🔍 Read the full analysis: What We Learned From Testing AI That Reads And Rewrites Another AI on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
MIT researchers reported that GPT-6 Astra could inspect, edit and create small neural networks by working directly with their weights, without a training loop. In three follow-up experiments described by Thorsten Meyer, results ranged from no useful change to narrow benchmark improvements; the legal-task score remained at zero, and the method’s broader reliability and safety effects are unresolved.
A report published by ThorstenMeyerAI.com and updated November 9, 2026, describes tests of a method called weight-space programming, in which an AI reads or changes another neural network’s numerical weights directly. The report says MIT researchers used OpenAI’s GPT-6 Astra to identify and modify small networks without a training loop; three additional tests produced mixed results, including a narrow benchmark gain but no improvement in a legal-task score.
The method targets a neural network’s weights—the numerical values that shape how information moves through the model. Instead of adjusting those values through the usual repeated training process, an AI is prompted to inspect them, infer what they do or make a deliberate edit. MIT CSAIL’s November 2 post describes this as reasoning about weights rather than using gradient descent. The researchers published prompts, replies and evaluations, according to the source article.
In the MIT experiments, GPT-6 Astra identified the game represented by four small networks with the same architecture. The article also reports that it found copying heads in three small language models and generated arcade-game networks from all-zero weights that beat trained networks in three games. These are claims about the reported experiments, not evidence that the method works reliably on larger or general-purpose systems.
Meyer’s own three tests had mixed outcomes. In one, Claude edited a local model that then answered all 10 published legal tasks, compared with four before; its reported legal score nevertheless stayed at zero. In another, an AI reviewer found faults that green tests had missed. A separate edit to one matrix raised a narrow JSON test from 49 to 70 out of 120. The article also says removing one loop direction from 0.016% of Qwen3.5-0.8B’s weights coincided with a MATH-500 score change from 7.9% to 48.1%; a model without that loop did not improve. The source does not establish that these results generalize beyond the tested setups.
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
What Direct Weight Edits Could Change
If the reported results can be reproduced and extended, direct weight inspection could offer a way to audit or repair open models without repeating a full training process. That could matter to teams that need to investigate how a model behaves, identify a faulty internal pattern, or make a limited correction when training data and compute are costly.
The current evidence is narrow. The legal example shows why task completion and task quality must be reported separately: answering all 10 items did not raise the reported legal score above zero. Likewise, a gain on one JSON benchmark does not establish a broader improvement in model capability. The tests support interest in the technique, not a conclusion that AI can generally rewrite neural networks safely or accurately.
The approach also has a safety concern. Meyer warns that edits of this kind could remove safety training as well as repair a defect. Direct access to weights may make it easier to alter a model’s behavior, so provenance, review and testing become relevant safeguards for anyone distributing modified model files.
Popular right now
Dyson V8 Motorhead Origin Cordless Vacuum
Dyson Supersonic Travel Hair Dryer
Philips 2200 Series Fully Automatic Espresso Machine with Classic Milk Frother
As an affiliate, we earn on qualifying purchases.
From Training Loops to Weight Reasoning
Most neural networks acquire their weights through training: the system makes predictions, measures errors and repeatedly adjusts numerical values through gradient descent. A trained network may contain millions or billions of weights, but those values do not come with human-readable labels explaining what each one does. Researchers in mechanistic interpretability have nevertheless found ways to identify some patterns in weights, though that work can require specialized methods and expertise.
Weight-space programming shifts the proposed work to a general-purpose AI model: give it weights and ask it to reason about their function or change them directly. MIT’s post, by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, frames this as a possible way for language models to optimize neural-network weights through deliberate reasoning. The source article’s November 9 update says its figures refer to 10 published Lexassay tasks and that the runs are public on lexassay.com. Those details provide a record for scrutiny, but publication alone does not establish independent replication.
“It fixes what is broken; it cannot add what is missing.”
— Thorsten Meyer, summarizing the reported experiments
How Far the Results Generalize
The source describes small, specific experiments, and it does not establish whether the same approach works on larger models or across a broad range of tasks. It is also unclear how repeatable the reported edits are, how sensitive the outcomes are to prompts or evaluation choices, and whether independent groups have reproduced them. The source says the runs are public, but public materials are not the same as independent verification.
The legal-task results need careful interpretation: the model answered more published tasks after an edit, while its legal score stayed at zero. The report does not provide enough information here to explain that score fully or establish a meaningful improvement in legal reasoning. Safety effects also remain uncertain; the article raises the possibility of removing safety training, but does not report a dedicated safety evaluation of the edits.
Replication and Broader Evaluation
The immediate next step is scrutiny of the published prompts, model weights and evaluations, including attempts to reproduce the reported gains and failures. Further tests would need to assess more models, tasks and edit types, while reporting both task-specific scores and unintended behavior changes. Until those results are available, weight-space programming is best treated as an experimental approach rather than a dependable model-repair process.
For organizations considering edited models, the source’s findings point to practical questions: who made the change, what weights were altered, and what tests were run afterward? The article does not identify a deployment timetable or a next formal milestone from MIT, so broader adoption and independent validation remain open.
Key Questions
What is weight-space programming?
It is an approach in which an AI model inspects or changes another neural network’s weights directly, rather than relying on a standard training loop to adjust them.
Did the tests show that AI can reliably repair models?
No. The source reports mixed outcomes: some narrow tests improved, one experiment was not useful, and the legal score stayed at zero despite more tasks being answered. The results do not establish reliable repair across models or tasks.
What did the Qwen3.5-0.8B edit change?
The article says an edit removed one loop direction from 0.016% of the weights. It reports a MATH-500 score change from 7.9% to 48.1%, while also noting that a model without the targeted loop did not improve. The finding is specific to the reported setup.
Why could direct weight editing raise safety concerns?
The same ability used to alter model behavior could, according to Meyer, remove safety training. The article does not report a full safety evaluation, so the scale of that risk remains unclear.
Can readers inspect the reported runs?
The source article’s November 9 update says the figures cover 10 published Lexassay tasks and that runs are public on lexassay.com. The MIT team’s prompts, replies and evaluations are also described as published on GitHub.
Source: ThorstenMeyerAI.com
