The curiosity algorithm

The two learning engines from the pool map straight onto his research, and the surprise is how much is already done. Context first: teaching a network to move its eyes is a reinforcement learning problem, an agent whose environment is the image and whose actions are gaze movements, up, left, right, zoom ▸ 16:49. RL’s classic weakness is the reward: games ship with scores built in, “pero en la vida no hay recompensas” ▸ 17:37, and hand-crafted rewards only carry you so far.

His answer came from watching children: make curiosity the reward, formalized as the inverse of the prediction error ▸ 19:23. If the agent predicts the next glimpse well, it already knows this region, low curiosity, move on; where prediction fails, novelty lives, look harder ▸ 20:06. It’s the child with the traffic light (once the sequence is learned, boredom), the puppy discovering the mirror ▸ 21:00, the toddler abandoning the understood teddy bear. He built it, got “resultados aceptables”, filed it as not-that-novel, and delivers the re-appraisal with a grin: “hoy en día lo vuelvo a ver y es como que: papá, soy recrack” ▸ 19:42.

curiosidad = 1/error de predicción; la disonancia aún no tiene fórmula →

Which leaves the second engine unbuilt and named: dissonance, the signal that fires not where prediction is merely uncertain but where confident predictions get contradicted, the swimming-breath moment rendered in loss functions. “Ya creo que tengo una idea de cómo hacerlo, pero después lo haré” ▸ 22:16, after the vaquita de leche, per the standing sequence. The archive notes the claim for posterity: curiosity plus dissonance, he believes, “ya son suficientes para finalmente darle ojos a la IA”…