Read the full analysis: How We Tested An AI That Reads And Rewrites Another AI on ThorstenMeyerAI.com
TL;DR
A November 2026 report tested whether an AI could inspect and directly edit another model’s weights without retraining. In three experiments, one review identified faults, one edit improved a narrow JSON test, and a legal-model edit increased task coverage without raising its legal score.
A November 2026 report on weight space programming describes three experiments in which AI systems inspected or directly edited another neural network’s numerical weights, without retraining it. The results were mixed: one edit raised a narrow JSON test score, a legal model answered more published tasks but kept a score of zero, and a separate review found faults that existing tests had missed.
The report, published by ThorstenMeyerAI.com and updated November 9, says the experiments used two AI systems and involved three attempts. In one, Claude edited a local model so it answered all 10 published legal tasks, up from four; the legal score nevertheless remained zero. The author says the public Lexassay runs now cover those 10 tasks. The report does not provide enough detail in the supplied material to independently assess the task answers or scoring.
In another test, a second AI reviewed a model and identified faults that its existing green tests had not caught. A separate edit to one matrix raised performance on a narrow JSON test from 49 to 70 out of 120. The report characterizes this as a limited gain, not evidence of a general improvement across tasks.
The article also summarizes MIT CSAIL research, described in a November 2 blog post by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. According to that account, OpenAI’s GPT-6 Astra was given raw weights and asked to interpret or alter networks without gradient-descent training. The MIT experiments reported successful identification of four small game-playing networks and copying heads in three small language models, as well as newly written game networks that beat trained counterparts in the tested games. These are claims about the study’s defined experiments, not a demonstration that the method works reliably on arbitrary large models.
MIT CSAIL · 2 Nov 2026
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 Repair
Model weights are the numerical parameters that shape a neural network’s outputs. They are usually adjusted through training, rather than by a person choosing what each number should do. If an AI can reliably inspect those numbers and make targeted changes, developers could gain another way to audit, diagnose or repair open models without repeating a full training run.
The reported tests also show why the distinction between a local benchmark result and broad capability matters. More answers on a set of legal prompts did not raise the legal score, while the JSON result concerned one particular test. Such edits may help identify a specific defect, but the available results do not show that a model becomes generally more accurate or dependable after weight changes.
There is also a safety concern. The source says an edit of this kind could remove safety training as well as repair a failure. That possibility makes weight provenance and post-edit evaluation relevant to organizations that use models whose parameters they can access. The report does not establish that its experiments removed safeguards; it raises this as a risk to check.
From Training Loops to Direct Edits
Neural networks contain large collections of weights. During conventional training, an algorithm adjusts them in small steps based on errors in the model’s outputs. The resulting values can look like an unstructured list, even though the model’s behavior depends on them. Researchers in mechanistic interpretability have studied ways to identify patterns in weights, but the source describes that work as requiring specialized methods and expertise.
Weight space programming refers to asking a general-purpose AI to reason directly about those parameters: to infer what a network does, find a defect, or write a new network. In the MIT account, the model receives weights as text and is asked to analyze or alter them without a training loop. The report’s examples include a rank-one edit, a change to a matrix built around one input pattern and one output direction.
The source also summarizes MIT tests involving a small open model, Qwen3.5-0.8B. It says an edit removed one looping direction associated with 0.016% of its weights and raised its MATH-500 score from 7.9% to 48.1%. It also says no improvement occurred on a model that did not have the relevant looping behavior. These figures are attributed to the source’s account of the research; they should not be read as evidence that the same edit would help other models or tasks.
“One experiment did nothing useful, one review caught real faults, and one edit produced a narrow, measured win.”
— Thorsten Meyer, describing the three experiments
How Far the Results Extend
The supplied source material does not establish whether these methods work consistently across different architectures, larger models or real-world workloads. The three independent experiments described by the author are small in number, and the specific AI systems, prompts, full evaluation procedures and all underlying results for the author’s tests are not included in the material provided here.
It is also unclear whether the reported improvements persist across repeated runs or on evaluations not used to guide the edits. The legal result is particularly limited: the model answered all 10 listed tasks, but its score remained zero. The source does not explain that score’s rubric or show that the answers were legally correct. The claimed risks to safety training are presented as a concern, not as an outcome demonstrated in these experiments.
Independent Tests and Broader Evaluation
The next useful evidence would be reproducible evaluations showing the same edits, their effects on a wider range of tasks, and whether gains remain after independent testing. The source says the MIT team published prompts, replies and evaluations on GitHub, while the updated report points readers to public Lexassay runs for the 10 legal tasks. Those materials may allow other researchers to inspect the procedures and results.
For model developers, the immediate practical question is whether a direct edit can be validated before deployment and checked for unwanted changes, including changes to safety behavior. Until broader testing is available, the reported findings support treating weight editing as an experimental method with possible diagnostic and repair uses—not as a replacement for training or comprehensive model evaluation.
Key Questions
What is weight space programming?
It is the use of an AI system to inspect or change another neural network’s weights directly, rather than relying on a training loop to adjust them.
Did the legal-model edit improve its score?
No. The report says the model answered 10 published tasks instead of four, but its legal score remained zero. The source material does not provide enough scoring detail to judge the answers’ quality independently.
What was the reported JSON test result?
A one-matrix edit raised the result on a narrow test from 49 to 70 out of 120. That is a result on the specified test, not evidence of a broad improvement across model tasks.
Does the report show that AI can safely edit any model?
No. The reported experiments are limited, and the source does not establish reliable performance across model sizes or architectures. It also identifies the possibility that direct edits could remove safety training, a risk that requires testing rather than an outcome demonstrated here.
Source: ThorstenMeyerAI.com