First screen, captured 3 days ago. Open the video.

What objective is reasoning solving?

What HN thinks

Half the thread debates whether reasoning is search or babble; the other half is petting the bird.

Reasoning extends computation per answer
Many argue chain-of-thought works by giving the model more forward passes and scratch space, citing the 'dot by dot' paper and restating as re-reading context.
RL search, not inference thinking, drives gains
Others back the article's view that RL finds successful paths during training and inference mostly imitates them, sometimes like a cargo cult that can wander off if it thinks too long.
Current reasoning is ad hoc
Some expect baked-in chain of thought to be replaced by latent reasoning, grounding frameworks, or direct question-to-answer mappings like JEPA.

Consensus: No firm consensus on the mechanism, though most agree reasoning traces are not human-like thought and that the site, especially the animated bird, is lovely.

Recommended in the comments

MathJax (library): Identified as the tasteful equation renderer on the article's page.
Let's Think Dot by Dot ↗ (other): Cited as evidence that even dummy tokens can improve reasoning results.

Summarised by Claude from the comment thread 2 days ago.

Points: 142
Comments: 65
Best rank: #5
Posted: 2026-11-10 by gmays
Category: AI & ML

reasoning-models reinforcement-learning chain-of-thought reward-hacking llm policy-gradient

All captures:

Full page

Captured 3 days ago. Click for the whole screenshot.

Looks similar

Screenshot of terrytao.wordpress.com: Lean Grimoire

Lean Grimoire (terrytao.wordpress.com)

51 points by smilelamp 1 day ago | 8 comments

Terence Tao blog post announcing Lean Grimoire, a coding-agent-rebuilt version of his QED logic crafting game with Lean proof display and a fantasy mode.

lean theorem-proving logic terence-tao

Screenshot of wiki.xxiivv.com: M/PC – A Concatenative OS

M/PC – A Concatenative OS (wiki.xxiivv.com)

71 points by caminanteblanco 4 months ago | 10 comments

Screenshot of andreasfragner.com: Three ways to solve problems

Three ways to solve problems (andreasfragner.com)

119 points by 42point2 10 months ago | 20 comments

Screenshot of erikgahner.github.io: Awesome repositories are no longer awesome

Awesome repositories are no longer awesome (erikgahner.github.io)

239 points by erikgahner 8 days ago | 116 comments

A blog post arguing that GitHub 'awesome' curated lists have become less useful due to outdated content and floods of AI-generated pull requests promoting vibe-coded sites.

awesome-lists github ai-slop pull-requests

Screenshot of perplexity.ai: Perplexity Decider v1.3: Training, eval, and use cases

Perplexity Decider v1.3: Training, eval, and use cases (perplexity.ai)

94 points by gmays 5 hours ago | 19 comments

Perplexity Research blog post announcing Decider v1.3 and Decider Flash v1.3 open-weight decision/classifier models, explaining their training, evaluation, and internal use.

perplexity decision-models classification model-distillation