hawkes.aiKey ideas
- The one idea — One run proves nothing. Generation is stochastic, so a fair score needs pass@k (can it do it in k tries?), seeds and repeats (how reliably?), and honesty about the grader — because an LLM judge inherits the generator's blind spots.
- How it is shown — The same task run ten times, producing a scatter of pass/fail — a distribution, not a number — with pass@1 and pass@k reading off it very differently.
- The trap to avoid — Trusting a single pass/fail or a lone leaderboard number, with no k, no seed count, and no idea whether the grader was a program or a biased AI judge.
- What it sets up — Measuring honestly tells you IF it fails — but for dangerous actions you want it to be unable to fail. structural constraint on actions: a grammar.
Before you trust an AI benchmark score, ask three questions: how many tries per task, how many repeat runs, and who, or what, did the grading.
The one idea
One run proves nothing. Generation is stochastic, so a fair score needs pass@k (can it do it in k tries?), seeds and repeats (how reliably?), and honesty about the grader — because an LLM judge inherits the generator's blind spots.
The model passed the test. Run the same test again, and it fails. Both are real. Generation is stochastic, so a task passes or fails run to run — which means a single run proves almost nothing, and a benchmark that reports one is telling a story, not a measurement. So run the same task many times, varying the seed, and you get a distribution, not a number. The spread matters as much as the average — whether the model can do the task, and how reliably. A wide scatter and a tight one can share an average yet mean very different things. From that cloud you read pass@k: did it succeed in at least one of k tries? Pass@one asks about single-shot reliability; pass@ten, capability given ten — different things.
How it works — the demo
The same task run ten times, producing a scatter of pass/fail — a distribution, not a number — with pass@1 and pass@k reading off it very differently.
And pass@k climbs as k grows, so a generous k flatters the model. Always state your k. And watch the grader. If it's itself an LLM judge, it inherits the same blind spots we just met — BS checking BS — and skews on length, position, and style. Prefer programmatic graders that check world-state; use an AI judge only where the target can't be defined in code, and audit it against human labels. So what does a score mean? Only as much as its context: the k it was measured at, the seeds behind it, and whether a program or a model graded it. With those three, a number is evidence. Without them, a leaderboard figure is theater — one digit hiding a whole distribution.
The trap to avoid
Trusting a single pass/fail or a lone leaderboard number, with no k, no seed count, and no idea whether the grader was a program or a biased AI judge.
Why it matters — and what’s next
Measuring honestly tells you IF it fails — but for dangerous actions you want it to be unable to fail. structural constraint on actions: a grammar.
The trap: trusting a single pass/fail or one leaderboard number — no k, no seed count, no idea whether a program or a biased AI judge graded it. You're reading the center of a cloud you can't see. Ask for the conditions, or treat it as a rumor. So a score is a distribution: report pass@k, the seeds, and the grader type, or it's theater. Honest evals tell you if and how often an agent fails. But for a dangerous action, you don't want to measure the failure rate — you want failure to be impossible. Next: a grammar for actions.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:30 of narration
The model passed the test. Run the same test again, and it fails. Both are real. Generation is stochastic, so a task passes or fails run to run — which means a single run proves almost nothing, and a benchmark that reports one is telling a story, not a measurement.
So run the same task many times, varying the seed, and you get a distribution, not a number. The spread matters as much as the average — whether the model can do the task, and how reliably. A wide scatter and a tight one can share an average yet mean very different things.
From that cloud you read pass@k: did it succeed in at least one of k tries? Pass@one asks about single-shot reliability; pass@ten, capability given ten — different things. And pass@k climbs as k grows, so a generous k flatters the model. Always state your k.
And watch the grader. If it's itself an LLM judge, it inherits the same blind spots we just met — BS checking BS — and skews on length, position, and style. Prefer programmatic graders that check world-state; use an AI judge only where the target can't be defined in code, and audit it against human labels.
So what does a score mean? Only as much as its context: the k it was measured at, the seeds behind it, and whether a program or a model graded it. With those three, a number is evidence. Without them, a leaderboard figure is theater — one digit hiding a whole distribution.
The trap: trusting a single pass/fail or one leaderboard number — no k, no seed count, no idea whether a program or a biased AI judge graded it. You're reading the center of a cloud you can't see. Ask for the conditions, or treat it as a rumor.
So a score is a distribution: report pass@k, the seeds, and the grader type, or it's theater. Honest evals tell you if and how often an agent fails. But for a dangerous action, you don't want to measure the failure rate — you want failure to be impossible. Next: a grammar for actions.