🔍 Read the full analysis: Weight Space Programming: 3 Questions Raised By Our AI-on-AI Experiment on ThorstenMeyerAI.com
Get the latest gadgets 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
A November 2026 report describes three experiments using AI to inspect or edit another neural network’s weights without a new training loop. One edit improved a narrow JSON test, another review identified faults missed by existing tests, and a legal-task score did not improve; the results do not establish that the method works broadly.
A report from ThorstenMeyerAI.com, published November 6 and updated November 9, describes three experiments in which AI systems inspected or edited another model’s weights without using a new training loop. The tests produced mixed results: one edit improved a narrow JSON task, a second AI review found faults missed by green tests, and a legal-task score stayed at zero.
The report calls the approach weight space programming: a general-purpose AI model is given the numerical weights of another neural network and asked to interpret or change them directly. The idea differs from conventional training, in which an algorithm adjusts weights through repeated feedback and gradient descent. The report says MIT CSAIL researchers presented related work in a blog post dated November 2, 2026, and that the team published its prompts, replies and evaluations online.
In the site’s own tests, an AI-generated edit to a local model allowed it to answer all 10 published legal tasks, compared with four before the edit, but the reported legal score remained zero. A separate AI review found problems that the existing green tests had not detected. In another experiment, changing one matrix raised a narrow JSON test result from 49 to 70 out of 120. These results describe the report’s particular tests; they do not establish general improvement across models or tasks.
The updated report also summarizes MIT’s experiments using OpenAI’s GPT-6 Astra. It says Astra identified the task behind four small game-playing networks from their weights and detected copying heads in three small language models. In a separate weight-writing test, the report says networks created from all-zero weights without training beat networks trained for 100 million steps in three games. Those findings are attributed to the MIT team and its published evaluations, not independently verified here.
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 Editing Could Change
If the reported results hold up in further tests, direct inspection of weights could offer researchers another way to audit and repair open models. Instead of relying only on outputs from a test set, an AI system might identify a specific internal structure or alter a targeted part of a network. That could be useful when a model behaves incorrectly but a conventional retraining run is costly or impractical.
The report’s own outcomes also show why the approach needs careful evaluation. Answering more tasks did not raise the legal score, while the JSON improvement was confined to a specific test and model edit. A change that improves one behavior may leave broader capabilities unchanged or affect them in ways that the test does not measure. The results are evidence of possible uses, not proof of a reliable repair method.
There is a safety concern as well. The report warns that edits to weights could remove safety training, just as targeted changes might address a defect. That makes the provenance and oversight of model weights relevant to organizations that download or adapt open models. The described experiments do not show that a safety measure was actually removed; this is a potential risk identified by the report.
Products Worth Considering
Samsung Galaxy Tab A11+ 8GB RAM, 256GB Storage, Gray
Sonos Arc Ultra Soundbar with Dolby Atmos, Black
Samsung Galaxy Tab A11+ (6GB RAM, 128GB Storage) – Gray
As an affiliate, we earn on qualifying purchases.
From Training to Weight-Space Reasoning
Neural-network weights are numerical values that shape how signals move through a model. In standard training, algorithms adjust many such values in small steps based on errors. In a trained network, behavior is distributed across the weights, which can be difficult for people to interpret by simply printing the numbers.
Researchers in mechanistic interpretability have studied ways to identify internal structures in networks. The report points to a 2021 Anthropic study that found some algorithms could be detected directly from weights, while noting that such work required specialist methods. Weight space programming proposes using a general-purpose AI to do some of that reading and writing through reasoning rather than a new gradient-descent training run.
The MIT project, as described in its November 2 blog post and the report’s November 9 update, covers both interpretation and construction: inferring a network’s behavior from its weights and creating or editing weights deliberately. The updated account says the MIT task runs are public on lexassay.com and that its figures refer to 10 published Lexassay tasks. The report’s separate local experiments should not be confused with those MIT evaluations.
“Our neural network weights will soon be natively optimized by LLMs that deliberately reason about their weights.”
— MIT CSAIL researchers, as quoted in the report
Limits of the Reported Tests
The report does not establish how well weight-space reasoning generalizes beyond the specific models, tasks and evaluations described. The available material also does not provide enough detail to assess the legal benchmark’s scoring rules or explain why answering all 10 tasks left its score at zero. That result should be reported as stated rather than treated as a demonstrated legal capability improvement.
The JSON result is a rise from 49 to 70 correct out of 120 on a narrow test, but the report does not establish performance across other formats or datasets. The reported game and language-model results are attributed to the MIT team; readers would need the published prompts, evaluation procedures and runs to judge their robustness. It also remains unclear how consistently an AI could make safe, useful edits across different model architectures.
The safety concern is prospective. The report says similar edits could strip safety training, but it does not describe a test proving that outcome. Independent replication, broader evaluations and clear information about edit side effects would be needed to establish the method’s practical limits.
Replication and Broader Evaluations
The immediate next step is scrutiny of the public materials. The MIT team’s prompts, responses and evaluations are reported to be on GitHub, while the updated account points to published Lexassay runs. Researchers and readers can compare the claimed results with the underlying tasks and check whether the evaluations support the conclusions.
Further tests would need to measure more than whether a target behavior changes. They should check for regressions in unrelated tasks, reproduce results across models and distinguish a genuine capability gain from a change that only affects a narrow benchmark. Safety-focused evaluation would also need to test whether edits preserve a model’s safeguards.
For now, the report presents weight space programming as an emerging research direction, with promising demonstrations alongside limited and mixed practical results. No broad deployment milestone or timetable is established in the source material.
Key Questions
What is weight space programming?
It is an approach in which an AI model examines or changes another neural network’s weights directly, rather than relying on a new gradient-descent training loop.
What happened in the three experiments?
According to ThorstenMeyerAI.com, a legal-task edit raised the number of answered tasks from four to 10 but left the score at zero; a second AI review found faults missed by existing tests; and a matrix edit improved a narrow JSON test from 49 to 70 out of 120.
Did the tests prove that the method works generally?
No. The results cover particular models and tasks. The report describes mixed, narrow findings and does not establish reliable gains across different models or applications.
Could editing weights affect model safety?
The report identifies removal of safety training as a possible risk of weight edits. It does not say its experiments demonstrated that outcome, so the concern remains a potential risk requiring further testing.
Where can readers review the research runs?
The report says the MIT team published prompts, replies and evaluations on GitHub and that the updated Lexassay task runs are public on lexassay.com.
Source: ThorstenMeyerAI.com
