🔍 Read the full analysis: How Does Weight Space Programming Work? Our AI Experiment, Explained on ThorstenMeyerAI.com
Get smart everyday buys 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
A ThorstenMeyerAI.com report tested AI systems that inspect and edit neural-network weights without a training loop. Its experiments found one measured improvement on a narrow JSON test and a change that expanded legal-task coverage but did not improve the score; the authors say a separate edit raised MATH-500 performance in Qwen3.5-0.8B. The method’s limits and its effects on safety remain open questions.
ThorstenMeyerAI.com reported on November 6 that it had tested AI systems that inspect and edit neural-network weights directly, without retraining the networks. The three experiments produced mixed results: one edit improved performance on a narrow JSON test, another expanded the number of legal tasks a model answered but left its legal score unchanged, and a separate review identified faults that existing tests had missed. The findings point to a possible way to audit or repair open models, but do not establish that the technique works broadly.
The report calls the approach weight-space programming: a general-purpose AI model examines the numerical weights inside another neural network and reasons about what to change. The MIT CSAIL researchers cited in the report describe it as a way to work with weights without gradient descent, the iterative process typically used to train networks. Their November 2 blog post outlined experiments in which OpenAI’s GPT-6 Astra read, wrote and edited networks. The researchers published prompts, replies and evaluations, according to the report.
In the publication’s own experiments, the authors say Claude edited a local model so it answered all 10 published Lexassay legal tasks, up from four. However, the legal score remained at zero. In another test, a change to one matrix raised the result on a specific JSON task from 49 to 70 out of 120. The article also says a second AI reviewer found faults that green test results had not revealed. These are reported results for the described tests, not evidence of general improvement across legal or software tasks.
The report separately summarizes MIT’s demonstrations. It says GPT-6 Astra identified the game behind four small game-playing networks from their weights and located copying heads in three small language models. For a math test, it reports that an edit to Qwen3.5-0.8B changed 0.016% of the weights and raised its MATH-500 score from 7.9% to 48.1%. The article says the edit removed one loop direction. The provided source does not include the full methods or results tables, so these figures should be understood as claims reported by the publication.
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
If the reported results can be reproduced and extended, direct inspection of weights could give developers another way to find faults or modify open models without running a full training process. That may matter where teams need to understand why a model behaves in a particular way, or where they want to test a targeted repair. The reported legal-task result also illustrates why coverage and quality are different measures: answering more prompts did not raise the score.
The same access creates a safety concern. The source warns that edits could remove safety training as well as alter unwanted behavior. A system that can modify weights may be useful to model developers, but the ability to make changes does not itself show that they are safe, reliable or easy to reverse. The report’s small and task-specific experiments do not establish how the technique behaves on large deployed models.
Products Worth Considering
Samsung 32-inch Odyssey OLED G8 (G80SH) 4K Gaming Monitor
Dell UltraSharp U3225QE 31.5-Inch 4K UHD 120Hz IPS Black Thunderbolt Hub Monitor
Samsung 34-inch ViewFinity S50GC Ultra-WQHD Monitor
As an affiliate, we earn on qualifying purchases.
From Training to Direct Edits
Neural-network weights are numerical values that shape how information moves through a model. During conventional training, an algorithm adjusts those values repeatedly in response to errors. The report compares them to many linked dials: the behavior of a network is distributed across its weights, rather than set by a person assigning a clear purpose to each one.
Researchers in mechanistic interpretability have studied how some internal patterns can be identified in network weights. The article refers to 2021 research from Anthropic that detected certain algorithms directly from weights, while noting that such work has typically required specialist tools and expertise. The MIT work described in the report asks whether a general-purpose AI can perform some of that analysis and editing through reasoning instead. The article’s November 9 update says the figures refer to 10 published Lexassay tasks and that the runs are public on lexassay.com.
“it fixes what is broken; it cannot add what is missing.”
— ThorstenMeyerAI.com, summarizing the reported limitation
Limits of the Reported Tests
The source describes a small set of experiments, but the material provided here does not include complete methods, statistical analyses, or independent replications. It is not clear whether the reported gains would persist across different prompts, model versions, evaluation sets or larger networks. The legal experiment’s score remained zero despite more tasks being answered, leaving the practical quality of those answers unresolved.
The report also says the technique did not help when the relevant behavior was missing rather than broken. That suggests the method may depend on an existing capability or pattern in a model, but the tests described do not establish a general rule. The implications for safety training and misuse are also not quantified; the source flags the possibility of removing safety-related behavior but does not report a systematic safety evaluation.
Replication and Safety Checks
The experiments and evaluations cited in the report are described as public, giving other researchers a basis for checking the results. The next meaningful test is whether independent teams can reproduce the findings and determine how reliably direct edits work across models and tasks. Further reporting would also need to show whether changes persist outside the original evaluations and whether they introduce new errors.
For organizations handling open model weights, the source raises a practical question about where weights come from and how edits are tracked. The report does not announce a deployment or a scheduled next milestone. Until broader tests are available, its results remain an early, task-specific demonstration rather than evidence that weight-space programming is a dependable substitute for training or conventional evaluation.
Key Questions
What is weight-space programming?
It is an approach in which an AI model reads or changes another network’s numerical weights directly, rather than relying on a training loop to make the changes.
What did the report’s own legal test find?
The report says an AI edit made a local model answer all 10 published Lexassay legal tasks, compared with four before. Its legal score remained at zero, so the greater task coverage did not amount to a higher score.
Did direct weight editing improve every test?
No. The report describes mixed outcomes: a narrow JSON score rose from 49 to 70 out of 120, while the legal score stayed at zero. It also says an edit did not help when the needed behavior was absent.
Could weight-space programming affect model safety?
The source warns that edits could strip safety training from a model. It does not provide a systematic measurement of that risk, so the extent and reliability of such effects remain unclear.
Are the reported results independently verified?
The source says the MIT team’s prompts, replies and evaluations were published, and that the Lexassay runs are public. The supplied material does not report independent replication of the results.
Source: ThorstenMeyerAI.com
