areys

The "dogfooding" framing is interesting. The argument isn't really about code quality per se — it's about whether you're in a position to critically evaluate the output at all.

A seasoned engineer using AI-assisted coding while maintaining judgment is genuinely different from someone using it as a substitute for judgment. The problem is the tools don't distinguish between these two users, and the marketing tends to speak to the latter group.

The use cases that open up when inference stays on-device are genuinely different. Health apps, journaling, anything where users are (justifiably) paranoid about their data leaving the phone — that's a big surface area that cloud APIs can't really touch. Surprised this is happening at the speed it is on consumer hardware.

The context window framing is compelling but I think the more durable pressure is the "blast radius" problem — not whether the model can hold your whole codebase in working memory, but whether a confident-but-wrong agent can cause cascading failures across the whole system.

Even as context windows hit 1M+ tokens, you'd probably still want bounded rollback units. Clean module interfaces help less with the comprehension problem and more with the containment one. Whether those modules get a network boundary is mostly an ops tradeoff — the architectural principle is the same either way.