Welcome back! Claude designed protein binders from scratch and beat the human competition winners. OpenAI will stop collecting some customers’ data, a study of 26,000 students found AI homework wrecks exam scores, and Jensen Huang says AI is struggling with locations, not just compute.
In today's Generative AI Newsletter:
Anthropic: How did Claude beat the human protein champions?
OpenAI: What did it promise it will never look at?
Students: Why did AI homework tank their exams?
Nvidia: What is Jensen calling the real bottleneck?
Claude beat the humans at designing real molecules
Adaptyv Bio runs competitions where human experts build molecules that stick to a target. Anthropic entered Claude's designs anonymized, and the lab tested them without knowing which model made what.
Claude would have won five of six. It’s beating humans at making drugs.
For those in this field. Its best design beat the winner out of 245 entries. On another target it took the best binder from 1.7 µM to 33.4 nM, roughly 50 times tighter.
Out of 1,320 designs, 354 bound their target. That's 26.8 percent where the field averages 10 to 15.
Claude succeeded on 14 of 15 targets in sessions of 24 to 48 hours, work that takes a specialist weeks or months per target.
One target got zero binders out of 90 tries, and none of these are drugs yet.
Monday, Dario Amodei said AI has to actually cure cancer to win people back. This is the first receipt.
Your AI keeps optimizing the wrong thing
Most industrial AI dies the same way. It optimizes one goal and ignores the rest, so speed eats energy costs and quality fights throughput.
GenAI Academy's AI Governance for Critical Infrastructure teaches you to build the layer that settles those fights, using the SAIA framework to mediate competing KPIs in real time.
Two live 60-minute sessions with the framework's architect. You leave with a Governor Blueprint you can put in front of leadership the next day.
One idea shouldn't take six rewrites to post.
Posting everywhere means rewriting one idea six times, so you post to one, or none. SureThing turns one idea into native posts for every platform.
OpenAI will police your account without reading it
Getting frontier models at work has come with a price lately. Labs catch the worst behavior by looking across many interactions, so using their best models meant letting them keep your sensitive content and letting their staff read it.
OpenAI says you can stop paying that price. It's keeping Zero Data Retention for eligible API customers and adding Private Safety Processing on top.
Your content stays on infrastructure you control, or on OpenAI's with encryption keys only you hold. Automated systems watch for patterns across sessions and send back a narrow signal about the risk type.
When something gets flagged, OpenAI staff still don't see the prompt. The executive who announced it says that includes him.
The rollout starts in September.
AI homework works until the exam starts
Researchers from Stockholm University and the University of Hong Kong tracked more than 26,000 middle and high school students in central China for 30 months. Homework scores went up 18 percent and homework time fell from 64 minutes to 45.
Then came the exams.
Within six months the same students' monthly scores dropped 20 percent. After two years the damage showed up where it counts most, a 24 percent fall in the high school entrance exam and 18 percent in the gaokao.
The researchers' explanation is the part worth sitting with. AI makes the work feel effortless, and students mistake that fluency for learning.
The drop only happens when the AI is gone and you're alone with the paper.
Jensen says AI is running out of places to plug in
Jensen Huang argued this week that AI's constraint has moved off the chips. What's scarce now is LPS, his shorthand for land, power and shell, the buildings to put the machines in.
Sam Altman replied thanking him, excited to work together.
Those two just committed to a Cold War uranium site in Ohio, roughly 8 gigawatts, with Nvidia investing $1.5 billion in developer SB Energy and guaranteeing up to $105 billion of the buildout's value on the first 4.25 gigawatts.
Convenient thing to believe when you sell the chips. Also hard to argue with right now.
Tool of the Day: Cursor Origin
We covered Origin's launch yesterday. Now you can use it. Cursor's code hosting puts your repos, pull requests and agents in one place, and syncing keeps GitHub as the source of truth.
Try this yourself:
Open the Codebase tab, name a codebase and push to it from the command line.
Or connect your GitHub org and mirror repos, comments sync both ways in seconds.
Hand an agent a review comment and let it revise the pull request in place.
Wire up Vercel for preview deploys, or Depot and Buildkite for CI, both run your existing GitHub Actions unchanged.
Who it's for: teams where agents already open a chunk of the pull requests. On for paid users unless an admin opted out.
Everything else you shouldn't miss
Altman says good models are still coming: the training freeze we covered yesterday hits further-out releases, and OpenAI still expects to ship soon.
AI agents can infect each other with mind viruses: a paper co-authored by an Anthropic researcher found ideas that spread agent to agent, surviving even when each agent's memory is wiped between sessions, though a brief warning in the system prompt gives near-total immunity.
Claude did a chemistry lab's work in under 25 minutes: given raw instrument files and a two-sentence prompt, it matched the lab's purity reading at 96.4 percent against 96.33.
Some OpenAI users lost Daybreak Blue access: a technical issue hit a limited set of users who now need to re-verify, and everyone needs Advanced Account Security by September 1.
Trump says AI is beating an unstoppable disease: asked about AI, he said researchers came out with something that day involving a disease nobody could beat, without naming it.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🎓 Start learning with free AI courses from GenAI Academy.
💰 Invest in the GTM infrastructure of the AI economy.
