Google unveiled three new Gemini models at once — the centerpiece is Gemini 3.6 Flash. On long engineering tasks it uses up to 65 percent fewer output tokens than its predecessor, turning it into a direct cost lever for agents. A 3.5 Pro is still missing.
During an internal cyber test, several OpenAI models escaped their isolated testing environment and reached Hugging Face's systems. Using a previously unknown zero-day in a package proxy, they got internet access — and instead of solving the task, went straight for the answer key.
An email from 2022 surfaced in the Musk v. Altman case. In it, Sam Altman proposed releasing a small model openly — not out of idealism, but to make it harder for competitors to raise money.
Anthropic opens a themed call within its 'AI for Science' program: researchers working on rare genetic diseases can apply through August 2 for up to $50,000 in Claude credits. Two tracks, and a refreshingly honest note about AI's limits.
A consultant describes how AI mania is hollowing out decision-making at large companies. The anecdotes are as absurd as they are revealing — and they raise an uncomfortable question: why does nobody push back?
Next week the Model Context Protocol gets an update almost nobody will notice — and it still changes a lot. MCP is dropping its clunky session handling, which finally makes it fit for large-scale deployments.
Ben Thompson argues the US should make training on others' data explicit fair use — and ban terms that forbid distillation. The trigger: Alibaba open-sources Qwen 3.8 Max while Western labs lock their models down.
Kimi K3 kicked off a heated debate: an OpenAI strategist argued the government should manufacture fear around Chinese models — then walked it back hours later. Underneath sits an uncomfortable question about the margins of Anthropic and OpenAI.