🔍 Read the full analysis: Can An AI Read And Rewrite Another AI? Exploring Weight Space Programming 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 and write or edit weights without a training loop. Separate experiments described by ThorstenMeyerAI.com found mixed results: one narrow benchmark improved, while a legal-task score did not. The findings suggest a possible approach to model inspection and repair, but do not establish reliable general-purpose editing.
Researchers at MIT CSAIL reported that OpenAI’s GPT-6 Astra could inspect, create and edit neural-network weights without using a training loop, a method they call weight space programming. In experiments reported separately by ThorstenMeyerAI.com, attempts to apply similar edits produced mixed outcomes, including a narrow benchmark gain and no improvement on a legal-task score.
The MIT team supplied Astra with the numerical weights of small neural networks and asked it to infer what they did. According to the team’s November 2 blog post, the model identified the game played by each of four small game networks, whose architectures were identical. The researchers also reported that it found copying heads—components that repeat patterns from earlier text—in three small language models.
The team also tested whether the AI could write networks rather than just inspect them. It reported that Astra created game-playing networks from all-zero weights, without training, that beat networks trained for 100 million steps in each of three games. The source material does not provide the full experimental details for those results, so they should be understood as the researchers’ reported findings rather than evidence that the method works broadly.
In a separate set of experiments, ThorstenMeyerAI.com said Claude edited a local Qwen3.5-0.8B model by removing one direction of a looping behavior, changing 0.016% of its weights. The model’s MATH-500 score rose from 7.9% to 48.1%, according to the report. However, a legal score remained at zero despite the model answering all 10 published legal tasks after the edit. Another one-matrix edit raised a specific JSON test result from 49 to 70 correct out of 120; a separate AI review also identified faults missed by existing 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
A New Route to Model Inspection
If the results hold up, weight space programming could offer model developers another way to inspect open models: ask an AI to reason about the parameters themselves, rather than only checking outputs from a series of prompts. That could help identify some internal behaviors or correct a narrow failure without repeating costly training. The reported results, however, do not show that this can reliably repair arbitrary models or improve their overall performance.
The same capacity could carry safety risks. ThorstenMeyerAI.com warned that edits of this kind might also remove safety training. If model weights can be changed directly, developers and users may need stronger ways to check what an edit changed, where it came from and whether it weakened safeguards. The report does not establish that safety protections were removed in these experiments.
Products Worth Considering
Trezor Safe 5 Crypto Hardware Wallet with Color Touchscreen
LG UltraGear 32GX870A-B 32-inch 4K OLED Gaming Monitor
Logitech MX Ergo S Advanced Wireless Trackball Mouse USB-C – Graphite
As an affiliate, we earn on qualifying purchases.
From Training to Direct Edits
Neural networks store learned behavior in large collections of numerical parameters called weights. During conventional training, an algorithm repeatedly adjusts those numbers based on errors, a process commonly associated with gradient descent. The logic of a trained network is distributed among its weights, which are generally not readable as ordinary instructions.
Researchers in mechanistic interpretability have studied how particular behaviors relate to parts of a network. The MIT work asks whether a general-purpose AI can perform some of that analysis and make changes directly. The team’s post, by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, describes an ambition for language models to reason deliberately about weights. The source says the MIT experiments used GPT-6 Astra and that prompts, replies and evaluations were published on GitHub.
The independent report describes three small-scale tests, not a broad evaluation across production systems. Its updated figures refer to 10 published Lexassay tasks, with runs made public on lexassay.com. Results on small models and selected benchmarks do not by themselves establish performance on larger models or real-world deployments.
“Our neural network weights will soon be natively optimized by LLMs that deliberately reason about their weights.”
— MIT CSAIL researchers, in their November 2, 2026, blog post
Limits of the Reported Tests
The available material does not establish how well the method performs on larger or more varied models, or whether independent researchers can reproduce the results. The MIT findings and the separate experiments also use different tests, so their outcomes should not be treated as a single head-to-head evaluation.
It is also unclear how consistently direct edits preserve other capabilities and safety behaviors. The source reports improvements on particular tests, but does not provide enough information here to determine the broader effects of each edit, the reasons for every result or how performance changes across additional benchmarks.
Reproduction and Broader Evaluation
The next useful step is independent replication of the published MIT experiments and the separate Lexassay runs, with clear reporting of prompts, evaluation methods and full results. Further tests would need to establish whether edits that help one task preserve performance on others, including safety checks.
For now, weight space programming remains a reported research approach with promising but limited demonstrations. The experiments do not show that developers can routinely rewrite neural networks by instruction alone, and the risks of unintended changes remain unresolved.
Key Questions
What is weight space programming?
It is an approach in which an AI model reads or changes another neural network’s numerical weights directly, rather than relying on a conventional training loop.
What did the MIT researchers report?
The team said GPT-6 Astra identified behaviors in small networks, found copying heads in three small language models and wrote game-playing networks from zeroed weights. These are the researchers’ reported results, not proof of broad performance.
Did the separate tests show consistent improvement?
No. ThorstenMeyerAI.com reported a large increase on one MATH-500 test and a gain on a narrow JSON test, but the legal score remained at zero. Another experiment did not produce a useful result.
Could direct weight edits affect safety?
Potentially, according to the source report, which warned that an edit could strip safety training. The material does not say that this happened in the reported tests; broader safety effects remain unclear.
Source: ThorstenMeyerAI.com
