🔍 Read the full analysis: What Is Weight Space Programming? An AI That Reads And Rewrites Another AI, And What Happened When We Tried It on ThorstenMeyerAI.com
Get movie nights 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 CSAIL researchers describe using OpenAI’s GPT-6 Astra to inspect and alter neural-network weights without a training loop, including building small game-playing networks from zeroed weights. A separate report by Thorsten Meyer describes mixed follow-up tests: some edits improved narrow benchmarks, while one legal-task score remained at zero. The work is preliminary, and the reported results do not establish broad model-repair or safety gains.
MIT CSAIL researchers have reported experiments in which OpenAI’s GPT-6 Astra inspected and changed the numerical weights of other neural networks without using a training loop. The work, described as weight space programming, includes tests in which the model identified what small networks did and wrote game-playing networks from zeroed weights; a separate report by Thorsten Meyer describes additional trials with mixed results.
In the MIT experiments, Astra received network weights as text and was asked to infer what the networks did. The researchers report that it identified the games behind four small game-playing networks and found copying heads in three small language models. These are reported results from the study, not evidence that the method can reliably interpret arbitrary large models.
The team also tested writing and editing. According to the source report, Astra produced networks from all-zero weights that beat networks trained for 100 million steps in each of three games. In an edit to the open Qwen3.5-0.8B model, it removed a looping behavior by changing 0.016% of the weights; the reported MATH-500 score rose from 7.9% to 48.1%. The source says the method did not improve a model that lacked the loop in the first place.
Meyer’s follow-up tests used two different AI systems and three experiments. One edit let a local model answer all 10 published legal tasks rather than four, but its legal score remained zero. A second AI review identified faults missed by existing green tests. A separate one-matrix edit raised a narrow JSON test score from 49 to 70 of 120. These results come from the report’s own tests and should not be treated as independent confirmation of the MIT findings.
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
Potential for Targeted Model Repairs
If the approach holds up under broader testing, direct weight inspection could give researchers a new way to examine or modify open models without repeating a full training process. That could help with auditing, debugging and narrow repairs, especially when a model’s behavior appears tied to a specific pattern in its weights.
The same ability carries risks. Meyer’s report warns that weight edits could remove safety training as well as unwanted behavior. A technique that changes a model’s internal parameters may be difficult to evaluate from a simple benchmark alone, so organizations would need to track where weights came from, what edits were made and whether safety behavior changed. The reported experiments do not show that the method is ready for routine deployment.
Popular right now
Bose QuietComfort Earbuds (2nd Gen) – True Wireless Bluetooth Earbuds with Active Noise Cancellation
Vitamix 5200 Blender with 64 oz. Container, Black
DREAME A3 AWD 1000 LiDAR Robot Lawn Mower
As an affiliate, we earn on qualifying purchases.
From Training to Direct Editing
Neural networks contain numerical weights that determine how information moves through their layers. Standard training adjusts those values repeatedly using methods such as gradient descent, which nudges parameters based on errors. Weight space programming instead asks a general-purpose AI model to inspect or change the parameters directly, by reasoning over them rather than running a training loop.
Researchers have previously studied how particular behaviors can be identified within network parameters. The MIT work’s distinction, as presented in the source, is its use of an AI model to perform reading and writing tasks on weights. The MIT post was written by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab; the source says the team published prompts, replies and evaluations on GitHub. Meyer’s November 9 update says figures in his report refer to 10 published Lexassay tasks, with runs available on lexassay.com.
“It fixes what is broken; it cannot add what is missing.”
— Thorsten Meyer, summarizing a limit observed in the reported tests
Limits of the Reported Tests
The available source material does not establish how well weight space programming works on larger, widely used models, or whether the reported benchmark changes will persist across independent replications. The MIT experiments used GPT-6 Astra, and the source does not provide enough detail here to assess how sensitive the outcomes are to prompts, model access or evaluation choices.
The follow-up legal test is also limited: the model answered more tasks after an edit, but the reported legal score stayed at zero. It is not clear from the source what that score measures, why it remained zero, or whether the change improved useful legal performance. The reported JSON result applies to one narrow test, and the safety effects of edits were not quantified.
Replication and Safety Checks
The immediate next step is scrutiny of the published prompts, model changes and evaluations. The MIT team says its experiment materials are public, while Meyer’s update points readers to the 10 Lexassay tasks and their public runs. Independent researchers can test whether the findings reproduce with different models and evaluation sets.
For organizations considering direct edits, the unresolved practical work includes checking performance beyond a single benchmark, documenting parameter changes and testing whether safety behavior has shifted. The source material does not give a schedule for further MIT results or a deployment plan, so whether the approach will become a practical model-maintenance tool remains unclear.
Key Questions
What is weight space programming?
It is an approach in which an AI model reads or changes another neural network’s weights directly, rather than changing them through a standard training loop.
What did the MIT researchers report?
The MIT team reported that GPT-6 Astra identified the tasks of small networks and created game-playing networks from zeroed weights. It also tested direct edits to model weights. The findings concern the experiments described by the team and do not establish general performance across models.
Did the reported edits improve every test?
No. Meyer’s follow-up report describes a narrow JSON test improving from 49 to 70 of 120, while a legal-task model answered more tasks but retained a score of zero. The source also says an edit did not help a model that lacked the behavior it was meant to remove.
Could direct weight edits affect AI safety?
They could, according to the risk raised in Meyer’s report: an edit might remove safety training along with an unwanted behavior. The available source does not quantify that risk or establish how reliably such changes can be detected.
Are these methods ready for general use?
The reported evidence does not show that they are. The experiments are limited, and broader replication, testing on larger models and safety evaluations remain open needs.
Source: ThorstenMeyerAI.com
