The finding about removing the 9.8 GB Metal LRU cache for a 38% speedup is the most interesting part. Same lesson as PostgreSQL's advice against application-level buffer pools that compete with the OS page cache : the hardware memory compressor doing 130K decompressions/sec was pure overhead.

Curious about the remaining gap: 5.7 tok/s vs 18.6 theoretical (from SSD bandwidth). Is the ~70% overhead mostly GPU compute on non-expert layers (attention, norm), or is there I/O scheduling room left?

Interesting tool. One thing I've noticed managing rules across Claude Code and Copilot: the same instruction produces very different results depending on the agent. Claude follows multi-step rules well, Copilot tends to ignore anything beyond the first line.

Does Skills Manager handle this at all, or is it purely format/distribution? Seems like the hard problem isn't syncing files — it's that the same "skill" needs different phrasing per agent to actually work.

This resonates. I spent years thinking I enjoyed coding, but what I actually enjoy is designing elegant solutions built on solid architecture. Inventing, innovating, building progressively on strong foundations. The real pleasure is the finished product (is it ever really finished though?) — seeing it's useful and makes people's lives easier, while knowing it's well-built technically. The user doesn't see that part, but we know.

With AI, by always planning first, pushing it to explore alternative technical approaches, making it explain its choices — the creative construction process gets easier. You stay the conductor. Refactoring, new features, testing — all facilitated. Add regular AI-driven audits to catch defects, and of course the expert eye that nothing replaces.

One thing that worries me though: how will junior devs build that expert eye if AI handles the grunt work? Learning through struggle is how most of us developed intuition. That's a real problem for the next generation.