Thirty times smaller
CIFAR-100 lands at 73% after 300 epochs, and the comparison that makes the paper is size: the smallest published model he can find at that accuracy carries ~600,000 parameters; his carries 20,000. The draft has a title with 'decision transformers' in it, and a publication plan with no journal in it.
// trace: where this idea came from
- ↳ video diary @ 33:45 (600.000 parámetros contra 20.000)
- ↳ video diary @ 35:41 (el título del paper)
- ↳ Entry 183-2: The nine gets a zoom (las estrategias de mirada que ilustran el draft)
The dataset escalation continues: CIFAR-100, a hundred classes, trained for 300 long epochs, reaches 73% accuracy ▸ 33:29. The number that matters arrives when he goes hunting for context: the smallest model he can find reporting that accuracy carries around 600,000 parameters, and his eye-based architecture does it with 20,000 ▸ 33:45, a thirtyfold efficiency gap, “inmensa” ▸ 34:10. The paper’s comparison tables push it further, ten-million-parameter baselines against his 24,000 ▸ 37:11. The whole thesis of looking-instead-of-ingesting, an eye that queries the image instead of a grid that swallows it, cashes out as parameters you don’t need.
The draft has a real title now, read aloud with visible pleasure: “Achieving state-of-the-art efficiency and robustness in active vision via decision transformers and intrinsic rewards” ▸ 35:41, transformers included, by his own admission, “porque soy cool” ▸ 35:53. The figures still fight him, one gaze plot he distrusts because that run really did repeat the same movement every time, against the nine’s zoom that varies by class ▸ 36:40. The tooling note for the record: Claude Opus 4.5 writing the experiment code “casi que no saca errores” ▸ 35:06.
lo publica como lo hizo: por hobby, sin peaje editorial →
And the publication philosophy stays exactly where he planted it weeks ago: no formal review pipeline, no revíseme-aquí-revíseme-acá, “yo estoy haciendo esto por hobby, lo voto, tómalo, déjalo, revísalo y ya”, LinkedIn as the venue ▸ 37:36. The work went from Reddit’s basurita to a state-of-the-art-efficiency claim in eleven months, and it will meet the public the same way it was built: in the open, unrefereed, take it or leave it…
Postscript, the day after: the paper never shipped. The audit found a data leak feeding the per-step error to the agent, and the state-of-the-art claim died with it; post-fix, EMNIST lands at 41% against a field around 60. The unrefereed publication plan is exactly what saved him, no journal, no readers, one day between draft and discovery. The full story is entry 186-1 ▸ 31:37.
// continued in
no entry has continued this idea yet: the arc is still open