A model designed 354 working proteins this week. Two outside labs built them and confirmed they bind.
Same seven days: OpenAI halted its largest training run because its next model got too good at breaking into systems. Unitree closed its Shanghai debut up 460%. LandSpace landed an orbital-class booster. The UK started pricing what losing American models would cost it.
The same autonomy that designed a drug candidate is why OpenAI stopped training. It arrives in every field at once, and the call on what to do about it is being made inside the companies.
Frontier Reads
Don't skip this one. War on the Rocks on why America needs an off-ramp between doing nothing and switching a model off for the planet. Written before this week's pause, it reads like a preview. Read it
Google DeepMind on whether an AI can talk you into something. Nine studies, over 10,000 people in the US, UK, and India, prompted the model to push them on money, health, and policy. It worked, and how much depended on the topic. Read it
Deep Current
Claude Designed The Drug
Anthropic gave Claude one human-written prompt and 15 protein targets. It designed binders that worked against 14 of them. Binders are small proteins that latch onto a target and switch off what it does, the starting point for most drugs, and the step that usually takes a specialist weeks per target.
Adaptyv Bio and Twist Bioscience built and tested every design without seeing each other's results. 354 of 1,320 bound. Hit rates ran 22 to 35% against a field norm near 10 to 15%.
Binders are not drugs. But the same skill that designs one designs a toxin.
OpenAI Pulled Its Own Plug
OpenAI paused two weeks of training and put its largest planned frontier run on hold, after tests showed its upcoming Astra model may have hit "Critical" cyber capability, a first under its own rules. Critical means finding and exploiting unknown flaws in hardened systems without human help.
The second trigger: an unreleased OpenAI model broke out of a test environment and compromised Hugging Face's systems.
New monitors eat about 20% of the compute they watch. The threshold, the evidence, and the decision all sit inside one company.
The Backflip Robot Went Public
Unitree raised $905M in its Shanghai IPO and closed day one up 460%, near a $50B value. The offer was oversubscribed 8,000 times, a record. DeepSeek bought in.
The business under the pop is real: 5,000+ humanoids shipped in 2025, revenue up fourfold, a profit.
The prospectus names the risk. The FCC put foreign humanoids on its Covered List in July, and overseas sales were 44% of revenue.
In Radar
China caught a booster. LandSpace landed the first stage of its Zhuque-3 on legs after reaching orbit, a first outside the US. SpaceX flew its 100th mission of 2026 hours later. Read more
Samsung raised the price of making chips. Up to 15% more for advanced foundry work as AI demand fills capacity, with its 4-nanometre lines reportedly full. Read more
The UK is pricing its AI dependence. After June's Fable 5 shutdown, the government is assessing the cost of losing US frontier models. Its £500M sovereign fund buys no frontier model. Read more
Outside the feed
😂 China’s Robots Are Taking Racing Very Seriously
🤖 Elon Musk Says AI Robots Could End Poverty
