🔍 Read the full analysis: We Tested An AI That Rewrites AI—Here’s How Weight Space Programming Fits In on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
An MIT CSAIL team reports that GPT-6 Astra could identify functions in small neural networks and create or edit weights without a training loop. A separate set of experiments described by ThorstenMeyerAI.com found mixed results, including a narrow benchmark improvement and tests that did not improve a legal score. The work is early, and the reported results do not establish broad performance or safety.
MIT CSAIL researchers say they used OpenAI’s GPT-6 Astra to inspect, generate and edit neural-network weights without running a training loop, an approach they call weight space programming. In a separate report published November 6, ThorstenMeyerAI.com describes three trials with two AI systems: one produced a limited improvement, one review identified faults missed by tests, and one edit did not improve a legal benchmark score.
In the MIT team’s reported experiments, Astra was given the numerical weights of small networks and asked to infer their function or modify them. The source says the researchers tested four small game-playing networks with the same architecture and 17,253 weights each; Astra identified the games from the weights. The account also says it found copying heads in three small language models. These are results reported by the team, not evidence that the method works reliably on large deployed systems.
The MIT post also describes generating game-playing networks from zero-valued weights without training, and editing a small open model. According to the source, removing one loop direction from 0.016% of Qwen3.5-0.8B’s weights raised its MATH-500 score from 7.9% to 48.1%. The figures are tied to that model, intervention and benchmark; they should not be read as a general improvement across tasks.
ThorstenMeyerAI.com reports separate local tests. One edit let its model produce 12 of 15 legal answers, up from four, but the legal score remained zero. Another AI review found faults that existing green tests had not caught. A one-matrix edit lifted a narrow JSON test from 49 to 70 of 120. The article says a further experiment did not help when the model lacked the relevant looping behavior. The supplied source does not provide enough detail to independently assess the setup or reproduce these trials.
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
Most neural-network development changes weights through repeated training, rather than having a general-purpose model inspect and deliberately alter them. If the reported approach proves reproducible, it could offer researchers another way to audit open models, identify specific learned behaviors and make targeted repairs without retraining from scratch.
Direct editing also creates a safety concern. The source warns that a similar intervention could remove safety training as well as unwanted behavior. A performance gain on one benchmark does not establish that other capabilities, refusals or safeguards remain intact. Any practical use would need testing across tasks, clear records of changes, and checks for unintended effects.
Products Worth Considering
LG UltraGear 34GX90SA-W 34-inch WQHD Curved OLED Gaming Monitor
Samsung Galaxy Tab A11 8.7-Inch Wi-Fi Tablet, 64GB, 4GB RAM, Silver
Kindle Paperwhite Signature Edition 32GB (2024) – Auto-adjusting front light, wireless charging, weeks of battery life – Metallic Black
As an affiliate, we earn on qualifying purchases.
From Training Steps to Weight Edits
Neural networks contain numerical weights that determine how information moves through their layers. During conventional training, an algorithm adjusts those values through many small updates, commonly using gradient descent. The resulting numbers are difficult to interpret individually, even when the model’s overall behavior can be measured.
Researchers in mechanistic interpretability have studied ways to identify patterns and algorithms within trained networks. The MIT post, credited to Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, presents weight space programming as a step toward asking a language model to reason about those values directly. Its stated ambition is that neural-network weights may be deliberately optimized by language models. The reported work is an early research demonstration, not a replacement for established training or evaluation practices.
““One experiment did nothing useful, one review caught real faults, and one edit produced a narrow, measured win.””
— Thorsten Meyer, ThorstenMeyerAI.com
Reproducibility and Safety Remain Open
The available source does not establish how well weight space programming performs beyond the specific small networks, model and tests described. It is also unclear how consistently another AI system would reproduce the results, how sensitive the edits are to prompts or model versions, and whether the benchmark gains persist under independent evaluation.
The report gives only a high-level account of the site’s local experiments. It does not supply full experimental protocols, model configurations or independent verification in the provided material. The safety consequences of successful edits are also unresolved: the source raises the possibility of removing safety training, but does not document a test demonstrating that outcome.
Independent Tests Will Set the Scope
The next useful step is independent replication of the MIT experiments and the reported local trials, with complete methods, unchanged baselines and evaluations across more than one task. Researchers and model operators would also need to check whether an edit changes safety behavior or degrades abilities outside the target benchmark.
The source says the MIT team published prompts, replies and evaluations on GitHub, offering material for scrutiny. No further milestone or deployment date is specified in the supplied account. Until broader testing is available, weight space programming is best treated as an experimental research technique rather than a validated way to repair production models.
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 for each change.
What did the MIT researchers report?
The MIT CSAIL post describes GPT-6 Astra identifying the functions of small networks, generating game-playing networks from zero-valued weights and editing a small language model. Those results are reported experiments, not proof of general performance.
Did the independent tests show a broad improvement?
No. ThorstenMeyerAI.com described mixed outcomes. One edit improved a narrow JSON test and another increased the number of completed legal answers, but the legal score stayed at zero.
Could direct edits affect safety?
The source says they could potentially remove safety training, but does not report a demonstrated safety-removal test. Whether edits preserve safeguards needs separate evaluation.
Is weight space programming ready for routine use?
The supplied reports do not establish that. The work is early, and its repeatability, effects across tasks and safety implications remain unclear.
Source: ThorstenMeyerAI.com
