The same three jokes

The provocation comes from Andrej Karpathy, godfather of the term vibe coding, who barely vibe-codes: he prefers autocomplete because the models are cuadriculadas, when he wants his own unconventional implementation, the AI keeps regressing to the way its training data always does it ▸ 19:42. His example claim: ask for a joke and you’ll get the same three jokes ▸ 23:13. So they test it live: three fresh ChatGPT chats, the zebra-is-the-oldest-animal joke three times ▸ 24:01. Gemini 3’s thinking mode is the interesting control: its visible reasoning brainstorms categories, rejects Pepito jokes, selects for accessibility ▸ 25:59, and still, in fast mode, repeats the job-interview joke verbatim ▸ 28:47. The mechanism is the seed lesson restated: probabilistic machines, same seed, same input, same answer ▸ 30:42.

His brother’s challenge gets its second round: aren’t your Suno songs creative? No, the creativity channels through your words, the model reflects it into music ▸ 22:31. The one exception he’ll grant is reinforcement learning: the DeepMind documentary where AlphaGo’s move sent Lee Sedol to the balcony for a cigarette, because in a lifetime of Go “nunca nadie había hecho esa jugada” ▸ 33:33, though even that may be brute-force search wearing creativity’s clothes ▸ 34:08.

lo que llamamos creatividad infantil es política de exploración sin colapsar →

The best idea belongs to the couple, built live: children aren’t creative by nature, they’re running the exploration policy, every environment genuinely unknown ▸ 40:53, Julia’s point being that only a baby truly acts without prior recipes ▸ 34:41. Adults switch to exploitation, he doesn’t even look at the closet anymore, he knows which shoes are there ▸ 40:20, and mastery periodically demands un-knowing, relearning technique you thought was finished ▸ 40:36. Where he breaks with Karpathy is the verdict on RL as OpenAI’s misstep: to him it’s a field still in diapers, pioneered against the current by Richard Sutton, the questioner-of-everything he recognizes himself in ▸ 43:25. And his own eye-tracking research is exactly that bet, still running…