Drills
One Million Tokens, Four Needles
A huge context window is capacity, not proof of recall. This 60-minute drill tests retrieval, synthesis, evidence, latency, and the bill.
Tim Cheung5 min read
Tag
evals
2 pieces
Sparring
A Two-Point Lead Is Noise Until Proven Otherwise
Coding-agent leaderboards report decimals. The VM, time limit, harness, and leaked answer keys can move the score more than the model.
Tim Cheung5 min read