Today in AI — November 6, 2026

An eighteen-year-old quantum conjecture fell to a model that isn't public, a 65-billion-parameter open model beat American rivals ten times its size, and the bill for all this intelligence came due in the same week: flat enterprise spend, a hostile ballot, and a chipmaker now tracking its own shipments.

Models & Research

PSI says an Anthropic model proved a 2008 quantum conjecture, and Lean checked the proof. Physical Superintelligence, Alexander Wissner-Gross's physics lab, announced Friday that an internal research system running instances of a non-public Anthropic model produced a complete proof of the Aaronson–Ambainis conjecture on the third day of a run on PSI's own infrastructure. Lean 4 accepted it; PSI then handed the system related open problems and it improved the result twice before the kernel signed off. The conjecture says every bounded low-degree polynomial on the Boolean cube has an influential variable, which implies any T-query quantum algorithm can be simulated classically in polynomially many queries on most inputs — quantum speedups stay confined to problems with exploitable structure. The caveats are real: the formal statement still has to be matched to the original by human mathematicians, the prose paper is not machine-checked, and PSI staffer David Turturean scoped the problem, prompted the model and is named as responsible for the paper's correctness. The four researchers behind an August paper on the classical-simulation consequence called it "a huge result."
Ripple-65B Preview: a 65B dense model that beats US open rivals ten times its size on agentic coding. Ricursive Intelligence, a startup building chip-design tools, pretrained the model from scratch on 12 trillion tokens with a 1M-token context and released it under Apache 2.0 in BF16 and NVFP4, per the model card. It scores 80.52 on Terminal-Bench 2.1, against 80.1 for Reflection Beam 501B, 76.4 for Meta's 816B Muse Spark Open and 63.8 for Thinking Machines' 975B Inkling. The card is candid about where it loses: 76.5 on SWE-bench Verified and 44.3 on DeepSWE 1.1, where Muse Spark Open takes 60.0. A team that grew from one person to five trained it at 40% utilization with no custom kernels; it fits on a DGX Spark and generates 250 tokens per second on two B200s.
OpenAI's new wet-lab benchmark shows agents still misread the gel. BioBasicsBench, built with biotech firm Ligo, scores 100 tasks drawn from real experiments — gel and blot images, colony plates, microscopy stacks, sequencing reads — against expert-written rubrics, 18,500 attempts across 37 model settings. GPT-6 Astra averaged 66.7% of available credit at high reasoning effort, against 43.8% for GPT-5.6 Sol, OpenAI reports. The company published the failures alongside the win: Astra reversed which colony-PCR lanes carried the new plasmid, counted colonies roughly 12% high by marking empty agar and missing neighbours, and inspected contradictory microscopy evidence before dropping it from its conclusion.

