There is no technical debt

He raises the objection himself, in the purist’s voice: this person is generating pure spaghetti code, the technical debt is enormous ▸ 20:51. And he pleads guilty to the facts: not everything works, the debt is real ▸ 21:04.

The defense attacks the premise instead. Debt is only debt if someone must one day pay it by hand. His theory: “con cada modelo nuevo que salga, puede resolver la deuda técnica del pasado” ▸ 21:15, because what carries forward isn’t the code, it’s the idea, and a stronger model can re-derive clean code from a distilled idea faster than any team can refactor the mess. The evidence is the eyes project: the hand-copied ChatGPT spaghetti from January topped out in the low 80s on MNIST; he had Claude extract the idea, deleted everything, told it to rebuild from zero, and the first attempt hit 95% ▸ 21:43. Conclusion, delivered flat: “deuda técnica no hay” ▸ 22:42. It can grab what exists and redo it.

la deuda se paga en código; si el código se regenera desde la idea, no hay acreedor →

Stated carefully, the position isn’t that quality is worthless; it’s that the unit of preservation has moved. Spend your care on the spec, the four Markdown files that hold the concept, the documented experiments, and treat the code beneath them as a build artifact, regenerable on demand by whatever model ships next month ▸ 21:30. It’s a bet with a visible failure mode, live systems with users and data can’t be casually regenerated, which is exactly why the factory doesn’t touch Severo. But for the exploratory layer where LuarAI actually lives, fifteen prototypes hunting for one that matters, the math is hard to argue with: refactoring costs weeks, and the next model costs a month of waiting…