Leonid M's articles
- about 1 month agoAI Coding
Subagent Configuration Is Least-Privilege Design
Every knob on a subagent answers the same question: what may this thing see and do? Read them that way and the defaults stop looking reasonable.
- about 1 month agoAI Coding
Your Streaming Benchmark Measures Two Things
Streaming cut time-to-first-audio from 1574 ms to 450 ms. Only one of those two mechanisms scales with input length, and conflating them makes the number useless.
- about 1 month agoAI Coding
When to Use an Eval Framework, and When to Build the Judge
Off-the-shelf metrics cover the failures every LLM app shares. Your domain failures are not on that list. A decision rule for which half you are in.
- about 1 month agoAI Coding
The Bugs Only Deploying Finds
Four defects that no unit test can see, found the first time a working service met a real container. Three of them were invisible from the code.
