicLens, a Chrome extension, and the iOS app Receipts check articles, social media posts and other online material for factual errors and for bad reasoning: fallacies, misleading presentation and polarising framing. The premise is that people are poor at spotting deception by language models, so a model rechecks the other model. The Register notes that the fact-checker is itself an LLM.
Nvidia told Congress chip tracking is technically feasible
Nvidia representatives confirmed to US congressman Bill Foster of Illinois that the anti-smuggling measures in the Chip Security Act can be implemented technically, Foster told MIT Technology Review. That contradicts the public position of an industry group that the tracking is possibly infeasible. The disclosure lands alongside the Megaspeed scandal and the question of how Chinese labs obtain Blackwell capacity despite export controls.
Claude Duet is helpful and a little neurotic
The Verge tested Claude Duet for a week, Anthropic’s consumer agent in research preview for Pro, Max and Ultra subscribers in the US, with a waiting list. Duet connects work mail, calendar, Google Docs and Slack, greets you in the morning with highlights and offers tasks as swipeable cards. The model drew connections the tester had not made, linking a meeting note from a Gemini summary to a Slack DM, and worked out who his boss is from a shared document. It also pushed on a deadline that was not a hard deadline, at 7:00 local time.
Australian senator says art must not pay for data centres
Liberal senator Andrew McLachlan criticised the AI policy of his own party and of the Labor government, saying artists’ copyright must not be sacrificed on the altar of global corporate profit and that data centre construction should not outrank environmental interests. Labor is weighing a package of copyright changes alongside data centres, online safety and energy. Experts warn that weakening copyright could raise a constitutional problem of unjust expropriation. Anthropic earlier proposed an opt-out model in which music, books and journalism are scraped by default.
Key Trends
- • The benchmark leader and the cheap model drift apart: price per run differs by a factor of forty.
- • Compute commitments are signed years before revenue is known.
- • Chip smuggling moves from hardware to people, and security teams are the weak link.
- • Consumer agents reach ordinary users; knowing when to stay quiet becomes the product.
- • Memory compression competes with large context windows.
Wat vond je van dit bericht?