🔍 Read the full analysis: A Closer Look At Weight Space Programming: We Tried AI Rewriting AI on ThorstenMeyerAI.com
Get hardware and tech essentials 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
MIT researchers report that GPT-6 Astra could identify behaviors in small neural networks from their weights and make some targeted changes without a training loop. In a separate set of experiments described by Thorsten Meyer, AI-assisted edits produced mixed results, including a narrow benchmark improvement; the methods and results still need independent scrutiny.
Researchers at MIT CSAIL report that they used OpenAI’s GPT-6 Astra to inspect and change neural-network weights directly, without a training loop, in experiments they call weight space programming. A November 6 account by Thorsten Meyer also describes three attempts to use AI to review or edit a local model: one did not help, one review identified faults, and one edit improved a limited JSON test.
In the MIT team’s account, Astra received the raw weights of small networks and was asked to infer what they did, identify patterns, and create or modify networks. The source says the team reported that it identified the game represented by each of four small game-playing networks and found copying heads in three small language models. The researchers also said it generated arcade-game networks from all-zero weights that outperformed networks trained for 100 million steps in each of three games. These are claims from the team’s reported experiments, not evidence that the approach works across models generally.
The reported editing result focused on Qwen3.5-0.8B. According to the source, an edit removed one direction associated with looping and changed 0.016% of the model’s weights; the model’s score on MATH-500 increased from 7.9% to 48.1%. The source characterizes this as a targeted fix and says the method did not improve a model that lacked the looping behavior. The results are specific to the reported test and do not establish broader gains in mathematical ability.
Meyer’s separate trials produced uneven outcomes. The account says a Claude edit helped a local model complete 12 of 15 legal answers, up from four, but its legal score remained at zero. Another AI review found faults that green tests had not caught. A one-matrix edit raised 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 independently reproduce or fully evaluate those tests.
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 reported approach proves reproducible, it could give researchers and developers a way to inspect open models or make narrow repairs without running costly additional training. It could also help reveal what a model has learned from its internal parameters, rather than relying only on its answers to test prompts. The reported experiments, however, involve particular models and tasks; they do not show that AI systems can reliably explain or repair arbitrary networks.
The same capacity could create risks. Meyer’s account warns that an edit intended to change a behavior might also remove safety-related training. Direct modification of weights could make it harder to track how a model changed, especially if the modified file is shared without clear records. Developers would need to test edited models and document their provenance rather than treating a narrow benchmark gain as proof of overall improvement.
Products Worth Considering
Apple MacBook Pro 16-inch with M1 Pro, 10-core CPU, 16-core GPU, 16GB RAM, 1TB SSD, Silver
Apple MacBook Pro MPHF3LL/A 14.2-Inch Laptop | M2 Pro 12-Core Chip 16GB Unified Memory 1TB SSD | 19-Core GPU Liquid Retina XDR Display Space
Apple MacBook Pro 16.2-inch with M5 Pro, 18-Core CPU, 20-Core GPU, 48GB Memory, 1TB SSD, Nano-Texture Display, Space Black
As an affiliate, we earn on qualifying purchases.
From Training Steps to Weight Edits
Neural-network weights are numerical parameters adjusted during training. In the conventional process described by the source, a model makes predictions, an algorithm measures errors, and gradient descent repeatedly updates the parameters. A model with hundreds of millions of weights can therefore be difficult to interpret by looking at the raw values alone.
Researchers in mechanistic interpretability have studied whether internal structures can be identified in trained networks. The MIT work, as summarized in the source, asks whether a general-purpose AI can reason over weight values directly and use that reasoning to read or alter a network. The team’s blog post was dated November 2, 2026; Meyer’s article appeared November 6. The supplied source says the MIT researchers published prompts, replies and evaluations on GitHub, but does not include a link to those materials.
Reproducibility and Safety Questions
The supplied material does not include the full MIT paper or repository details, so the experimental setup, evaluation procedures and independent replication cannot be assessed here. It is also unclear how well the method performs on larger or differently structured models, how much human guidance each task required, and whether the reported results hold across repeated runs.
The local trials are described by Meyer, but the source excerpt does not give enough information about prompts, model settings, test construction or evaluation standards to establish that the changes caused the measured differences. The potential to weaken safety behavior is presented as a risk, not as a demonstrated outcome in these experiments. Wider effects on reliability, security and model provenance remain unresolved.
Independent Tests Will Matter
The next useful step is scrutiny of the MIT team’s published prompts, outputs and evaluations, followed by independent attempts to reproduce the results. Those tests would need to examine not only the reported tasks but also failures, repeatability, and behavior after edits outside the target benchmark.
For developers considering direct weight changes, the reported results point to a need for before-and-after evaluations, records of every edit, and checks for unintended changes to safety and other capabilities. The available source does not specify a publication or deployment timeline beyond the MIT post and Meyer’s reported experiments.
Key Questions
What is weight space programming?
It is an approach in which an AI model examines or changes another neural network’s numerical weights directly, rather than relying on a conventional training loop for that task.
What did the MIT researchers report?
The source says the team used GPT-6 Astra to identify behaviors in small networks, generate game-playing networks from zero-valued weights, and edit a small language model. These findings come from the researchers’ reported experiments and have not been independently established by the material provided.
Did the reported edit improve every test?
No. The source describes a rise in the Qwen3.5-0.8B MATH-500 score and a separate narrow JSON-test improvement, but also says one local experiment did not help and the legal score remained zero.
Could editing weights affect a model’s safety behavior?
It could, according to the risk raised in Meyer’s account, because edits may alter behaviors beyond the specific target. The supplied material does not report a demonstrated safety failure from these experiments.
Can the results be independently checked?
The source says the MIT team published prompts, replies and evaluations on GitHub, but the supplied text does not provide the repository link. Independent reproduction and broader testing remain important.
Source: ThorstenMeyerAI.com
