Intelligent Machines (Audio) IM 844: Poob Has It For You - Spiky Superintelligence vs. Generality
28 snips
Nov 6, 2025 
Jeremy Berman, a post-training researcher at Reflection.ai, dives into the evolution of AI's capabilities. He addresses the challenge of 'spiky superintelligence'—smart in specific tasks but limited overall. The discussion contrasts pre-training and post-training methods, underscoring the potential of reinforcement learning for developing generality in AI. Berman shares insights on ARC AGI benchmarks, the risk of AGI, and why open-weight models are crucial for adoption. Their outlook on how AI could learn tasks easier for humans promises a fascinating future.
AI Snips
Chapters
Books
Transcript
Episode notes
Pre- vs Post-Training Distinction
- Pre-training compresses internet knowledge into models but leaves them as document completers.
- Post-training is required to make models useful for human tasks and personalities.
Why ARC Exposes LLM Weaknesses
- ARC puzzles revealed that LLMs trained only to predict tokens fail at novel reasoning tasks.
- Solving such puzzles requires models to generate new solutions, not memorize web patterns.
Iterative Code Evolution Won ARC Briefly
- Jeremy described his ARC 2024 entry: he generated many Python programs and iteratively refined the best ones.
- His approach won briefly until reinforcement-learned models outperformed it.
