Writing & Projects
Articles, notes, side projects, random thoughts.
Benchmark testing frontier models at driving a car in real life.
-
DrivingBench report↗
The evaluation design, cone course, results, and what we learned.
Reasoning effort
2 articles
For some reason I've been obsessed with reasoning effort levels and post-training length penalties, so here are some interesting notes on that.
-
"Juice Numbers" and the awkward way GPT reasoning efforts work
An investigation of how GPT reasoning-effort settings map to hidden juice values, affect prompt caching, and may be learned during post-training.
-
How models learn the price of thinking more
DeepSeek V4.1 and Cognition SWE-2's math behind reasoning effort