AI writes the code, humans are told to write the spec

Cisco gave one engineer unlimited AI tokens to rewrite legacy code, and the top replies asked what the rewrite is for, since the article never says why 98% of the code needs replacing. One reply from someone who has worked on Cisco gear for years said AI can hardly be worse. The Lean creator Leo de Moura interview made the complementary argument that as AI generates more software, people will spend their effort on specifications and properties instead. A Show HN put a working NES inside roughly 1GB of Qwen weights, and commenters immediately asked how anyone knows it is correct without testing every input. In the Ask HN thread on replacing Claude or GPT with a local model, the report was that Qwen 3.6 27B is about Haiku 4.5 level, but tokens per second on a Mac still lag the cloud.

The pattern: the argument has moved from whether AI can produce code to how you trust what it produces. Formal verification, specs and tests are becoming the scarce skill. Anthropic's move to stop Claude Code subscriptions working with OpenClaw, which it justified as outsized strain, also reminds everyone how much of this runs on vendor-controlled capacity.

One personal note from the archive thread is a 60-year-old saying Claude Code reignited his passion for building. Others say the work feels unfulfilling. Both camps agree the job has changed.