Sanjeev Arora on Parrots No More: How AI Models Learn, Reason, and Self-Improve

Video & Timestamps

On Thursday, May 22, Sanjeev Arora joined Markus’ Academy for a conversation on Parrots No More: How AI Models Learn, Reason, and Self-Improve. Arora is the Charles C. Fitzmorris Professor of Computer Science and the Director of Princeton Language and Intelligence at Princeton University.

Watch the full presentation below. You can watch all Markus’ Academy webinars on the Princeton BCF YouTube channel.

Timestamps:

[0:00] Markus’ introduction

[4:15] Next-word prediction is more powerful than it seems

[18:02] Skills and implications for originality

[41:27] Metacognition: are LLMs aware of how they are solving tasks?

[46:07] Current AI techniques

[1:03:34] Peeking 5-7 years ahead

Summary

• A summary in four bullets
  • LLMs don’t just memorize their training data—they learn to combine abstract skills in new ways, showing signs of metacognition (“thinking about thinking”)
  • Synthetic data—generated by another LLM—improves the quality of a model’s training data. Human data need not be the gold standard!
  • Post-training (fine-tuning on curated Q&A or reinforcement learning) teaches models how to access and apply their internal knowledge more effectively
  • Building on these ideas, the talk covered six current AI techniques: (1) extrapolation of scaling laws for training, (2) multimodal models, (3) models helping improve future models, (4) distillation, (5) self-improvement loops, and (6) AI agents.
• Click here for the full summary