While everyone watched Astra, this quietly dropped: DeepSeek is building a 160,000-chip cluster in Inner Mongolia. Powered entirely by Huawei Ascend chips. Zero Nvidia. US export controls were meant to slow China’s AI. China’s answer: just build the cluster without them.

While the Kaggle competition demonstrated AIRA₃’s capabilities in a specific domain, the system itself can generalize across distinct domains: changing only the task specification. In an internal benchmark, AIRA₃ achieved a 27% latency reduction on production GPU kernels, and gold-level performance in another Kaggle competition translating 4,000-year-old Akkadian clay tablets into English. We're early, and hard problems are still ahead of us. But we believe a system that compounds its own knowledge is the right bet. As we continue to develop and scale, we’re excited about its potential to accelerate AI research and unlock recursive self-improvement.

While speaking yesterday at the G20 innovation summit about AI, I answered a question about the military actions in Iran, I misspoke. 18 members of our military have died, and to those families my heart breaks and I am so very sorry for their loss. I was thinking about the Venezuelan military operation where no service members died. As the 25th anniversary of 9/11 approaches, I am deeply thankful for our military and the sacrifices the men and women who serve our great country make, so such an attack, where I lost my brother Gary (36), my best friend Doug (39) and 656 other close friends and colleagues in the attacks on the World Trade Center, never happens again. Never!

While ‌a trip to a galaxy far, far away may still be out of reach, the Lucas Museum of Narrative Art in Los Angeles offered a closer look at the art and storytelling behind George Lucas' Star Wars franchise and countless others

While most are building tools on top of existing models, we’re taking a different path. We’re building our own model from the ground up, specifically for smart contract security. Not another wrapper. Not another skill. A model built for one purpose: finding vulnerabilities.

While I'm cleaning up my AI skills I have a question for the experts here. Sometimes I end up in this pattern: 1. I run my skill and it doesn't get it perfect in one shot 2. I do manual iteration with AI to get it right 3. I then ask AI something like: "How would you update the skill so we can one-shot this?" before reviewing and approving its edits. The problem is AI often overfits on this one thread and over time the skill drifts. Any good solutions here?