🔍 Read the full analysis: Our Weight Space Programming Experiment: One AI Reads, Another Is Rewritten on ThorstenMeyerAI.com
Get audio and creator gear delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
An MIT CSAIL team reported that GPT-6 Astra could inspect and directly edit neural-network weights without a training loop. A separate set of experiments described by Thorsten Meyer produced mixed results: one edit improved a narrow test, while a legal benchmark score remained at zero. The work may aid model auditing and repair, but its scope, reproducibility and safety implications remain open questions.
The MIT team calls the method weight-space programming: an AI model receives the numerical weights that determine another network’s behavior and reasons about them directly. The team says Astra identified the purpose of four small game-playing networks from their weights and located copying heads in three small language models. The source article says the researchers published prompts, outputs and evaluations on GitHub.
The team also reported generating game-playing networks from all-zero weights, without training, that beat networks trained for 100 million steps in each of three games. In an editing test on Qwen3.5-0.8B, the article says an edit removed one direction of a looping behavior from 0.016% of the weights; the model’s MATH-500 score rose from 7.9% to 48.1%. Those are reported results from the described tests, not evidence that the technique will improve models generally.
Meyer’s own experiments had mixed outcomes. An edit by Claude led a local model to answer all 10 published legal tasks, up from four, but its legal score remained zero. A separate AI review found faults missed by green test results. An edit to one matrix increased performance on a narrow JSON test from 49 to 70 out of 120. The article does not provide enough detail in the supplied material to assess how broadly those results generalize.
Infographic 01 · What weight space programming showed
Read, write, edit: what GPT-6 Astra did with raw weights
A blog post with complete public logs, not a peer-reviewed paper. Small networks, one research group, one run per setting.
MATH-500 score of the small open model Qwen3.5-0.8B after GPT-6 Astra removed one loop direction from 0.016% of its weights. Before: 7.9%. A Hugging Face re-run of the edited model: 51.0%.
Three things an AI did with raw weights
No training loop, no gradient descent
- Read: named the game behind 4 of 4 small game networks (17,253 weights each) and found every copying head in three small language models with their first and last layers removed.
- Wrote: arcade-game networks from all-zero weights that beat networks trained for 100 million steps in all three games. Breakout: 517.4 against 25.3 for 500 million steps.
- Fixed: a car network that never backed up: 0 successes in 256 episodes before, 1,024 of 1,024 after, with 3.5% of the weights changed.
Fixes what is broken, adds nothing new
Gemma 3 270M rarely loops (8.5% of answers hit the limit). No edit moved its MATH-500 score by more than about a point.
Some things stay invisible
Honest model or yes-man? From the weights alone: not found in 4 of 4 attempts. With a view of the running model: found in 3 of 3 pairs.
Sources: MIT CSAIL blog post, 2 Nov 2026 · logs on GitHub · Hugging Face re-run.
Powered by Thorsten Meyer AI
Direct Edits Could Change Model Audits
If the approach proves reliable beyond these tests, it could give researchers a way to inspect or repair open models without repeating costly training. Direct weight analysis may also help identify internal behaviors that are hard to spot through ordinary input-and-output testing.
The same capability could create risks. Meyer warns that an edit could remove safety training as well as repair a behavior. The reported experiments do not establish that the method can reliably distinguish desirable from harmful changes, or that it is ready for use on large deployed systems. For organizations handling model weights, the immediate point is to treat their origin and subsequent modifications as matters for review.
Products Worth Considering
Sonos Arc Ultra Soundbar with Dolby Atmos, Black
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
Amazon Kindle Scribe (3rd Generation), 64GB, 11-inch Display, Graphite, with Premium Pen
As an affiliate, we earn on qualifying purchases.
From Weight Inspection to Editing
Neural-network weights are numerical values that shape how signals pass through a model. Training adjusts those values repeatedly, usually through gradient descent. Although the resulting network can perform useful tasks, its behavior is distributed across many weights rather than written as human-readable rules.
Researchers in mechanistic interpretability have studied how specific behaviors can be represented in model internals. The source article cites a 2021 Anthropic study that found some algorithms could be detected directly in weights, while noting that such work relied on specialist methods. The MIT team’s reported contribution is to test whether a general-purpose language model can reason about weights and make changes itself.
“Our neural network weights will soon be natively optimized by LLMs that deliberately reason about their weights.”
— Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, in the MIT CSAIL post as quoted by Thorsten Meyer
Results Still Need Broader Testing
The supplied account does not establish how well the method works across larger models, different architectures or tasks beyond the reported tests. It also does not provide enough methodological detail about Meyer’s local experiments to independently judge their repeatability or the reason the legal score remained at zero.
It is also unclear how robust direct edits are to different prompts and evaluation methods, whether changes persist without unwanted side effects, and how reliably the technique can preserve safety behaviors. The reported results should be read as specific experimental outcomes, not a general measure of model performance.
Published Runs Invite Replication
The MIT team has published its prompts, replies and evaluations on GitHub, and the source article says the runs are public on lexassay.com, with figures updated to cover 10 published tasks. Those materials give other researchers a basis to examine and try to reproduce the reported results.
The next useful evidence will be independent replications across additional models and tasks, including tests for unintended effects and changes to safety behavior. No further milestone or release date is specified in the source material.
Key Questions
What is weight-space programming?
It is an approach in which an AI model reads or edits another neural network’s weights directly, rather than changing them through a conventional training loop.
What did the MIT CSAIL team report?
The team reported that GPT-6 Astra identified small networks from their weights, found copying heads in small language models, generated game-playing networks without training, and edited a model in a test where its MATH-500 score rose from 7.9% to 48.1%.
Did the independent experiments improve legal-task performance?
According to Meyer’s account, a model answered all 10 published legal tasks after an edit, compared with four before it, but its legal score stayed at zero. The source does not explain that score in enough detail to interpret the result further.
Does this show AI can safely repair any model?
No. The reported tests are limited, and the source warns that edits could also remove safety training. Broader replication and checks for side effects are still needed.
Source: ThorstenMeyerAI.com
