Cursor dropped their Composer 2 deep dive today. Finally — actual technical details instead of marketing fluff. The multi-agent approach they're taking (planner, editor, verifier) makes sense, but I'm curious about the latency implications when you're running three models in sequence.

Simon shipped two things worth trying: a vulnerability checker that Claude built in one session, and a text layout library that skips DOM rendering entirely. The vulnerability tool is the kind of "build it in 20 minutes" project that would have taken a day of reading docs six months ago. The text library solves a real performance problem — calculating text height without triggering layout is one of those annoying web dev bottlenecks.

The LessWrong piece on "folie à machine" hit different. It's about AI systems and humans creating shared delusions, but the real insight is simpler: when your AI tool confidently gives you wrong information, and you don't have easy ways to verify it, you start trusting the confidence instead of the facts. We're already seeing this in code generation — the AI writes plausible-looking functions that subtly break edge cases.