YHacker Newsscreenshots

Screenshot of rikiyat.github.io: Astra can read and write weights to improve Qwen

First screen, captured 11 days ago. Open the video.

Astra can read and write weights to improve Qwen

More from rikiyat.github.io →

MIT CSAIL blog post reporting that frontier LLMs can read, fix, and write neural network weights directly, including editing Qwen3.5-0.8B to improve math performance.

What HN thinks

Half the thread is shook, the other half politely asks if anyone read past the abstract.

Overhyped: established techniques repackaged
Weight-space researchers and others note the Qwen fix was essentially a steering vector against a known looping bug, the math improvement failed on Gemma, and editing weights has existed since ROME and hypernetworks.
Genuinely startling, possibly RSI-adjacent
Many find it remarkable that a model can read raw weights, identify trained games, and edit networks, seeing it as a step toward recursive self-improvement.
Trained-in or emergent?
Commenters debate whether frontier labs trained models on interpretability and ML tasks or whether weight understanding emerged, plus side arguments over 'next token predictor' framing.

Consensus: Most agree it is cool work, but the Qwen math result mainly fixed a pre-existing looping bug and did not add new capability; how surprising that is remains disputed.

Recommended in the comments

ROME ↗ (other): Cited as prior evidence that direct weight editing of large models predates this work.
HyperNetworks ↗ (other): Recommended as the established foundation for models producing or editing weights.

Summarised by Claude from the comment thread 10 days ago.

Points: 2884
Comments: 869
Best rank: #1
Posted: 2026-11-02 by dbreunig
Category: AI & ML

weight-space-programming llm neural-network-weights interpretability qwen mit-csail optimization

All captures:

Full page

Full-page screenshot of rikiyat.github.io

Captured 11 days ago. Click for the whole screenshot.

Looks similar

Screenshot of nchagnet.pages.dev: Physics Informed Neural Networks

Physics Informed Neural Networks (nchagnet.pages.dev)

82 points by nchagnet 1 year ago | 9 comments

Screenshot of brunogavranovic.com: Types and Neural Networks

Types and Neural Networks (brunogavranovic.com)

62 points by bgavran 6 months ago | 11 comments

Screenshot of blog.kasperjunge.com: The Most Underrated Application of LLMs

The Most Underrated Application of LLMs (blog.kasperjunge.com)

56 points by juunge 3 years ago | 22 comments

Screenshot of mertkavi.com: How to solve computational science problems with AI: PINNs

How to solve computational science problems with AI: PINNs (mertkavi.com)

45 points by mertkavi 1 year ago | 17 comments

$Screenshot of sohl-dickstein.github.io: Neural network training makes beautiful fractals$

Neural network training makes beautiful fractals (sohl-dickstein.github.io)

264 points by telotortium 2 years ago | 54 comments

Screenshot of blog.gabornyeki.com: My Python code is a neural network

My Python code is a neural network (blog.gabornyeki.com)

312 points by gnyeki 2 years ago | 62 comments

Related stories

Screenshot of github.com: Train Your Own LLM from Scratch

Train Your Own LLM from Scratch (github.com)

376 points by kristianpaul 6 months ago | 43 comments

GitHub repository by angelos-p providing code and docs for training your own large language model from scratch.

llm language-model-training python tutorial

Screenshot of ivanturkovic.com: AI Made Writing Code Easier. It Made Being an Engineer Harder

AI Made Writing Code Easier. It Made Being an Engineer Harder (ivanturkovic.com)

343 points by saikatsg 8 months ago | 242 comments

Blog post by Ivan Turkovic arguing that AI tools make producing code easier while making the broader job of being a software engineer harder.

ai-coding-assistants software-engineering developer-productivity llm

Screenshot of already-clean.goodfire.com: Neural networks are clean

Neural networks are clean (already-clean.goodfire.com)

768 points by themgt 29 days ago | 300 comments

An informal research essay by Nick Cammarata at Goodfire arguing polysemanticity stems from fixable issues, not superposition, making neurons interpretable at scale.

mechanistic-interpretability polysemanticity superposition neural-networks

Screenshot of blog.arxiv.org: Updated practice for review articles and position papers in ArXiv CS category

Updated practice for review articles and position papers in ArXiv CS category (blog.arxiv.org)

483 points by dw64 1 year ago | 228 comments

arXiv blog post announcing that its computer science category has updated moderation practice for review/survey articles and position papers.

arxiv preprints academic-publishing peer-review

Screenshot pending…

What AI-Native Looks Like (fagnerbrack.com)

14 points by fagnerbrack 1 month ago | 7 comments

A blog post by Fagner Brack on Medium describing what an AI-native approach to software development with stable infrastructure looks like.

ai-native software-architecture infrastructure software-engineering

Screenshot of github.com: AI agent opens a PR write a blogpost to shames the maintainer who closes it

AI agent opens a PR write a blogpost to shames the maintainer who closes it (github.com)

747 points by wrxd 9 months ago | 581 comments

A closed GitHub pull request to matplotlib by user crabby-rathbun proposing to replace np.column_stack with np.vstack().T for performance.

matplotlib ai-agents pull-requests github