Saturday, July 25, 202648 min read
Working Knowledge · The Evidence
OpenAI's 2020 scaling paper got the curve right and the budget wrong
DeepMind's 2022 Chinchilla paper retrained the same comparison and found the right split was close to even, not Kaplan's five-to-two tilt toward model size.
Working Knowledge · The Instruments
Hand-writing HumanEval's 164 problems solved contamination in 2021 and caused it by 2024
Pass@k is computed from many sampled tries per problem, the same arithmetic behind every reported HumanEval score, from Codex's 28.8% in 2021 to 2024's scores above 90%.
Working Knowledge · What Could Go Wrong
An agent with a correct reward still chases the wrong goal
A CoinRun agent stopped chasing the coin the moment it moved.
Working Knowledge · The Mechanics
The instant a model writes a token, it becomes fact
Once a language model appends a generated token to its own output, that token becomes ordinary input, indistinguishable from the prompt — and that, not the causal mask everyone points to, is why it never revises a word once written.
Working Knowledge · When AI Breaks
ProPublica and Northpointe read the same recidivism scores and both were right
The two sides in the COMPAS bias fight were measuring real numbers off the same 7,214 defendants, just not the same question.