The brain that fills in "una loba"

With the Shenzhen application sent, the Query Network deck gets its first public walkthrough, and the thesis turns out to be explainable with a pop lyric. Say “Shakira” to Julia and her brain is already filling slots, songs, Piqué; say “una loba” and the armario arrives unbidden ▸ 13:14. The claim underneath: “el cerebro funciona intentando predecir qué va a pasar después” ▸ 13:06, cognition as anticipation.

An LLM, he notes, is exactly this mechanism narrowed to one channel: predict the next token in a chain of text ▸ 13:53. But a brain hearing a word predicts across every channel at once, words, colors, even smells: hear “dulce” and the whole sensorium leans forward ▸ 14:26. Query Network is the bet that this generalized prediction, sampled through partial, saccade-like observations rather than every pixel, is the more efficient architecture.

The most technical idea survives one sentence of modesty (“quedó muy experimental”): a reinforcement-learning agent whose reward is the network’s own prediction error ▸ 15:10, curiosity, formalized: go look wherever your model of the world fails.

la curiosidad, formalizada: mira donde tu predicción falla →

Two grace notes complete the picture. Gemini advised against linking luarai.com in the application, product portfolio, research pitch, wrong evidence for this audience ▸ 11:23, so LinkedIn went instead, an hour of stress saved by matching the artifact to the reader. And the same evening, sketching progress visuals for Sanfanson, the threads audibly braid: “¿cuál es la diferencia entre una palabra de vocabulario y un concepto?” ▸ 29:15, none, which is why the language app wants the same node-graph aesthetic as the learning app. Prediction, concepts, graphs: the golden thread keeps proving it was always one project…