You learn the way a neural net learns
Eight months of papers distilled into one principle, and why the language game grades you instantly.
// trace: where this idea came from
- ↳ video diary @ 21:05 (por qué dejó el trabajo, la historia completa)
- ↳ video diary @ 23:21 (la conclusión del feedback rápido)
- ↳ Entry 1-4: Do you want to be remembered as the Coupa guy? (la otra mitad de por qué se fue del trabajo)
Last year, when I left my job, part of the reason was that AI was calling me. I wanted to understand it from the ground up: reinforcement learning, the DeepMind results (AlphaGo, AlphaStar, AlphaFold), the transformers behind GPT ▸ 21:05. The legacy question in entry 1-4 was one half of that decision; this hunger was the other.
So I spent seven or eight months reading papers. At the beginning I understood nothing. I read “Attention Is All You Need”, the most famous transformer paper, about three times and got nowhere ▸ 21:51. Little by little it clicked, and today I understand it all well.
The principle that survived
Here’s what all those papers reduce to. Deep learning works because feedback is instant: the model predicts, the prediction is compared against reality, and the gap becomes the adjustment. Like tensioning a rope full of little points until it holds exactly the shape you want ▸ 23:06.
And the conclusion that mattered wasn’t about machines: you learn when feedback is fast ▸ 23:21.
predice, compara, ajusta →
The principle, shipped
That’s the entire design of the language game: you translate a sentence, and the app instantly tells you which words were right and which were wrong ▸ 23:28. No lesson plans, no streaks: prediction, comparison, adjustment. Used passively, it still compounds vocabulary.
Design your practice so feedback arrives in seconds, not semesters.
Whatever you’re learning, ask where the loop is. If the gap between your attempt and knowing whether it worked is measured in weeks, the rope never gets tensioned…