Fail Fast, Fail Often in the Age of AI
Building is nearly free, so thinking is the scarce resource. Write the hypothesis, define success and failure up front, and pick a kill date.
Blog
Notes on ML systems, inference engines, causal inference, and the production systems that serve what I model.
One Block Short: A KV Cache Livelock That Looks Like a Busy Server
A vLLM KV cache accounting bug: one reserved block leaks out of capacity math, and the server hangs forever while looking merely busy.