An AI Agent Designed Working Drug Candidates in a Lab, and No Scientist Was Watching

Anthropic let its Claude models run an entire drug discovery pipeline with no scientist steering the work, and the results held up when two outside labs tried to reproduce them. Broadcom is lining up as much as $100 billion in debt to keep the chip supply flowing to the companies building the next generation of models. In Beijing, more than two thousand robots opened a five day games built entirely around watching machines compete against each other. And in New Delhi, two Indian badminton players ended a wait that had lasted fifteen years.

An AI agent ran a drug discovery campaign with nobody watching

Anthropic gave Claude a single prompt, written by a human expert and running to about 30,000 tokens, then let it work alone. The agent researched 15 disease-relevant protein targets, picked which publicly available design tools to use for each one, generated candidate binders, and ranked the results for synthesis. No human scientist guided it during execution.

The campaign produced 354 confirmed working binders out of 1,320 designs, succeeding against 14 of the 15 targets. Depending on the setup, the hit rate landed between 22.6 and 35.1 percent, more than double the 10 to 15 percent that Anthropic says is typical for the field. Two outside firms, Adaptyv Bio and Twist Bioscience, independently built and tested the designs to confirm the numbers weren’t an artifact of Anthropic’s own lab.

Anthropic was careful about what this does and doesn’t mean. A protein binder is not a drug. Finding one that sticks tightly to its target is an early step in a much longer process of turning it into something that could ever be given to a patient. Still, doing that early step faster and more accurately than most human-run campaigns is a real result, not a demo.

Broadcom wants to borrow up to $100 billion to keep AI’s chip supply growing

Broadcom is in talks with a group of lenders, including Blackstone and Apollo Global Management, over a debt package that could reach $100 billion. The structure being discussed pairs roughly $30 billion in junior debt with a senior secured tranche of $60 billion to $70 billion, issued through a special purpose vehicle that Broadcom would partly guarantee.

The money is earmarked for custom AI chips and data center infrastructure for Anthropic and other AI labs. Under the arrangement, Anthropic wouldn’t buy the hardware outright. Investors would finance the purchase and lease the equipment to the company instead, a structure that keeps the debt off Anthropic’s own balance sheet.

This builds on a partnership Broadcom struck with Apollo and Blackstone in June, which raised $35 billion in its opening round to expand Anthropic’s computing capacity using Broadcom’s custom chips and networking gear. Together the partners have talked about financing more than 20 gigawatts of AI computing power by 2028, roughly the output of 20 nuclear power plants. Broadcom, which also designs custom chips for Google and Meta and supplies both Anthropic and OpenAI, expects its AI chip revenue alone to top $100 billion next year.

Beijing just opened a five day games built entirely around robots

The second World Humanoid Robot Games opened on Saturday at Beijing’s National Speed Skating Oval, drawing 666 teams and 2,056 robots from 16 countries. That’s up 138 percent in teams and 311 percent in robots from the first games a year earlier.

Over five days, the robots will compete in 51 events spanning 1,301 individual competitions. Thirty of those events are straightforward sports, things like sprinting and soccer built for machines instead of people. The other 21 are scenario based challenges meant to test how a robot handles a task rather than how fast it can do it, closer to a practical skills trial than a race. More than a thousand humanoid robots took part in the opening ceremony alone, under the theme “Intelligence Competing Toward the Future.”

It’s a showcase more than a scientific result, and the pace of the sport still lags well behind the marketing footage that circulates from these events. But the scale of the jump from last year, in both teams and machines, says something real about how fast humanoid hardware is being built and fielded right now.

India ends a 15 year wait for a badminton world medal

At the BWF World Championships in New Delhi, India’s Treesa Jolly and Gayatri Gopichand, ranked 39th in the world going in, beat three seeded pairs on their way to the women’s doubles semifinal. That alone guaranteed them a medal, since the tournament awards bronze to both losing semifinalists rather than playing a third place match.

They ran into the defending champions and top seeds, China’s Liu Shengshu and Tan Ning, and lost a tight three setter 21-18, 13-21, 8-21 after 71 minutes on court. The defeat still closed out a genuine milestone. It’s the first World Championships medal for an Indian women’s doubles pair since Jwala Gutta and Ashwini Ponnappa won bronze at the 2011 edition in London, and only the second time an Indian pair has medaled in the event at all.

None of these four stories wraps up cleanly. Anthropic’s protein binders are still years from anything a doctor could prescribe, Broadcom hasn’t closed its financing yet, the robots in Beijing are still more exhibition than industry, and Treesa and Gayatri will be back on court chasing gold at the next major event rather than resting on a bronze. Taken together, they’re a fair snapshot of where things stand right now: AI quietly doing real scientific work, enormous sums of capital chasing the hardware to run more of it, and, in New Delhi, two players proving that a 15 year gap can close in one tournament if you’re good enough on the day.