🔍 Read the full analysis: Weight Space Programming: Inside Our Experiment With AI-to-AI Rewriting on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
MIT CSAIL researchers demonstrated experiments in which GPT-6 Astra read, wrote and edited neural-network weights without a training loop. A separate three-part trial reported one narrow benchmark improvement and a successful fault review, but its legal-task score did not improve; the results do not establish broad model repair or safety benefits.
The MIT team’s November 2 post describes tests in which Astra was given network weights and asked to reason about them directly. According to the source report, it identified the game behind four small networks from their weights and located copying heads in three small language models. The report says the team published prompts, replies and evaluations on GitHub. Those are reported experimental results, not evidence that the method works across model architectures or at production scale.
For a weight-editing test, the researchers asked Astra to modify a small open model, Qwen3.5-0.8B. The report says the edit removed one looping direction represented by 0.016% of the weights, while the model’s MATH-500 score rose from 7.9% to 48.1%. The same source stresses an important limit: the edit addressed an existing failure mode, and did not improve a model that did not exhibit that loop. The scores are specific to the reported setup and should not be read as a general measure of model capability.
The article also recounts three trials by its author using two AI systems. One edit reportedly changed performance on a local model from answering four of the 10 published legal tasks to answering all 10, but its legal score remained zero. A second AI review found faults that green tests had missed. In a third trial, an edit to one matrix raised a narrow JSON test from 49 to 70 of 120. The report provides no basis for treating these results as broad improvements; one test showed no score gain, while the JSON result covered a limited task.
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 repeatable, it could give researchers another way to inspect and modify open model weights without running a full training process. That may be useful when a model has a specific, understood defect, or when researchers want to examine whether a behavior is represented in its parameters. It could also make some forms of model testing faster, though the reported trials do not establish how well the approach scales.
The same capability has a safety dimension. The source article warns that an edit could remove safety training as well as alter an undesirable behavior. Direct access to weights is therefore not automatically beneficial: who can edit them, how edits are checked and where weights came from matter. The experiments point to a research possibility, not a ready-made safety tool or a substitute for broader evaluation.
Products Worth Considering
Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB‑C Charging, Up to 2X More Active Noise Cancelling, Transparency Mode, Adaptive
Anker SOLIX C300 DC Portable Power Station with 60W Solar Panel, 288Wh, 300W
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
As an affiliate, we earn on qualifying purchases.
From Training Loops to Weight Reasoning
Most neural networks acquire their weights through training: an algorithm repeatedly adjusts many numerical parameters to reduce errors. A model’s behavior is distributed across those values, which can make the raw weights difficult for people to interpret. Researchers in mechanistic interpretability have developed methods to study parts of networks, but the source report describes weight space programming as an effort to have a general-purpose AI read and write weights through reasoning rather than gradient descent.
The MIT post, attributed to Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, frames that goal as direct, deliberate optimization of weights by language models. The report says the experiments used GPT-6 Astra and that the researchers made their materials public. The described work therefore sits at an early experimental stage: the central question is whether an AI can infer useful structure from parameters and make controlled changes, not whether ordinary users can safely edit any model.
How Far the Results Extend
The source material does not establish whether the results have been independently replicated, whether comparable edits work on larger or differently structured models, or how reliably they hold up across broader evaluations. The reported task counts and benchmark scores apply to the experiments described, and the author’s follow-up trials are not a substitute for a controlled, independent study.
It is also unclear how researchers can reliably distinguish a targeted change from unintended effects elsewhere in a model. The article warns that weight edits could remove safety training, but it does not report a measured safety evaluation of the edits. The extent of that risk, and the safeguards needed to manage it, remain open questions.
Public Runs and Independent Checks
According to the source article’s November 9 update, figures refer to 10 published Lexassay tasks, and the runs are public on lexassay.com. That gives readers a way to examine the reported task results, though public availability alone does not establish independent replication.
The next useful evidence would include independent attempts to reproduce the MIT tests, evaluations across more models and tasks, and checks for unintended or safety-related changes after editing. Until such results are available, weight space programming is best understood as an early research approach with promising examples and clear limits, rather than a proven general method for repairing models.
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, using reasoning rather than a training loop. The reported work tests whether this can identify learned behavior or make targeted edits.
What did the MIT team report?
The source report says GPT-6 Astra identified the game behind four small networks, found copying heads in three small language models and made an edit associated with a higher MATH-500 score on Qwen3.5-0.8B. These results apply to the described experiments.
Did the follow-up trials improve every task?
No. In the reported legal-task trial, the model answered all 10 published tasks after an edit, but its legal score stayed at zero. Another narrow JSON test rose from 49 to 70 of 120, and a separate review found faults missed by green tests.
Does this prove AI can safely repair models?
No. The report describes early, task-specific experiments and does not establish reliable general repair or safety. It also warns that edits could remove safety training, while the extent of that risk remains unmeasured in the described trials.
Source: ThorstenMeyerAI.com
