AI News Top 10
3 November 2026
Anthropic has disclosed nearly $900 billion in compute commitments, a total that could pass $1 trillion, while its revenue growth slowed in the third quarter.
Anthropic's compute bill could pass $1 trillion, filing shows
A revised S-1/A filing lists nearly $900 billion in compute and infrastructure commitments, and estimated extra payments push the total past $1 trillion. About $70 billion falls due before September 2027, and only about $213 billion of the contracts can be cancelled. Revenue over the first nine months of 2026 reached $33.2 billion, up 1,206%, but third-quarter growth of 46% trailed a 77% rise in compute costs. Deals with Microsoft, Google, Amazon and Broadcom total $437 billion and run whether or not a single GPU is switched on.
Anthropic warns investors it is close to self-improving AI
In the same S-1/A filing, Anthropic writes that its internal estimates suggest it could hit the threshold for dramatically accelerated progress attributable to AI in its Responsible Scaling Policy in the near term. That threshold is the company's marker for recursive self-improvement. No current model has crossed it, the filing says, though Anthropic concedes it has spotted acceleration in its own development only after the fact. The company says it does not expect to let Claude design and build frontier models without meaningful human supervision.
Gemini 4 Carbon trails Claude Opus 5.5 but costs a third
SemiAnalysis published GDPx, a frontier benchmark of 206 tasks across finance, materials science, cybersecurity, software engineering, law and biology, with a score and a cost figure per model. Claude Opus 5.5 ranks first and Gemini 4 Carbon sits just behind at about $16 per task against $51 for Claude. OpenAI's GPT-6 Astra scores lower and costs more, around $54 per task. Remove cybersecurity, biology and alignment from the set and GPT-6 Astra moves to the top while Gemini 4 Carbon drops below GPT 6.1 Sol and Claude Sonnet 5.5.
MIT lets GPT-6 Astra rewrite another model's weights directly
Researchers at MIT CSAIL described weight-space programming, in which a language model inspects and edits the numerical parameters of another neural network without conventional training. With GPT-6 Astra they lifted Qwen3.5-0.8B from 7.9% to 48.1% on MATH-500 by changing about 0.016% of its weights. Most of the gain came from correcting repetitive reasoning. A Databricks researcher called the technique gene editing for neural networks.
New York Times and other publishers sue Brave over scraping for AI
The New York Times is suing Brave, the company behind the privacy browser and its own search index, for scraping copyrighted articles without permission and passing them on through its search API to AI customers including Anthropic and Perplexity. The complaint targets the intermediary instead of the model builder. Brave positions itself as a privacy-friendly alternative to Google.
UK prosecutor cites invented case in High Court after AI search
A senior crown prosecutor at the UK Crown Prosecution Service referred the High Court to CPS – R v J [2024] EWCA Crim 3529 in a case about prosecuting an 11-year-old. The case does not exist. Two High Court judges called the explanation unsatisfactory, and the Guardian reports the citation came from an AI summary produced during a web search. A tracker run by Doughty Street Chambers counts 101 registered AI-related errors in UK court cases.
OpenAI offers its data center power plants as grid backup
OpenAI announced a Grid Protectors Program, with a first pilot at a data center in Effingham County, Georgia. During extreme weather or emergencies the company promises to free up to 1,000 megawatts, enough to cool about 700,000 households. OpenAI frames its own generation as reserve capacity that should improve reliability for neighbours instead of adding load to the grid. The plan follows growing resistance to the power demand of AI data centers and to grid congestion.
VOLTA infostealer uses Claude, Gemini and Kimi to rank victims
Researchers found VOLTA-AUTODOXER, a Windows infostealer that harvests passwords and cookies and then uses a language model to sort the haul by resale value. The model works as an analyst that decides which stolen data can be sold or abused first, and the malware falls back to other models when one refuses. The builder kit automates a step that used to take a human.
Zyphra presents hierarchical attention as sub-quadratic route
Zyphra AI introduced Zyphra Hierarchical Attention, a tree-indexed sparse attention method that is applied to existing frozen models with 4 billion distillation tokens. The company says ZHA matches dense attention on reasoning and retrieval benchmarks up to 256K context, while compute scales differently from classical quadratic attention. The claims have not yet been tested independently.
Wall Street executives reject the idea that AI is a bubble
On a Bloomberg panel, Goldman Sachs chief executive David Solomon, Franklin Templeton's Jenny Johnson and Morgan Stanley's Ted Pick rejected the suggestion that the AI boom is in bubble territory. Pimco's Emmanuel Roman said the debt wave in the sector has not yet disturbed broader credit markets. Solomon described demand for AI deals as very robust and spoke of a technology supercycle that will dominate capital markets.
Key Trends
- • Anthropic's filing turns the bubble debate into a balance-sheet question: third-quarter revenue grew 46% against a 77% rise in compute costs.
- • Two of the ten items come from that same S-1/A, including the warning that recursive self-improvement is near.
- • Cost per task enters model selection: Gemini 4 Carbon scores just below Claude Opus 5.5 at a third of the price.
- • Research moves from prompting models to editing them: 0.016% of the weights lifted MATH-500 from 7.9% to 48.1%.
- • Publishers now sue the data layer as well as the model builders.
Wat vond je van dit bericht?
