Back to all writing in Evaluation

Every model launch comes with a chart. The chart almost always shows the new model on top. The question this pillar takes seriously is: does the chart mean anything?

Sometimes yes. Often no. The gap between “scored well on a benchmark” and “is actually useful for what you want to do” is the single most expensive misunderstanding in applied AI today.

Read these in order

  1. Why benchmarks lie — contamination, saturation, and the incentive to teach to the test.
  2. Evals vs. vibes — why you still need to write your own evaluation set, and what a good one looks like.

Where this pillar is going

Future essays will cover LLM-as-a-judge (using one model to grade another, and where that breaks), red-teaming (adversarial evaluation of safety), and human preference data (what RLHF and DPO actually measure).

Citations

  1. Measuring Massive Multitask Language Understanding (MMLU) — Hendrycks et al., 2020 · accessed May 17, 2026 · view archived copy
  2. Beyond the Imitation Game (BIG-bench) — Srivastava et al., 2022 · accessed May 17, 2026 · view archived copy
  3. Chatbot Arena Leaderboard — LMSYS · accessed May 17, 2026 · view archived copy

Essays in this pillar

  • Why benchmarks lie

    Contamination, saturation, and the incentive to teach to the test — a field guide to reading model launch charts skeptically.

    ~2 min read

  • Evals vs. vibes

    Why every serious AI project ends up writing its own evaluation set, and what a good one looks like.

    ~3 min read