The curriculum that waits for you
The curriculum idea had already failed, 90% against 95 without it, and he almost skipped phase four as redundant. He ran it anyway and the self-paced variant scored 97: show only the easiest 30% first, add harder examples only as they're mastered. Then he asks the obvious question about universities.
// trace: where this idea came from
- ↳ video diary @ 6:07 (el 97% del self-paced)
- ↳ video diary @ 14:31 (el 30% más fácil primero, el resto invisible)
- ↳ Entry 162-1: From 69 to 95 (la investigación que esto corona)
The original intuition, from a year ago, was a curriculum: rank all 60,000 images by prediction error and feed the easy ones first ▸ 4:09. The experiments said no: 94-95% without it, 90 with, so he ran everything else curriculum-free ▸ 4:51, and when phase four said “try different curricula” he nearly skipped it as already-refuted ▸ 5:28. The code existed, so he ran it anyway, and one variant Claude had proposed without his full understanding, self-paced, came back at 97% ▸ 6:07.
He has Antigravity explain the difference to him on camera ▸ 9:03. His old baseline was a soft preference: half the training samples random, hard images visible from epoch one ▸ 12:10, quantum physics on day one before classical, before the math ▸ 13:47. Self-paced is a hard barrier: epoch one sees only the easiest 30%, the difficult 70% completely invisible, the window expanding toward 100% only as learning progresses ▸ 14:31. The control condition seals it: hardest-first scored 80 ▸ 8:03. Julia names the human version before the AI does: universities end the semester and march on whether you learned or not; self-paced holds you on the topic until you own it ▸ 7:39.
el experimento que casi se salta por “ya refutado” dio el mejor número del año →
The discipline around the result matters as much: he caught the agent silently skipping two of five experiments and made it rerun them ▸ 11:05, and before publishing he’ll audit everything in case a flattering formula crept in, “puede que esté haciendo trampa” ▸ 11:24. Next: a report, a draft paper for Bryan the vision PhD, then public regardless ▸ 3:07. And the question the title promised, already itching: “así deberían ser las universidades”, so how would you build the curriculum that adapts to what each student actually knows ▸ 16:17…