Forty pages, one page

The reading list from the research standby now has a pipeline. He keeps a queue of pending papers ▸ 30:33, and when one arrives at forty pages ▸ 30:42, the whole PDF goes into Gemini 3 for a structured draft ▸ 29:22, which Nano Banana Pro then renders as a single dense page ▸ 29:40. His standard, said straight at the authors: “conténseme, hágame un paper de máximo 20, 10 si es mejor, y ese sí me lo leo con mucho gusto” ▸ 47:26. Julia pushes back for reading itself, and he concedes the real cost: skim the summary and you never collide with the stray concept that would have changed you ▸ 46:10. Her synthesis is bibliographic: books already ship with this feature, “por eso en los libros existe la sinopsis” ▸ 46:56. The one-pager is a sinopsis generator; the deep read is still the prize for the papers that earn it.

la calificación aproximada no vale el doble de cómputo →

The demo is January’s DeepSeek R1 paper ▸ 38:47, and the one-pager surfaces exactly the ideas worth keeping. First, GRPO: standard reinforcement learning pays for two networks, an actor and a critic that only approximates a grade, so DeepSeek asked why keep the expensive approximator at all ▸ 40:02 and replaced it with scoring answers relative to the group’s average, Group Relative Policy Optimization ▸ 41:49. Second, emergence: the model started saying “wait, wait, esto podría estar mal” and correcting itself mid-chain ▸ 42:25, behavior that existed nowhere in the training data ▸ 42:41. Third, distillation, the big model teaching small open-source students ▸ 44:21. All of it under the US chip veto ▸ 45:30, on second-tier hardware that ended up performing like first ▸ 45:45. Next in the queue: Sam Altman’s Startup Playbook…