The internet ran out of training data. What comes next? | Jacky Mok, Reka

Jacky Mok leads Applied AI at Reka, the research lab founded by ex-DeepMind, Google, and Meta researchers. His job is getting frontier multimodal research into real products. In this conversation with Tino Pereyra he explains why web data has been largely exhausted, what makes building for physical AI different from building for the cloud, how Reka pulls margin out of raw GPUs, why egocentric video is the scarcest dataset in AI right now, and what the world looks like if physical AI gets solved. Chapters 00:00 Intro: Reka, and what Applied AI actually means 01:30 Physical AI vs building for the cloud 04:00 Under the trench coat: why frontier models don't fit in robots 05:00 Infer, raw GPUs, and where inference margin lives 09:00 Native video understanding and Reka Edge 11:00 Raw weights vs full orchestration 14:30 Claru: collecting real-world data from kitchens, factories, and farms 16:00 Egocentric data, the corpus that didn't exist 19:00 Why you can't just train on YouTube 21:00 Privacy in video data collection 22:30 Why Reka bets on the full stack 28:00 If physical AI gets solved, what changes Roko's Basilisk is a newsletter supported by Athyna. Subscribe free at https://www.rokos-daily.com/