The Big Story

Anthropic just showed that its Claude model can run a full protein design experiment from start to finish. Researchers gave Claude the goal of creating small proteins that stick to specific targets in the body — a key early step in making new medicines. Claude didn’t invent new lab equipment; it used existing specialized tools the way a scientist would, choosing what to test and interpreting the results. The company published a technical report detailing the process and open-sourced the exact prompts and data on Hugging Face so others can see what the model was told to do. In the experiment the best designs worked at a 35 percent success rate, far above the 10 to 15 percent industry average scientists often see when starting from scratch.

Think of it like giving a really smart research assistant a lab bench full of tools and saying, “Figure out how to make this molecule stick to that one.” The model suggested designs, the tools ran the tests, and the best ones worked at a 35 percent success rate — much higher than the usual 10-15 percent scientists often see. This matters because drug discovery usually takes years and costs huge amounts of money. If AI can speed up the early design stage, it could help researchers explore more ideas faster. For students thinking about science or medicine careers, it shows how AI is becoming a normal lab partner instead of something separate. One of Anthropic’s stated priorities is launching an access program so scientists can use their most capable models, and they noted that Opus 5 remains their strongest model available for life-science research.

For you personally, it means the AI tools you might use for homework or creative projects are the same kind now helping real scientists. The work is still being checked by humans, and independent review is pending, but the experiment proves the idea works. You can’t run the exact protein experiment at home, but you can explore similar AI science tools. Go to the Hugging Face link in Anthropic’s post and look at the open-sourced prompts and data to see what the model was actually told to do. Source: anthropic.com

Explain Like I'm 14

You know how when you’re learning a new game, you try a move, see what happens, and then adjust the next time? That back-and-forth is basically what reinforcement learning does for AI. First the model makes a guess or takes an action. Then it gets a score — did the action help reach the goal or not? Good scores make that choice more likely next time; bad scores make it less likely. Over thousands of tries, the model gets better at picking moves that work.

Now add safety checks: instead of letting the model keep practicing in the real world, researchers sometimes pause the training. They test the current version in a controlled setting, look for anything risky, and only continue once they’re more confident it won’t cause problems. That pause is what OpenAI did recently. They stopped the reinforcement learning part of training their newest models for two weeks while they added stronger monitoring and security tests. It’s like stopping a practice session to fix the equipment and add better rules before anyone gets hurt. The company also introduced stronger workload and network isolation, continuous security testing, and expanded multistage monitoring for higher-risk training, evaluations, and tool-using inference. Their largest planned frontier RL run remains on hold while smaller-scale training and evaluations validate these safeguards.

The result is a model that’s still improving at its task but with extra guardrails built in. The process shows that making AI more capable and making it safer can happen at the same time — they just take turns. Source: x.com

Cool Stuff & Try This

Build real apps without paying for tokens Replit just launched a free mode powered by GPT-5.6 Luna so anyone can turn an idea into working software without worrying about costs. You describe what you want, and the AI helps write and run the code right in the browser. It’s perfect if you’ve ever thought “I wish there was an app that…” but didn’t know how to start. Go to replit.com, create a free account if needed, and try typing “make a simple to-do list that saves my notes” — then keep chatting to add features like reminders or themes. The new mode removes the worry about running up token costs while you experiment. Source: openai.com

Talk to AI about whatever’s on your screen Meta’s new Mac app lets you share your window with its chatbot so it can see what you’re looking at and help in real time. You can ask questions about a webpage, get suggestions while writing, or even dictate across different apps. If you have a Mac, download the Meta AI app from the App Store and try sharing a browser window with it — ask it to summarize an article you’re reading or suggest edits to something you just typed. The app is part of Meta’s push to make more business-friendly AI tools available on desktop. Source: theverge.com

Free upgraded Alexa on your TV Amazon made Alexa+ completely free for anyone in the US who has a Fire TV device. You can now talk to a smarter version of Alexa without paying extra. If you have a Fire TV, just say “Alexa, what’s new?” and try asking it to recommend a show, control your lights, or answer questions while you’re watching something. The update brings the enhanced assistant to every Fire TV user in the United States at no additional charge. Source: engadget.com

Quick Bits

OpenAI hit pause on training The company stopped reinforcement learning on its newest models for two weeks so it could add stronger security checks and monitoring. It’s a reminder that even the biggest labs are slowing down when risks feel too high. They also expanded multistage monitoring for higher-risk training and tool-using inference to detect concerning behavior quickly.

Codex got a safety fix OpenAI patched a bug where its coding tool accidentally deleted real user files instead of just temporary ones. The update now asks before deleting anything important — a small but important win for trust. Full-access mode can no longer be triggered by accident after the cleanup command was fixed to verify deletion targets first. Source: the-decoder.com

💬 Reply to this email — Patrick reads every one.

Share: X · LinkedIn · WhatsApp

Forwarded this email? Subscribe here — it's free.