From 69 to 95

The tool is Claude Code’s cloud sibling, discovered inside the Max plan: a sandboxed terminal wired to his GitHub, holding 1,000 credits that spend like Windsurf’s per-task sends ▸ 5:20. He pointed it at the abandoned eye-movement research repo, and the first impressive thing was a failure handled well: the Jupyter notebook blew the token limit, and where Cline “te manda a comer mierda” ▸ 11:10, the agent diagnosed, counted the file’s 3,373 lines, and read the structure instead, processing the JSON locally and sending its brain only a map, never the 35,000 tokens of training logs ▸ 15:41.

The workflow that followed is the entry’s method. Run one: distill, four Markdown files that abstracted the whole concept from his spaghetti ▸ 21:13, so faithful he concluded the AI could now out-implement him ▸ 23:57. Run two: a fresh private repo seeded with only those four files, zero inherited code by deliberate choice ▸ 38:39, and one credit later, 3,313 lines of experiment code ▸ 26:46 that ran locally, CUDA setup included, without a single error ▸ 28:02.

la especificación destilada valió más que todo el código que la precedió →

The scoreboard makes it a story. In January his hand-rolled version hit 69% on MNIST and Reddit’s top comment called it what it was, most algorithms clear 90 ▸ 32:09. The rebuilt experiment converges to about 95%, the confusion matrix finally a clean blue diagonal ▸ 33:55, first of ten planned experiments, documented toward a possible paper ▸ 34:51. He catches his own overconfidence mid-sentence, trusting the code so much he feels no need to review it, then reversing, because it would be very funny if the model had the answers sitting beside it ▸ 35:04. The generalization: in product the new bottleneck is market and sales, but in research there’s no marketing at all, so the only scarce input left is the idea, structured well ▸ 35:53. Julia gets the method as homework: five days to write PickBar’s complete spec files, vision first, carcasa sin motor if the food database isn’t found yet, and let the cloud agent pour the concrete ▸ 41:38…

Postscript, five days later: the curve keeps climbing. Re-running the experiments, the model reaches 97% accuracy, and he recites the whole arc in one breath, “pasé de 60, a 86, a 95, ahorita ya voy en 97”, within sight of state of the art on the dataset that once called his work basurita ▸ 1:56.

Postscript, two weeks on: three of four experiment phases done and the verdict lands: “está relativamente confirmado que funciona utilizar el error como recompensa” ▸ 27:02. The robustness runs add a biological wink: horizontal flips hurt the most, but blurred digits actually score better*, which he reads as evidence the eye-like mechanism is doing eye-like work, you recognize a number in the corner of your vision without looking at it* ▸ 28:07. Publication plan, chosen over formality that bores him: a mini paper on LinkedIn and Reddit, feedback from Bryan, because “lo que me interesa es que la gente sepa que esto es posible y que lo utilice” ▸ 29:56.

Postscript, the correction: the Claude-rebuilt pipeline turned out to carry a data leak, the per-step error fed to the agent as input, so the 95, the 97, and everything after belong to a contaminated series; only the original hand-built 86% and the post-fix numbers stand. The full anatomy of the discovery is entry 186-1 ▸ 14:57.