🔍 Read the full analysis: Could One AI Understand Another’s Weights? Our Weight Space Programming Test on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
MIT CSAIL researchers reported that GPT-6 Astra could identify behaviors in small neural networks from their weights and create or edit networks without a training loop. A separate set of trials described by ThorstenMeyerAI.com found mixed results: some reviews and narrow edits helped, while another change did not improve a legal benchmark score. The experiments do not establish that the approach generalizes to larger models or can reliably improve them.
MIT CSAIL researchers reported that OpenAI’s GPT-6 Astra could inspect and change the numerical weights of small neural networks without using a training loop. In a separate set of tests, ThorstenMeyerAI.com reported a mix of results, including a narrow benchmark improvement and an edit that did not raise a legal-task score, underscoring that the method remains experimental.
The MIT team’s November 2 post describes what it calls weight space programming: asking a general-purpose AI model to reason about another neural network’s weights directly. The researchers said Astra identified the game played by each of four small networks from their weights alone. The networks had the same architecture and 17,253 weights apiece, according to the source article.
The team also reported tests involving three small language models, each with 7.16 million weights. Astra identified copying heads—components that can reuse patterns from earlier text—in those models. For a separate writing test, the researchers said the AI produced arcade-game networks from all-zero weights, without training, that beat networks trained for 100 million steps in each of three games. These are results reported by the MIT team, not evidence that the method works across arbitrary models or tasks.
ThorstenMeyerAI.com said it ran three experiments using two different AI systems. In one, a Claude edit changed a local model’s answers on the 10 published legal tasks from four correct responses to all 10, but the legal score remained at zero. A second AI review found faults that existing green tests had missed. In another test, a one-matrix edit raised a narrow JSON benchmark result from 49 to 70 out of 120. The article’s November 9 update says the task set and public runs are available on Lexassay.com.
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
What Direct Weight Edits Could Change
If the reported results hold up in further testing, weight space programming could offer a new way to inspect or adjust open neural networks. Instead of relying only on a model’s answers or running additional training, developers might use an AI system to examine internal parameters, flag possible faults, or make a targeted change. That could matter where a model is difficult to retrain or where its behavior needs closer examination.
The current evidence is narrow. The MIT experiments described in the source concern small networks and specific tasks; the independent tests also produced mixed outcomes. A change that improves one benchmark does not show that a model is more capable overall, and the legal test illustrates why scores and task outcomes must be reported separately. The method may also carry a safety risk: the source article warns that edits of this kind could remove safety training. Any practical use would require checks for unintended effects and careful tracking of model changes.
Products Worth Considering
LG UltraGear 27GX704A-B 27-Inch QHD OLED Gaming Monitor, 240Hz
EcoFlow DELTA 3 Max Portable Power Station, 2048Wh LFP Battery
DJI Osmo Mobile 8 3-Axis Phone Gimbal Stabilizer
As an affiliate, we earn on qualifying purchases.
From Model Training to Weight Editing
Most neural networks acquire their weights through training. An algorithm adjusts numerical parameters repeatedly in response to errors, a process commonly involving gradient descent. The resulting weights are not a readable set of instructions: a model’s behavior is distributed across many values and connections.
Researchers in mechanistic interpretability have developed ways to identify some internal patterns, but the source article describes that work as requiring specialized methods and expertise. Weight space programming proposes using a general-purpose AI to perform some of that analysis and editing by reasoning directly about the parameters. The MIT post, by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, frames the idea as a possible future in which language models deliberately reason about neural-network weights. The experiments and their reported limits, rather than that longer-term ambition, are the evidence available here.
“It fixes what is broken; it cannot add what is missing.”
— Thorsten Meyer, ThorstenMeyerAI.com
How Far the Results Generalize
The available source does not establish whether the method will work on larger, widely used models, across more tasks, or when tested by independent researchers. It also does not provide enough detail here to assess how the reported benchmark gains compare with repeated runs, alternative edits, or changes to the evaluation setup. The MIT team says it published prompts, replies and evaluations on GitHub, while the separate article points to public runs on Lexassay.com; those materials would be needed to examine the results directly.
It is also unclear how reliably an AI can predict the full effects of changing weights. The source article reports a narrow JSON improvement and a legal-task change without a higher legal score, but does not establish whether either result transfers to other settings. The risk that an edit could affect safety behavior is raised by the article; the provided material does not quantify how often that might happen.
Public Runs and Further Replication
The immediate next step is scrutiny of the published prompts, outputs and evaluations cited by the researchers and the public Lexassay runs. Independent replication could test whether the results persist across model versions, datasets and repeated attempts, and whether apparent improvements come with regressions elsewhere.
For developers considering direct weight edits, the reported tests point to a need for broader evaluation before deployment: compare behavior before and after an edit, test safety-related responses, and report both task-level outcomes and benchmark scores. The source material does not announce a timetable for further MIT experiments or a planned product release, so whether the work advances beyond these demonstrations remains open.
Key Questions
What is weight space programming?
It is an approach in which an AI model examines or changes another neural network’s numerical weights directly, rather than changing them through a conventional training loop. The reported research tests this idea on small networks and specific tasks.
What did the MIT researchers report?
The team said GPT-6 Astra identified the tasks behind four small game networks from their weights, found copying heads in three small language models, and created networks that beat trained comparators in three games. These are reported experimental results; the source does not show that the approach works generally.
Did the separate tests show that weight edits improve models?
They showed mixed results. One narrow JSON test rose from 49 to 70 out of 120, while a legal-task edit changed the number of correct answers but left the legal score at zero. Another AI review reportedly caught faults missed by existing tests.
Could editing weights affect a model’s safety behavior?
The source article warns that an edit could remove safety training. The provided material does not quantify that risk or show how often it occurs, so safety effects remain an open concern requiring direct testing.
Source: ThorstenMeyerAI.com
