The Orthogonal Bet Rohit Krishnan on Training AI to Write Better
8 snips
Oct 15, 2025 
Join Rohit Krishnan, a creative force behind the Strange Loop Canon newsletter and an AI practitioner, as he dives into the world of artificial intelligence and writing. He reveals the intriguing flaws of large language models in generating prose and discusses the role of reinforcement learning in training AI like his project, Walter. Rohit also explores how AI can reshape our future work dynamics, likening it to managing playful video games. Above all, he emphasizes the importance of experimentation and play in unlocking AI's full potential.
AI Snips
Chapters
Books
Transcript
Episode notes
Great Writing Hides Trainable Patterns
- High-quality writing has deeper, subtler patterns than surface token statistics.
- Rohit proposes extracting surprisal and sampling patterns from great literature to create training gradients.
Craft Reward Proxies For Hard Tasks
- Try to design reward functions for non-verifiable human tasks like writing by finding proxy signals.
- Rohit argues improving writing is a key step toward models that reason deeper across domains.
LLMs As Fuzzy Processors
- 'Fuzzy processor' remains a useful metaphor for LLMs as probabilistic compute substrates.
- Rohit sees current systems as chips combined with tool use and multi-loop internal reasoning.
