Blog
Premature Compilation: Why agent developers write too much code
Most programmers building AI agents reach for frameworks before they understand the problem. The instinct to write code is the bottleneck. The primary artifact of agent development isn't code — it's natural language heuristics. Here's why, and what to do instead.
You don't understand vCPU
Two servers. Both spec'd at "2 vCPUs." One runs 4-5x faster. The vCPU was an internal scheduling abstraction for hypervisors. Then it migrated to pricing pages and became the number buyers compare across providers. Same label, five different hardware realities, and nobody told you which one you're buying.
What Claude Code skills actually are (and aren't)
Anthropic positions skills as the application layer of computing. Reddit calls them "literally just prompts." The evidence sits closer to Reddit. But the process of creating skills reveals something neither side talks about: an involuntary audit of your own knowledge. Here is what skills actually are, what they can't do, and the one test that separates signal from noise.
Deep research has no bridge to fall
Deep research tools have no feedback mechanism. When the report is wrong, nothing falls down. The errors accumulate silently, and the market selects for the features that make this problem worse.
Ultrathink is back. It doesn't do what you think.
The community optimizes effort settings they can't observe against outcomes they can't attribute. I checked every quantitative claim I could find. Most don't hold up.
Why science works, and why that's a strange question
Science requires trust, a cooperative universe, and imperfect correction mechanisms piled on top of each other. Here's what actually makes it work — and where it breaks.