The Bar-Exam Claim, Expanded
This video presents the same text shown beside it, spoken and on screen. It adds nothing the text does not say.
Expanding "our model passed the bar exam" into its three clauses turns a headline into a list of answerable questions.
The task clause: which exam form, scored by whose rubric, at what passing line? Passing a multiple-choice section rewards answer-picking — a different skill from practicing law, so the measured task must be named before the verb "passed" means anything. The distribution clause: which questions circulated publicly before training? A model that read the exam's text holds its residue, and a lawyer's Tuesday does not resemble exam conditions — the score transfers exactly as far as the resemblance runs. The conditions clause: how many attempts, what tool access, what time limits, what human formatting help? Each vanished condition was load-bearing. Three clauses, maybe ten questions — and the claim either survives them with its meaning sharpened, or dissolves into "a system scored well on something, somehow, once."
The expansion is symmetric — it deflates the boast and it deflates the sneer, because both skip the same three clauses.
Builds on
Unlocks
- Nothing yet depends on this.