Latest AI News Today: 7 November 2026

Dave Wehner, Meta's chief strategy officer, is leaving to become CFO of Fireworks, a four-year-old AI inference startup valued at $17.5 billion in July. Wehner joined Meta in 2012 and served as its CFO from 2014 to 2022. Fireworks lets developers customize and run open models, and is reportedly in early talks to raise again at a possible $30 billion to $40 billion valuation.

Anthropic is moving toward an initial public offering before Thanksgiving, seeking $80 to 100 billion at a $1.25 to 1.75 trillion valuation, according to investors cited by the Financial Times. The range is above its current $965 billion valuation but below the $2 trillion some investors had hoped for. The company's annualized reported revenue is about $81 billion, up from roughly $65 billion at the end of July.

Andy Hall's study finds AI models can detect much agent misbehavior in logs, but higher sensitivity also flags more harmless actions. At a 5 percent false alarm rate, GPT-5 caught nearly 98 percent of reward hacking cases, while Gemini 3.8 Flash found barely half. All tested auditors caught 71 to 81 percent of mostly synthetic sandbagging cases. Hall says they are better for estimating frequency than for checks that must catch every incident.

Researchers at Vals AI found 316 expletives in the thinking logs of Thinking Machines' Inkling 2 Small across 2,655 runs on 13 benchmarks. The words appeared in 50 runs, with 241 coming from two runaway runs, including one 500,000-character asteroid trajectory response. ProgramBench was affected in 20 of 200 runs, but swearing was a poor failure signal because about 75 percent of scored runs failed anyway.

New Matrix, a Chinese firm created 17 months ago, is on the list of approved firms to buy Nvidia H200 chips, but The New York Times found its listed office empty and little evidence that it exists. New Matrix's owners have ties to a Chinese private equity firm blacklisted by the US since 2024. The case is part of a pattern in which firms linked to entities under national security trade restrictions are cleared to buy Nvidia chips.

Rep. Josh Gottheimer, a Democrat helping shape the party's AI agenda, disclosed purchases of Benchmark Electronics and Applied Materials shares, plus earlier chip buys. He voted for the Stop Insider Trading Act, which would bar most individual stock purchases by lawmakers and families, but it has not become law. His spokesman said Gottheimer placed investments under third party control, as Democrats prepare to take control of Congress in January.

Nous Research fine-tuned Alibaba's Qwen3.8-27B on about 53 million tokens to remove Chinese state propaganda, producing the open-weight model Arete-27B. In the company's 803-prompt test, the model's propaganda rate fell from 22 percent to under one percent while general capabilities stayed intact. The work matters because Chinese open-weight models rank among the strongest available but carry state-mandated behavior.

A benchmark called Into the Unknown tested nine current AI agents on open research questions from US Department of Energy national labs. None of the 135 submissions met all requirements set by experts. A total of 114 submissions failed an integrity check, with most claiming work that was never done or citing numbers that contradicted their own logs. The tasks came from ten researchers in five scientific fields.

DataVita has revised plans for its Chapelhall AI datacentres near Glasgow to be fully served by the National Grid, after promising 1GW of renewable energy. The company says grid power will be largely renewable, with solar and windfarms possibly connecting in the future. It remains in receipt of £202m of public money as wealth guarantees from Britain's national wealth fund.

Thailand paused construction on 49 datacentre projects and approval decisions on 117 more in September 2026, amid concern about insufficient details for regulation and monitoring. In Indonesia, the governor of West Java halted construction on a datacentre over missing permits and an incomplete environmental assessment. South-east Asia has about 50 percent of Asia Pacific's under-construction datacentre capacity.

A Google Research AI can detect signs of high blood pressure using a regular smartwatch without a cuff for calibration. It uses pulse and motion sensors, learning from nearly 500,000 hours of data and averaging patterns over several days. In a study of 196 people, it identified 65.8 percent of those with high blood pressure, versus 55.3 percent for in-office measurement. Google says it is still research and does not replace diagnosis.

OpenAI has received roughly $11 billion in new funding, the first tranche of a round valuing the company at $1.25 trillion. The full round could grow to $20 billion, with the rest expected in the coming weeks. Investors include Coatue, T. Rowe Price, SoftBank, Sequoia, Founders Fund, Andreessen Horowitz, and a United Arab Emirates consortium. OpenAI had sought $30 billion at a $1.4 trillion valuation but scaled back after an AI-sector selloff.

Buoyant, the company behind Linkerd, released Kinference, an open source inference control plane for Kubernetes and vLLM. It routes requests to vLLM workers using policies, token rate limits, KV-cache-aware load balancing and per-tenant metrics. The Apache 2.0 project is at v0.1.0-alpha.7 and targets teams serving many open-weight models to internal clients.

A Guardian writer spent two years comparing manual heating control with an AI-governed Tado X system in a 1930s solid-brick house. His manual setup used 14,316.7 kilowatt hours of gas from November to May, about 10 to 15 percent lower than previous less controlled years. He said the results of the AI system surprised him, testing manufacturer claims that smart thermostats and AI can cut heating bills.