🔍 Read the full analysis: What Our Trial Revealed About Weight Space Programming on ThorstenMeyerAI.com
TL;DR
An account published by ThorstenMeyerAI.com on Nov. 6 describes three tests of AI systems that read or edit neural-network weights directly. Results were mixed: one review caught faults, one targeted edit improved a benchmark score, and another test showed no gain; the claims are reported by the source and are not independent verification of MIT’s findings.
The report describes weight space programming as asking a general-purpose AI model to read, edit or create another neural network’s numerical weights through reasoning, rather than adjusting them through gradient descent. It says MIT CSAIL researchers presented related work in a post dated Nov. 2, with experiments using OpenAI’s GPT-6 Astra and publishing prompts, replies and evaluations on GitHub. These details are reported by ThorstenMeyerAI.com; the source material does not provide independent verification of the results.
In the publication’s own trials, an edit to the open model Qwen3.5-0.8B removed one direction of a looping behavior and was associated with a MATH-500 score rising from 7.9% to 48.1%. The report says that change affected 0.016% of the model’s weights. It also says a Claude edit helped a local model complete 12 of 15 legal answers, up from 4, while its legal score remained zero. A separate one-matrix edit raised a narrow JSON test result from 49 to 70 out of 120.
The report characterizes its other trial as an unsuccessful edit, and says a second AI review found faults that existing green tests had missed. The source does not give the full protocol or results for each trial in the supplied material. The scores and descriptions should therefore be treated as the publisher’s account of particular experiments, not proof that direct weight editing will improve models generally.
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 method works reliably, direct inspection of weights could offer developers another way to audit or repair open models, particularly when a behavior appears linked to a specific internal pattern. The reported edit affecting a small fraction of Qwen’s weights alongside a higher MATH-500 result illustrates the possibility of a focused change, but a single reported benchmark result does not show that the edit is robust across tasks or settings.
The same capability could carry risks. ThorstenMeyerAI.com warns that an edit could also remove safety training. That is a concern raised by the article, not a demonstrated outcome in the trials described here. For organizations using open models, the report points to the importance of knowing where model weights come from and testing any modifications for both performance and safety effects.
Popular right now
Breville Bambino Espresso Machine, Brushed Stainless Steel
Hisense 40-Inch Class QD4SR Series Hi-QLED FHD Smart Roku TV (2026 Model)
De'Longhi La Specialista Arte Evo Espresso Machine with Grinder
As an affiliate, we earn on qualifying purchases.
From Training to Direct Weight Edits
Most neural networks acquire their weights through training: an algorithm repeatedly adjusts numerical values as it measures errors. The resulting parameters are spread across a model, rather than functioning as a set of human-readable instructions. The article compares them to many dials whose combined settings shape the output.
Researchers in mechanistic interpretability have studied how some model behaviors are represented in weights. The source cites a 2021 Anthropic study that detected certain algorithms directly from weights, and presents the newer proposal as an effort to have a general-purpose AI perform some of that reading and editing. The MIT CSAIL post named in the report is dated Nov. 2; the report’s own experiments were published four days later.
„Our neural network weights will soon be natively optimized by LLMs that deliberately reason about their weights.“
— Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, as quoted in the MIT CSAIL post described by ThorstenMeyerAI.com
Limits of the Reported Trials
The supplied source material does not include full experimental protocols, complete evaluation tables or enough information to independently reproduce the three publisher-run trials. It is also unclear how the edits performed on other benchmarks, with different prompts, or after repeated testing. The reported MATH-500 change is a result on that benchmark; it does not establish broad improvement in mathematical reasoning.
The source says the method could remove safety training, but does not report a test demonstrating that risk. It also does not establish whether an AI can reliably infer what weights represent across larger or differently structured models. The distinction between the MIT team’s reported experiments and ThorstenMeyerAI.com’s separate trials matters when interpreting the findings.
Replication and Safety Testing Needed
The next useful milestone is independent replication with published methods, baselines and evaluations across multiple models and tasks. That would help clarify whether the gains survive broader testing and whether unintended behavior appears after an edit.
For now, the article presents weight space programming as a research direction, not a ready-made replacement for conventional training or model evaluation. Developers considering modified weights would need to test task performance and safety before relying on them. The source material does not identify a scheduled follow-up study or deployment plan.
Key Questions
What is weight space programming?
It is an approach in which a general-purpose AI model reads or changes another neural network’s weights directly, rather than using a conventional training loop to adjust them.
What did the reported trials find?
ThorstenMeyerAI.com describes mixed results across three experiments: one edit did not help, one review found faults missed by green tests, and targeted changes improved selected measures in two tests. The results are the publisher’s account and are not independent confirmation.
Did the Qwen edit improve performance?
The report says a targeted edit affecting 0.016% of the weights coincided with a MATH-500 score increase from 7.9% to 48.1%. That is a result on one reported benchmark, not evidence of general improvement.
Does the method make models safer?
That has not been established by the material provided. The report raises the possibility that weight edits could remove safety training, but does not describe a trial demonstrating that outcome.
What remains to be tested?
Independent replication, full experimental details, results across broader benchmarks, and checks for unwanted or safety-related changes are still needed to assess how reliable the method is.
Source: ThorstenMeyerAI.com
