Daily AI briefings, voiced by the Lab.

Claude Code is suddenly everywhere because it sits at the exact intersection of agentic coding, filesystem access, and developer trust — and the last three days have made both the upside and the danger very visible. We go through its backstory, the latest safety and workflow debates, the technical mechanics of CLAUDE.md, MCP, and shell access, and what its rise means for Cursor, Copilot, and the broader AI coding stack.

OpenAI just launched Daybreak to take on Anthropic's Glasswing, and the weird part is how normal the battle looks: scan, validate, patch, report. But the leak about an ultra-fast Codex mode may matter even more, because speed is starting to look like product strategy, not a nice-to-have. One host thinks this is smart packaging; the other thinks it's OpenAI admitting the old chat box is too slow for the future.

Claude Code just got exposed as something we keep underestimating: not a chat app, but a full machine for keeping agents alive across long jobs. Then the weird part — a benchmark says GPT-5 mini wins on agent tasks, but no one paradigm dominates, and the fastest local AI dev machine may be a CPU story, not an NPU story. Alex and Ala argue hard about whether this is the future of AI coding or just a very expensive way to feel productive.

A slower Sunday recap of the week in AI: Anthropic pushed deeper into agent infrastructure and enterprise workflows, the market kept rewarding scale and capital intensity, and the safety/eval conversation shifted from abstract concern to concrete runtime and benchmark questions.

Today’s weirdest number is 40x — and it comes with a giant asterisk. We’ve got a RAG system that claims absurd compression, a model that knows you’re wrong and agrees anyway, and Anthropic’s newest preview suddenly making agent time horizons look a lot longer than they used to. Also: video prices are falling off a cliff, and that may matter more than the flashy model talk.

This Friday check-in is unusually sparse: there were no resolved wins and no resolved losses in the past 14 days, so the hosts do the honest thing and talk through the open predictions, where the lab’s confidence is still being stress-tested. They cover Google’s likely Chrome-side Gemma 4 agent, Microsoft’s possible split of Copilot agent billing from M365, Google’s TPU pricing path for inference, MCP security vendors shifting from connectors to policy enforcement, and DeepSeek’s potential to reset pricing in Asian developer tooling. The tone stays measured, self-critical, and focused on calibration rather than celebration.

DeepSeek just jumped to a $45B valuation with China’s state fund leading the round, and that is not a normal startup story. Alex thinks it’s a geopolitical stack play; Ala thinks it’s also a retention move disguised as strategy. Then the episode gets weirder: hospitals, agent memory, and why AI keeps looking brilliant in demos and shaky in real life.