An Anthropic researcher quit over safety and its alignment lead agreed
Jacob Coxon resigned from Anthropic after three years of pretraining research at OpenAI and Anthropic, saying the labs are racing to self-improving superintelligence, and Evan Hubinger, who leads alignment science there, replied that he is right and that Anthropic has no plan yet to solve alignment for superintelligence; OpenAI shipped ChatGPT Images 2.5 to every tier including free; and OpenAI said an internal model proved Navier-Stokes solutions can blow up, in a priority dispute with the two mathematicians working on it.
An Anthropic researcher quit over safety and a colleague agreed publicly Jacob Coxon resigned from Anthropic on Tuesday after three years of pretraining research, first at OpenAI and then Anthropic, posting that the labs are “racing straight to self-improving superintelligence and gambling with our lives”. Evan Hubinger, who leads alignment science at Anthropic, replied that Coxon is correct and that the company does “not yet have a plan to solve alignment for superintelligence and are not clearly on track to”; Anthropic has not responded as a company. Source
ChatGPT Images 2.5 reached every tier, including free OpenAI shipped Images 2.5 on Tuesday to all ChatGPT, ChatGPT Work and Codex users across desktop, mobile and web, and added Sketch, which lets you draw inside ChatGPT and use the drawing as a reference for the generated image. Two API models came with it: GPT-Image-2.5 Flare, which OpenAI says beats GPT-Image-2 on quality at half the latency, and Sunburst, a premium option aimed at holding detail steady across successive edits. Source
OpenAI says a model proved Navier-Stokes solutions can blow up OpenAI said on Tuesday that an unreleased internal model, running roughly 10,000 agents in parallel over 88 hours, produced a proof that solutions to the Navier-Stokes equations can develop a singularity in finite time — a result on one of the seven Millennium Prize Problems. NYU mathematician Tristan Buckmaster says OpenAI learned in the past week that he and Anthropic’s Levent Alpöge were making progress and adopted the method they were using; OpenAI’s Sébastien Bubeck says the run began on 1 September after the company heard rumours that two Millennium problems had been solved, and denies using their work. Source
DeepSeek routes all V4 Pro API traffic to V4.1 Flash today
From noon Beijing time today, every request DeepSeek receives on its deepseek-v4-pro endpoint is served by V4.1 Flash instead, at Flash pricing and with no code change on the developer’s side, until V4.1 Pro ships. The two-day test endpoint that let people try it first — deepseek-v4.1-flash-expires-on-0910 — opened on Tuesday afternoon and stops working today. Source
Suno launched v6, trained on licensed Warner and BMG music Suno released v6, v6-wild and v6-mini on Wednesday in partnership with Warner Music Group, BMG and Believe, and says rightsholders will be paid each time a subscriber generates a track with them; recordings from Universal and Sony, which are still suing, are not in the training data. Every earlier Suno model is being retired, as agreed in the company’s November 2025 settlement with Warner; v6 and v6-wild are limited to Pro and Premier subscribers and v6-mini is open to everyone. Source
Google is putting €13bn into Finnish AI data centres Google said on Wednesday it will spend at least €13bn over the next two years on data centres and supporting infrastructure at Hamina, Kajaani, Muhos and Vaala, which it calls its largest single investment in Europe. It has been in Finland since 2009, when it bought a shut-down paper mill at Hamina and turned it into a data centre. Source
US agencies named six Chinese AI firms as industrial-scale distillers The NSA, CISA and FBI put out a joint advisory on Tuesday accusing DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI of pulling billions of tokens across millions of requests out of Claude, GPT, Gemini and Grok since late 2024, and saying the campaigns are the core of those companies’ development strategy rather than a supplement to it. It tells US providers to watch subscription-to-usage ratios and immediate maximum usage from new accounts, and to deploy targeted response changes that reduce what a distillation campaign gets back. Source
DeepMind published the predicted effect of 9 billion DNA variants Google DeepMind released AlphaGenome Atlas on Tuesday, roughly a petabyte of predictions covering every one of the 9 billion possible single-letter changes to human DNA, free to academic researchers. It comes with an AlphaGenome Variant Impact score, on which 10 marks the most impactful tenth of variants and 30 the strongest one in a thousand. Source
An autonomous agent framework harvested credentials in under six hours Google’s Threat Intelligence Group said on Wednesday that a financially motivated group used an AI coding chatbot, one prompt and a set of markdown instruction files to plan, build and run a mass credential-harvesting campaign in less than six hours, with the agents managing the scanning pipeline, live troubleshooting and IP rotation without a human in the loop. GTIG also found an exposed command-and-control server running a production-style dashboard, called Recon, that had sorted and validated more than 23,800 harvested secrets, including credentials for cloud and AI services. Source