Industry

Thinking Machines is raising $5 billion at a $40 billion pre-money valuation. Business Insider reported Accel is in talks to lead the round, more than tripling the lab's valuation from its last public one. The timing cuts both ways: it lands the same week its Inkling 2 Small topped the Western open-weight table, and the same week the card data below showed enterprise AI budgets flattening.
TSMC is weighing a move into memory at Nvidia's request. The Information reported, citing two people, that TSMC has begun drafting plans to manufacture 3D DRAM — a step that would put the world's largest chipmaker in direct competition with Samsung and SK Hynix just as the AI buildout has memory sold out. Nothing is signed, and TSMC has not commented.
Ramp's payment data: companies use more AI than ever and spend almost nothing extra for it. Token spend grew 1.5% across August and September, after 42% in July, then re-accelerated 8.9% into October, per Ramp's index. Lead economist Ara Kharazian points at the top 1% of firms — the cohort that had been clearing $10,000 per employee per month — slipping below that mark, which he says is enough on its own to flatten total revenue. Price cuts did the rest: OpenAI halved GPT-6 Sol and Luna pricing in September, and Anthropic trimmed Fable 5.1 and Opus 5.5 at a cost it put at about $600 million in forgone growth. Bloomberg reported Friday that Anthropic's IPO valuation expectations slid to $1.2 trillion–$1.3 trillion from roughly $2 trillion in early October.
Palo Alto Networks is in talks to buy Clutch Security for $500 million. Clutch, three years old, guards AI agents' access to enterprise systems; it was last valued at $85 million in January 2025 and is approaching $10 million in annual recurring revenue, putting the price near 50 times sales, TechCrunch reported. It would be Palo Alto's seventh acquisition this year. Neither company commented; The Information had the talks first.
Cyera is reportedly raising about $800 million at a $20 billion valuation. The Information reported the cloud-security startup's round would be led by Cyberstarts, up from $12 billion in June — a re-rating that pairs neatly with the Clutch talks. Agent access is becoming the line item CISOs cannot defer.
Microsoft has offered buyouts to 60,000 employees, about a quarter of its workforce. Only a sliver are expected to accept, the company says. The New York Times reported Friday on staff long used to stability now reading the offer as a statement about AI-era staffing, with the buyouts framed as cost control while capital flows into compute.
Cleanview: 40% of planned US data-center capacity sits in states whose incoming governors want a pause. Bloomberg reported the figure from the market-intelligence firm after Democratic governors-elect campaigned on slowing AI development. Voters backed restrictions or rejected pro-development measures in all but one of 41 local ballot questions across 10 states this week, per Ballotpedia's count cited by CNBC.
Protesters broke into Microsoft's Amsterdam construction site. About 100 people organised by the group Geef Tegengas entered the site, some scaling a crane to hang a banner, The Verge reported; Reuters put the crowd inside at roughly that number. The project is three 85-meter towers built by Pure Data Centres with Microsoft as sole tenant. Amsterdam banned new data centers until 2030, but this one was planned in 2016 and has navigated the loopholes — which is the pattern activists are now targeting across Europe.

Policy

The Justice Department charged three foreign nationals over an AI impersonation scheme that took $129 million. Prosecutors say they used AI to impersonate chief executives and attorneys at major international law firms, and charged wire fraud and aggravated identity theft. The indictment was unsealed Thursday; NBC News and the Guardian both reported it.
Nvidia will track where its chips go in Southeast Asia. Bloomberg reported that customers seeking Nvidia Cloud Partner status must enroll in a program that monitors the location of their accelerators, after criticism that the company has not done enough to stop chips reaching Chinese firms. Export policy is being enforced through the sales contract now, not just the licence.
A Democratic Congress is picking targets, and they are large. The Washington Post reported that early oversight targets include Apple, Meta and Elon Musk's companies, with threads running through Grok, an ICE-tracking app distributed in Apple's App Store, and Meta's agreements with local governments over data centers. Two years of hearings that went nowhere under Republican chairmen now have subpoena power behind them.
Elizabeth Warren is a top contender to chair Senate Banking, and Nvidia is in the frame. Bloomberg reported that Democrats are poised to step up oversight of the chipmaker over export controls and its circular financing arrangements with customers. The circular-deals question is the one Wall Street has been asking all autumn; a committee chair can make Nvidia answer it under oath.

Tools

Perplexity's Teammates are always-on agents, not one-off jobs. Each runs on its own virtual machine with a desktop and a browser, reports progress in its own thread and takes instructions through Slack, Perplexity says. It requires the $200-a-month Max plan, the same price as the top tier of this whole category — OpenAI's Dots starts at $100, Anthropic's Duet at $20 — and the cheaper tiers tend to hit their usage ceilings fast.
Amazon will let you text Alexa to shop. Select US customers are being invited to search, compare, add to cart and buy in natural language, or send a photo to find similar items, with a wider rollout promised in the coming weeks, TechCrunch reported. The timing is deliberate: it arrives before the holiday season, two months after Amazon blocked Meta's Muse from querying its product pages. Every platform wants the agent, and none wants to be the catalogue underneath someone else's.

If a proof can be machine-checked but the statement it proves still needs a human to confirm, what exactly has been verified? Write to us via our contact page.