The Product Manager Are We Overpromising and Under-delivering on AI?
Nov 18, 2025
Dhruv Batra, AI researcher and entrepreneur and co-founder & Chief Scientist at Yutori, brings 20 years in deep learning, computer vision, and embodied AI. He unpacks AI’s “jagged” abilities and why human expectations misalign with machines. They cover the challenges of browser automation, UX patterns machines miss, building trust without risky permissions, and what remains unsolved in vision and agents.
AI Snips
Chapters
Transcript
Episode notes
AI Capabilities Are Jagged Not Uniform
- AI capabilities are jagged rather than uniformly improving across tasks.
- Dhruv Batra illustrates this with models that solve Olympiad math yet make basic numeric errors like saying 9.11 is greater than 9.9.
XKCD Example Shows Sharp Task Jumps
- Dhruv retells an XKCD-style example showing sharp jumps from easy to hard problems.
- The PM asks if a photo is in a national park (easy) then asks if it's off a board (requires $50M research), illustrating jaggedness.
Web Automation Is Sequential And Fragile
- Browser automation is hard because websites are designed for humans and actions form long sequential decision chains.
- Mistakes cascade and some actions (like submitting forms) are irrecoverable, requiring simulation training.