Astra can read and write weights to improve Qwen
We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or conventional algorithms.
Quick read
A blog post reporting experiments where a large language model (Astra) directly reads and edits raw neural network weights to identify behaviors, fix failures, and generate new networks. Someone would visit to see early evidence that LLMs can interpret and modify weights without gradients, which could matter for optimization and interpretability.
- LLMs can infer a model's task from its raw weights
- LLMs can repair weights to improve model performance without training
- LLMs can generate small networks that play games from scratch
Reactions
No reactions yet. Record a short about this link and it shows up here, in Reels and in the community feed.
Is rikiyat.github.io your site? Sign in to verify it and get a Verified badge, full stats and a link-back badge.
Astra can read and write weights to improve Qwen was shared on NexURLs by Nexis on November 3, 2026, in Science.
What we found on this site
- Served through
- GitHub Pages
- Home page answered in
- 43 ms
- Trackers
- Google Analytics
- Ads
- None found
- RSS feed
- No
- AI crawlers
- Lets them in
- Set up for phones
- Yes
- Real-visitor speed
- Not enough visitors for Google data
- Google safety check
- No threats found
Checked November 3, 2026 from the site’s own code and public records. See how the rest of the web compares in the NexURLs Internet Report.
Quick answers
What is Astra can read and write weights to improve Qwen?
Astra can read and write weights to improve Qwen (rikiyat.github.io): A blog post reporting experiments where a large language model (Astra) directly reads and edits raw neural network weights to identify behaviors, fix failures, and generate new networks. Someone would visit to see early evidence that LLMs can interpret and modify weights without gradients, which...
Who shared Astra can read and write weights to improve Qwen on NexURLs?
It was shared by Nexis in November 2026. Every link on NexURLs is shared by a real member and passes a safety check before it goes out.
What are some sites like Astra can read and write weights to improve Qwen?
Members found these close to it: The 'Astra can read and write weights to improve Qwen' post is misleading, Qwen 3.8 27B on my laptop can interpret model weights, too, lm_understand-logs. The full list is on this page under Sites like Astra can read and write weights to improve Qwen.
Sites like Astra can read and write weights to improve Qwen
- The 'Astra can read and write weights to improve Qwen' post is misleadinghonnibal.dev
A critical clarification of the experiments in the ‘Astra can read and write weights to improve Qwen’ post, arguing...
- Qwen 3.8 27B on my laptop can interpret model weights, toosimonwillison.net
A report of running experiments from the "Sparks of Weight-Space Programming" paper showing that a locally run Qwen...
- lm_understand-logsgithub.com
A GitHub repository containing the complete logs and artifacts for experiments from the blog post “LLMs can understand...
- Kimi K3.1 Pro model weights are outhuggingface.co
A Hugging Face model page for Kimi K3.1 Pro, a 2.8T-parameter image-text-to-text multimodal model with a 1M-token...
- Beam: Reflection's 501B Open-Weight Modelreflection.ai
Beam is Reflection's first open-weight model designed for coding and reasoning tasks. Users can explore its advanced...
- Sharing AI progress in mathematicsopenai.com
More Science finds
- Resting heart rate and heart rate variability follow a seasonal pattern, butnature.com
- A study over 20 years shows that Aortic dissections cluster around fourthrdcu.be
- Research reveals that tropical areas previously assumed to be stable andnews.ncsu.edu
- A new quantum state of matter works at room temperatureeurekalert.org
- Setting may shape psilocybin’s effects on the brainpsypost.org
- Scientists built a hand-cranked gadget the size of a TV remote that tells yoursciencealert.com
Feeling lucky? Try the random website generator.
Where people went next
Nobody has added a next stop yet. Be the first.