Super Data Science: ML & AI Podcast with Jon Krohn · Jon Krohn

972: In Case You Missed It in February 2026

·27 min·1 clip

Nova Act targets agents that are “60%” reliable today, which Amazon says is “0% useful.”

1. Super Data Science: ML & AI Podcast episode 972 is an “In Case You Missed It” roundup from February 2026. 2. Jon Krohn hosts the episode and frames it as a monthly highlights reel from prior conversations. 3. The episode asks which clips from recent interviews best show AI product building, human intelligence, and automation. 4. Will Falcon, co-founder and CEO of Lightning AI, explains how PyTorch Lightning grew from open source into a startup with over $500 million in annually recurring revenue. 5. Falcon says he did not want to start a company and preferred research, reading, and “thinking about things.” 6. He describes working on what he calls “pre-training world models” before that phrase had a name. 7. Falcon says VCs started emailing after PyTorch Lightning took off, and he flew out on a Thursday and had 10 term sheets by Monday. 8. He says he almost shut down PyTorch Lightning in September because people kept dragging him back into coding and support work. 9. Tom Griffiths, professor of computer science and psychology at Princeton and author of The Laws of Thought, talks about human intelligence as rational adaptation under tight constraints. 10. Griffiths ties human learning to limited decades of life, a couple pounds of neurons, and communication through “weird honking noises” or keyboard typing. 11. He says humans need small-data learning, efficient search, and conventions such as writing and starting companies to pool resources. 12. Griffiths contrasts that with AI systems that can add compute, transfer state, and share training data much more directly. 13. He says AI and human intelligence may become different kinds of intelligence because they solve different computational problems. 14. He lays out two paths: keep engineering the best systems possible, or try to build human-like attributes into AI. 15. He names meta learning as one way to add inductive biases that are more human-like. 16. He says scaling is easier than engineering inductive bias, which is why the field is in its current moment. 17. Jon Krohn’s interview style is conversational and highly guided, with detailed follow-ups and occasional summaries of the guest’s point. 18. The episode’s tone is technical and practical, with topics ranging from fundraising and research to agent reliability and routing optimization. 19. Listeners interested in ML systems, AI agents, and research-to-startup stories would likely get value from this episode. 20. Listeners seeking a single deep-dive topic rather than a highlights package may skip it.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Super Data Science: ML & AI Podcast with Jon Krohn