Yesterday we talked about Google's "big week." Today we're seeing what happens when AI companies stop talking about what their models can do and start measuring how well they actually do it in the real world.
01
Former Meta PM Peter Yang breaks down how Anthropic builds Claude
Product manager Peter Yang shared his top five takeaways from Anthropic's Alex Albert on building the next Claude model. The key insight: the model and its "harness" (the prompt and tool setup) are coupled together, so the same model gives different responses depending on whether it's running in Claude, Cowork, or Claude Code. Albert also revealed that Claude is "starting to dream" when agents aren't actively running tasks.
Why it matters: This is the first detailed look at how Anthropic thinks about deploying one model across multiple products. Every AI company building agents is wrestling with this same coupling problem.
Developer Thariq talks long-running agents on Code with Claude
Developer Thariq appeared on Anthropic's Code with Claude podcast to discuss "staying in the loop with long running agents." The conversation covered how to monitor and manage AI agents that run tasks over extended periods.
Why it matters: Long-running agents are the next frontier for AI coding tools, but nobody's solved the "how do I know if my agent is stuck or making progress?" problem yet. This is one of the first public discussions of real solutions.
Together AI claims 76% cost advantage over Claude for coding agents
Together AI published new benchmarks showing their inference platform delivers 31% more tokens per second than TensorRT-LLM, 2× better time-to-first-token at saturation, and 76% lower cost than Claude Opus for coding agent workloads. These are the first real-world performance numbers we've seen for coding agents at scale.
Why it matters: If you're running coding agents in production, these numbers suggest you could cut your AI bills in half by switching platforms. That's the difference between a viable business and burning cash.
AI developer Swyx suggested that someone should try live-coding a specific project as a 2-3 hour workshop for people to learn from. He's looking for volunteers to attempt the challenge publicly.
Why it matters: Live-coding with AI tools is becoming a skill worth teaching, but most "AI coding" tutorials are still edited demos that hide the messy debugging reality.
Former Everlane executive Nan Yu reflects on brand resurrections
Nan Yu, who used to work at fashion brand Everlane, posted about the company's struggles and how brands can come back from the dead, citing Ray-Ban, J.Crew, and Abercrombie as examples of successful resurrections after appearing to be finished.