🔍 Read the full analysis: We Let One AI Rewrite Another—Here’s What Happened on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A report from ThorstenMeyerAI.com describes MIT CSAIL experiments in which GPT-6 Astra read and edited neural-network weights without training, and three further tests by the report’s author using two AI systems. The results included a narrow benchmark improvement and useful fault detection, but one legal-task score remained at zero; the findings do not establish general-purpose model repair.
MIT CSAIL researchers tested whether a general-purpose AI could read and change neural-network weights directly, without using a training loop, and a separate report says follow-up trials produced mixed results. The experiments point to a possible way to inspect or make targeted changes to open models, but the reported gains were limited to specific tests and do not show that AI can reliably repair models in general.
In a post published November 2, the MIT team described asking OpenAI’s GPT-6 Astra to inspect model weights and infer what networks had learned. According to the team’s account, Astra identified the game behind four small game-playing networks from their weights. It also located copying heads in three small language models. The researchers published prompts, responses and evaluations on GitHub, according to ThorstenMeyerAI.com.
The team also reported writing arcade-game networks from all-zero weights, without training, that beat networks trained for 100 million steps in three games. That result is specific to the tested tasks and networks; the source does not establish that the same approach can produce capable models across broader applications.
The report’s author then describes three experiments using two AI systems. In one, an edit to the open model Qwen3.5-0.8B removed one direction associated with looping from 0.016% of its weights. Its score on the MATH-500 benchmark rose from 7.9% to 48.1%, according to the report. In another trial, an AI review found faults that green tests had missed. A separate one-matrix edit raised a narrow JSON test result from 49 to 70 out of 120. These are reported results from particular tests, not evidence of a general performance improvement.
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
A New Route to Model Inspection
If the method holds up beyond the reported experiments, direct weight inspection could give developers another way to audit or modify models, including models whose training data or internal behavior is difficult to examine through ordinary testing. A system that can identify a learned behavior from weights might help researchers locate specific mechanisms without retraining the whole network.
The report also describes a hard boundary: the edit improved a model that already had a looping behavior to remove, but did not add a missing capability in a model that did not loop. That distinction matters for companies considering weight edits as a shortcut to model improvement. The benchmark results are narrow, and the legal-task example reported in the source showed more answers but no improvement in the score.
There is also a safety concern. ThorstenMeyerAI.com says a similar kind of weight edit could remove safety training. The possibility cuts both ways: direct editing might repair a defect, but it could also weaken safeguards. The report therefore frames the method as a reason to scrutinize the provenance and subsequent modification of model weights, not as a ready-made safety solution.
Popular right now
Breville Bambino Espresso Machine, Brushed Stainless Steel
Dyson Airsmooth Styling Brush
Roborock RockMow X130H Wire-Free 4WD Robot Lawn Mower
As an affiliate, we earn on qualifying purchases.
From Training Loops to Weight Edits
Neural-network weights are the numerical values that determine how signals move through a model. In conventional training, an algorithm repeatedly adjusts those values based on how far the model’s predictions are from a target. The process, commonly called gradient descent, can involve millions of updates. The resulting weights often look like unstructured numbers even though they encode the model’s learned behavior.
Work in mechanistic interpretability has sought to identify some of that behavior inside networks. The report cites a 2021 Anthropic study finding that certain algorithms in small language models could be detected directly from weights, while noting that such analysis required specialist methods. The MIT team’s term, weight-space programming, refers to using an AI model to reason about another network’s weights and read or change them directly, rather than adjusting them through a training loop.
The MIT post, by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, presented the work as an early demonstration. ThorstenMeyerAI.com later updated its account on November 9 to say its figures referred to 10 published legal tasks and that the runs were public on lexassay.com. The supplied source does not include the full methods or results for those tasks.
“It fixes what is broken; it cannot add what is missing.”
— ThorstenMeyerAI.com
How Far the Results Extend
The available account does not establish whether the reported gains will reproduce across model sizes, tasks or independent evaluations. The benchmark changes concern specific tests, and the source does not provide enough methodological detail here to judge how robust they are or how the comparison conditions were set.
The legal-task results also require careful reading. The updated report says an edited local model answered all 10 published tasks instead of four, while its legal score remained at zero. The source does not explain the scoring in enough detail to determine why the answer count changed without a score increase. Nor does it establish whether the edits affected other capabilities, reliability or safety behavior.
It remains unclear how reliably an AI can identify the right weight change, whether human review is needed before edits are deployed, and how developers can detect harmful modifications. The supplied material reports a safety risk but does not describe a tested safeguard against edits that weaken model protections.
Reproducibility and Safety Checks
The MIT team’s prompts, replies and evaluations are reported as public on GitHub, providing material for other researchers to inspect and attempt to reproduce. ThorstenMeyerAI.com says its own runs are public on lexassay.com. Independent replication across additional models and tasks would help show whether the findings extend beyond the examples described.
For developers, the immediate follow-up is to test any proposed edit against a broad set of capabilities and safety evaluations, not just the task that prompted the change. The source does not announce a deployment plan or a next research milestone. Until more evidence is available, weight-space programming remains a research approach with promising demonstrations and substantial open questions.
Key Questions
What is weight-space programming?
It is an approach in which an AI model reads or edits another neural network’s numerical weights directly, rather than changing them through a conventional training loop.
What did the MIT researchers report?
They said GPT-6 Astra identified the task behind four small game networks, found copying heads in three small language models and wrote networks from all-zero weights that beat trained counterparts in three games.
Did the reported edits improve model performance?
Some specific tests improved. The report says a Qwen3.5-0.8B edit raised its MATH-500 score from 7.9% to 48.1%, and a separate edit improved a narrow JSON test from 49 to 70 out of 120. Those results do not establish broad improvements.
What happened in the legal-task test?
According to the report’s November 9 update, the edited local model answered all 10 published tasks, compared with four before, but its legal score remained zero. The source does not provide enough detail to explain the unchanged score.
Could direct weight editing affect safety?
The report warns that a similar edit could remove safety training. It does not describe a validated protection against such changes, so any proposed edit would need separate safety testing.
Source: ThorstenMeyerAI.com
