A math problem that stayed open for 90 years just got solved in 88 hours. Roughly one hour of AI compute for every year humans were stuck. The Navier-Stokes equations describe how fluids move. Weather forecasts, airplane wings, blood flow, and ocean currents all run on them. In 2000, the Clay Institute put a $1M bounty on one question about these equations, and until today only one of the seven Millennium Prize Problems had ever been solved. The open question was whether smooth fluid motion can spontaneously break the math. OpenAI's agents say yes. They found a vortex that spirals inward and stretches like spaghetti until, at a finite moment in time, the equations produce infinities. The model of the fluid fails. Past that point you'd have to track individual molecules. Here's the detail most people will scroll past. When Perelman solved Poincaré, the proof took him 7 years and mathematicians spent about 4 more verifying it. OpenAI shipped this proof with a Lean formalization, so a computer can check every logical step mechanically. The verification bottleneck that used to take years compresses to running a program. The fight ahead is over whether it counts. The official problem statement accepts four routes to a solution, and this proof uses a smooth external force, the door many expected to crack first. Clay hasn't ruled. 10,000 coordinating agents. 88 hours. A model they say is well beyond GPT-6 Astra. And they're not even claiming the $1M.

Google have, dare I say... RELEASED A GOOD MODEL 🥹 Gemini 3.8 Flash provides ~Opus 5 performance at much lower cost, while being super fast

one side effect of treating LLMs like humans is people keep using them ineffectively you can't break a task up into 5 pieces and work on all of them in parallel and non-linearly but the agent can. you don't even have to tell it to, if it has the right environment it will do it

The software industry is entering into a security debt crisis, though many have yet to realize it. The exploitation of a slew of Bitcoin projects was just the canary in the coal mine.

Thinking Machines dropped Inkling-Small - a 266B multimodal model with open weights and 168K downloads last month. Here's what it actually does: > Accepts text, image, and audio inputs - outputs text > Built for agentic and tool-use systems, coding assistants, RAG pipelines, and general conversational use > Multilingual with strong English as the primary lane - Open weights, released for research, fine-tuning, and third-party integration A 266B open-weight multimodal model that handles three input types and targets agentic workflows is not a small drop. The open source multimodal stack is catching up faster than most people realize. HuggingFace link: huggingface.co/thinkingmachi…

Holy shit! This might be one of the most underrated releases of the year. Anthropic just gave Claude the ability to control physical lab equipment like microscopes, robotic arms, liquid handlers, lasers. The results are wild: – QuEra had Claude fix quantum computer lasers overnight, unsupervised. A fix that took human experts 5-10 minutes, Claude got down to 6 seconds. Success rate: 58% → 99.3%. – Genentech had Claude self-optimize lab pipetting by scoring its own results against expert baselines, and converged on the right settings on its own. – A university lab went from manual 4 am plate-swapping to Claude Code running the whole workflow hands-free. – Claude even ran qPCR experiments that helped detect human sewage contamination in a California creek. Automated science labs are closer than most people think. Anthropic is somewhat back!