The Top Story
On Monday, August 18th, Anthropic (the company behind Claude) published something that’s a genuinely bigger deal than the usual model-launch noise. They had Claude design protein binders from scratch, little proteins built to latch onto a specific target, which is the starting point for a lot of modern drugs. Then two independent labs, Adaptyv Bio and Twist Bioscience, actually built the things and tested them.
The results: Claude designed working binders against 14 of 15 targets. Out of 1,320 designs, 354 actually bound, about a 27% hit rate. The field typically runs 10 to 15%. On one target, Claude’s designs would’ve beaten 5 of 6 human teams in a real competition.
Why this matters for the rest of us: this is AI doing science, not just chatting about it. The old way took expert humans weeks or months per target. Claude did the design part largely on its own, orchestrating existing tools, in sessions measured in hours.
But before you get too excited, a protein binder is not a drug. It’s step one of a very long, very failure-prone road. Binding tightly in a dish is a world away from something safe and effective in a person. One former pharma exec called the results “not impressive” and picked at the details, which is a useful reminder that a company grading its own homework, even with outside labs checking, is still a company with a lot riding on the headline.
Quick Hits
The US wants countries to pick a side. Washington is reportedly leaning on other nations to choose between the American and Chinese AI ecosystems, warning that cozying up to China could cost them a seat at the US table. On August 19th, China pushed back, calling for “digital sovereignty.”
Qwen is running away with the open-model crowd. Alibaba’s Qwen hit around 2 billion downloads this year, passing the open models from both Google and Meta. If you’re tracking who’s winning in open-weight AI, it isn’t an American lab.
Copilot had a nasty security hole. Microsoft patched a bug (nicknamed CoSnitch) where one malicious link could hijack Copilot and quietly siphon your connected Gmail and Drive. It’s fixed, but it’s a clean real-world example of the prompt injection problem, an AI tricked by a link into leaking your stuff.
Okay, that’s the recap. See you tomorrow.