Dwarkesh Patel's crisp 12-minute video argues that AI's real bottleneck is sample efficiency, not compute — humans learn from ~200M words while frontier models train on trillions of tokens. Notes on scaling laws, distillation, and whether AI can solve its own data hunger.
Read more →
Federico UlfoSakana AI launched Fugu, an LLM that orchestrates a swappable pool of frontier models — pitched as resilience against vendor lock-in and export controls. Two variants (Fugu / Fugu Ultra) ship behind one OpenAI-compatible API.
Read more →Federico UlfoMidjourney — the AI image lab — unveiled Midjourney Medical and a whole-body "Ultrasonic CT" scanner that maps your body in ~60 seconds, with plans for a "Midjourney Spa." It's a first-gen prototype with no regulatory clearance.
Read more →Federico Ulfo