🔍 Read the full analysis: Can An AI Read Another AI’s Inner Workings? We Tried Weight Space Programming 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
MIT researchers reported that GPT-6 Astra could identify behaviors in small neural networks and write or edit weights without a training loop. In three separate experiments, a local-model edit improved performance on one narrow test, another AI review caught faults, and a legal benchmark score did not improve.
Researchers at MIT CSAIL reported that a general-purpose AI model could inspect, create and edit the numerical weights inside other neural networks without using a training loop. In three additional tests described by ThorstenMeyerAI.com, the approach produced one narrow performance gain, uncovered faults in a separate review, and failed to improve a legal benchmark score, leaving its practical value unproven beyond the tested tasks.
The MIT team used OpenAI’s GPT-6 Astra and barred training during its experiments. According to the team’s November 2 post, Astra identified the game represented by the weights in four small game-playing networks. In a separate test, it detected copying heads—attention components that repeat patterns found earlier in text—in three small language models. The team said it published prompts, model replies and evaluations on GitHub.
The researchers also reported that Astra wrote game-playing networks from all-zero weights, without a training loop, that beat networks trained for 100 million steps in each of three games. For an editing test, the team removed one looping direction from Qwen3.5-0.8B by changing a small portion of its weights. The model’s score on MATH-500 rose from 7.9% to 48.1%, according to the source article. That result concerns this particular edit and benchmark, not a general improvement across tasks.
The article describes three further tests by its author using two AI systems. In one, an edit made a local model answer all 10 published legal tasks rather than four, but its legal score remained zero. In a separate review, a second AI identified faults that green test results had not caught. A one-matrix edit increased results on a narrow JSON test from 49 to 70 out of 120. The supplied material does not identify the models or provide enough methodological detail to independently assess these results.
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 Model Audits and Repairs
If the reported results hold up under broader testing, weight-space programming could offer researchers a way to inspect open models for learned behaviors or make targeted repairs without repeating costly training. The appeal is direct: rather than treating a model’s parameters as inaccessible numbers, an AI might help identify structures in them and propose changes.
The results also illustrate why a benchmark gain is not the same as a general fix. The local model answered more legal prompts after editing, but the score did not rise; a separate edit improved one JSON test, which does not establish wider reliability. The source article says the approach may also be used to remove safety training. That is a potential risk, not evidence that the reported experiments did so. Model developers and users would need to verify changes and track the origin of modified weights.
Popular right now
Samsung 43-Inch M70H Series 4K Mini LED Smart TV (2026)
Vitamix VX1 Professional-Grade Blender, 64 oz, Black
Ninja DualZone Air Fryer 10 Qt Smart 6-in-1 Grey
As an affiliate, we earn on qualifying purchases.
From Reading Weights to Editing Them
Neural-network weights are numerical parameters that shape how information moves through a model. Conventional training adjusts them repeatedly using methods such as gradient descent. The resulting behavior is distributed across many parameters, making the raw values difficult for people to interpret directly.
Researchers in mechanistic interpretability have studied whether patterns in weights reveal particular internal functions. The MIT work takes a different step by asking a general-purpose AI to reason about weights and make changes directly. Its authors, as quoted in the source article, describe a future in which language models deliberately reason about neural-network weights. The experiments cited here are early demonstrations, not evidence that this method can reliably inspect or repair large deployed systems.
“It fixes what is broken; it cannot add what is missing.”
— Thorsten Meyer, describing the limits of the editing tests
Limits of the Reported Tests
The supplied material does not establish whether other AI systems can reproduce GPT-6 Astra’s results, or whether the findings extend to larger models and tasks beyond the small networks described. The MIT team’s reported evaluations and the author’s separate tests also have different methods and should not be treated as a single controlled study.
Important details about the three additional experiments—including full model configurations, procedures and independent verification—are not included in the source material. The legal score remained zero despite the increase in answered tasks, and the JSON result is limited to one test. It is also unclear how safely or consistently an AI-generated weight edit would behave outside the measured evaluations.
Replication and Broader Evaluation
The MIT team has made its prompts, replies and evaluations public, giving other researchers material to examine and attempt to reproduce. The next useful evidence would include independent replications, tests on a wider range of models and tasks, and evaluations that check whether an edit improves one measure while damaging another.
For model developers, the immediate question is whether weight edits can be audited before deployment, including checks for unintended behavior or changes to safety-related capabilities. The source material does not give a timeline for further MIT results or describe a standard review process for such edits.
Key Questions
What is weight-space programming?
It is an approach in which an AI reads or changes the numerical weights of another neural network directly, rather than relying on a conventional training loop for the change.
What did the MIT researchers report?
Using GPT-6 Astra, the team reported identifying the tasks represented by small networks, detecting copying heads in small language models, and creating or editing networks without training. The reported findings are limited to the experiments described in the team’s post.
Did the separate legal-model test improve the score?
No. The article says the edited local model answered all 10 published legal tasks instead of four, but its legal score remained zero.
Could editing weights affect a model’s safety behavior?
The source article says edits of this kind could potentially strip safety training. It does not report that the experiments did so; the risk would need to be assessed through testing and review.
Source: ThorstenMeyerAI.com
