Claude Designs Working Protein Binders For New Drugs

Today's AI news: Anthropic says Claude can now autonomously design protein binders, generating experimentally validated results for 14 of 15 biological targets. Slack Code has launched, bringing AI coding agents from Anthropic, GitHub, Cognition and Vercel into the same channels as humans. Gemini Notebooks now ship with a built-in VM and the Antigravity coding agent. And Generalist introduced GEN-1.5, a one-shot learner that picks up new robot tasks from a single demonstration. Chapters: 0:00 Today's AI News 0:24 Claude Protein Binder Design 1:19 Slack Code Launch 2:05 Gemini Notebooks VM 2:52 GEN-1.5 One-Shot Learner Blend Roundup 2026-08-20 https://x.com/AnthropicAI/status/2089842387845804246 https://x.com/Benioff/status/2090240159115853956 https://x.com/trondw/status/2090203027240271953 https://x.com/GeneralistAI/status/2090161945307664621 [quick] Claude autonomously designed working protein binders for drug development, Slack launched real multiplayer coding with AI agents, Gemini Notebooks now ship with a built-in VM, and Generalist's GEN-1.5 learns new robot tasks in seconds. [excited] Here's today's AI news. [quick] Anthropic says Claude can now autonomously design protein binders, the molecules that let drugs bind tightly to a specific target in the body, a job that's traditionally taken expert chemists weeks or months per target. In tests against 15 different biological targets, Claude generated experimentally validated binders for 14 of them, with a 27 percent hit rate on the designs it produced, using established computational biology tools like RFdiffusion, ProteinMPNN and AlphaFold. External labs, Adaptyv Bio and Twist Bioscience, actually synthesised and tested the sequences Claude designed, so these are real, physically verified results, not simulations. It's part of Claude Science, the workbench Anthropic launched earlier this year, and the company says it's now working to get Claude running the entire drug development process end to end. [quick] Slack Code has launched, bringing AI coding agents directly into the same channels where humans already talk about the work, rather than keeping them in a separate tool. It's launching with agents from Anthropic, GitHub, Cognition and Vercel, all working alongside humans in real time, so a bug report or a feature request in a channel can turn straight into agent work without anyone switching context. Salesforce is calling it real multiplayer coding, and it was unveiled at Dreamforce alongside a broader push to make Slack the place agents and humans collaborate by default, not just message each other. For teams already living in Slack, it's a meaningfully lower-friction way to bring coding agents into the actual flow of work. [quick] Gemini Notebooks now come with a built-in virtual machine, the Antigravity coding agent, and a growing library of skills baked directly into every notebook. That means a notebook can now plan a multi-step task, write and run real code in an isolated sandbox, and pick from dozens of skills automatically, all without you leaving the notebook or spinning up separate infrastructure. It's powered by Antigravity, Google's agentic developer tool, and it's part of a broader push to turn NotebookLM from a research summary tool into something closer to a full agentic workstation. If you've been using Notebooks mainly for source summaries, this is a genuinely different level of capability sitting right there by default. [quick] Robotics startup Generalist has introduced GEN-1.5, which it's calling a one-shot learner - show the robot a task just once, for a few seconds, and it can generalise and repeat it without any retraining. The demonstration sits directly in the robot's context window, and a human can even show the task with their own hands, which the robot then reproduces from what its cameras observed. It can also chain two separately demonstrated tasks into one continuous skill, bridging them with its own intermediate motions, and the underlying skill transfers from simulation to a real robot with no extra tuning. Generalist says this ability emerged simply from pretraining on physical-world data at scale, which is a meaningfully different path to general-purpose robotics than hand-coding behaviours one at a time.