First screen, captured 10 days ago. Open the video.

The 'Astra can read and write weights to improve Qwen' post is misleading

What HN thinks

A post about a misleading post, inviting a post about how that post is misleading.

Title and abstract oversell
The Qwen result was a standard steering/directional ablation done via Python tooling, not raw weight reading; mixing it with toy raw-weight experiments implies more than was shown.
Original paper was clear enough
Anyone who read the full write-up knew exactly what was done; the rebuttal adds nothing and the hype came from readers and clickbait.
Still useful and exciting
Practitioners report real gains using LLM-guided weight editing on their own small models, and fast reverse engineering is impressive regardless of novelty.

Consensus: Most agree the raw-weight reading only happened on small networks and the Qwen fix used known interpretability techniques via code; they disagree on whether the original framing counts as misleading.

Recommended in the comments

Tracr ↗ (library): Cited as prior work showing rule-based logic can be written directly into transformer weights.

Summarised by Claude from the comment thread 10 days ago.

Points: 994
Comments: 203
Best rank: #1
Posted: 2026-11-03 by syllogism
Category: AI & ML

gpt-6-astra qwen weight-space-editing llm-evaluation interpretability research-critique

All captures:

Full page

Captured 10 days ago. Click for the whole screenshot.

Looks similar

Screenshot of naut.ca: The Hackintosh Reborn

The Hackintosh Reborn (naut.ca)

385 points by rubatuga 8 days ago | 223 comments

Blog post arguing AI coding tools like Claude Code are reviving the Hackintosh community, citing macOS ports to Snapdragon, ARM SBCs, and Nvidia GPU support.

hackintosh macos claude-code arm

Screenshot of allenpike.com: LLMs aren't "trained on the internet" anymore

LLMs aren't "trained on the internet" anymore (allenpike.com)

127 points by ingve 2 years ago | 81 comments

Screenshot of lacuna.tiptreesystems.com: Bf16 might be freezing your weight updates

Bf16 might be freezing your weight updates (lacuna.tiptreesystems.com)

9 points by martincsweiss 11 days ago | 3 comments

An article showing that training bf16 weights directly with AdamW silently drops small updates, acting as a sparse mask, demonstrated by diffing the Quiet-STaR checkpoint.

bf16 floating-point-precision adamw fine-tuning