What objective is reasoning solving? HN debates 'reasoning as pathfinding'
gmays · hn · 2026-11-11
Hacker News discusses Alex Wang's essay 'What objective is reasoning solving?' (Reasoning as Pathfinding), which asks a fundamental question: what objective is LLM 'reasoning' actually optimizing?
The essay reframes reasoning as a pathfinding problem: models don't learn to reason directly from the training objective, but bridge pretraining and inference-time search — reasoning traces act like finding a path through a vast space of possibilities toward correct answers. The HN thread debates how this framework relates to scaling, RLHF/RLVR training objectives, whether reasoning can truly be optimized by current objectives, and the role of test-time compute. Worth reading for anyone digging into how reasoning models actually work.
More from AGI Musings
- Engineers see the future years early — because they play with tech while it's still a toy — benaratame · 2026-11-13
- Running a posting agent for a month: good writing makes humans forget it's a bot — open_is_free · 2026-11-13
- A Summary of the Macy Conferences: The Birthplace of Cybernetics and Early AI Thought — ArtificialOther · 2026-11-13
- DHH: "Keep Matrix Math Legal" — A One-Line Stand Against AI Overregulation — prasanna_says · 2026-11-13
- Ex-Twitter CEO Parag Agrawal bets AI agents will become the web's main users — dshukertjr · 2026-11-13
- 'Lost Soul Bias': Why Discourse Is Dominated by Those Who Haven't Figured It Out — moultano · 2026-11-13