Odd Lots Jack Morris on Finding the Next Big AI Breakthrough
184 snips
Sep 26, 2025 
Jack Morris, an AI researcher and Ph.D. candidate at Cornell, delves into the unpredictable nature of AI breakthroughs. He discusses the intricate balance between open and closed research labs, revealing how proprietary data can set teams apart. Jack also explores the complexities of measuring model performance with benchmarks and the distinction between supervised and reinforcement learning. With thoughts on the future of personalization and the challenges of monetizing curiosity-driven research, he offers a fresh perspective on the evolving AI landscape.
AI Snips
Chapters
Transcript
Episode notes
ELO Ratings Capture Qualitative Gaps
- Human pairwise comparisons (ELO-style) capture qualitative differences models' metrics miss.
- ELO ladders help rank models on style and subjective quality beyond raw benchmarks.
Reinforcement Learning Was A Turning Point
- Reinforcement learning from human feedback unlocked major gains like improved math and reasoning.
- This training shift is a key recent scientific breakthrough for language models.
Progress Is Lumpy And Unpredictable
- Progress in AI is lumpy and unpredictable across capabilities like agents versus math.
- Breakthroughs often appear in unexpected subdomains, not the hyped ones.