🔍 Read the full analysis: Weight Space Programming: How An AI Can Work With Another AI’s Weights on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
MIT researchers reported that GPT-6 Astra could inspect and alter other neural networks’ weights without a training loop. Source-reported tests include a narrow benchmark improvement, but other experiments showed no legal-score gain and no benefit when the relevant behavior was absent. The results are early and do not establish a general-purpose way to repair models.
Researchers at MIT CSAIL reported that GPT-6 Astra could inspect and change another neural network’s numerical weights without using a conventional training loop. The method, called weight space programming, was also tested in experiments reported by ThorstenMeyerAI.com, which found mixed outcomes: one narrow benchmark improved, while a legal-task score did not.
Neural-network weights are the numerical parameters that shape how a model processes inputs. In ordinary training, an algorithm adjusts those parameters repeatedly using examples and error measurements. Weight space programming instead gives a general-purpose AI model access to another network’s weights and asks it to interpret or modify them through reasoning, according to the MIT CSAIL post by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab.
The MIT team said GPT-6 Astra identified the tasks performed by four small game-playing networks from their weights alone. It also reported that Astra could identify copying attention heads in three small language models. In separate tests, the team said Astra created game-playing networks from zeroed weights without training, and those networks beat models trained for 100 million steps in each of three games. The post’s experiments and evaluations were made public, according to the source article.
The reported editing results were narrower. In Qwen3.5-0.8B, Astra removed one direction of a looping behavior by changing 0.016% of the weights; the source reports that the model’s MATH-500 score rose from 7.9% to 48.1%. ThorstenMeyerAI.com also describes its own three experiments: a Claude edit made a local model answer all 10 published legal tasks rather than four, but its legal score remained zero; a second AI identified faults missed by existing tests; and a one-matrix change raised a JSON test result from 49 to 70 out of 120. These are source-reported results, not evidence that the method works reliably across models or tasks.
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 Gains, Narrow Evidence
If the approach proves repeatable, direct weight inspection could give developers another way to audit open models, locate unwanted behavior or make targeted repairs. That could matter where retraining is costly or where teams need to understand a model they did not create. The reported tests suggest that some behaviors may be represented in ways an AI can identify and alter.
The results also show why claims of broad model repair would be premature. In the source article’s legal experiment, expanding the number of tasks the model answered did not raise its score. A reported edit also helped only when a relevant behavior was already present in the network; it did not add missing capability. The method may be better understood, for now, as a possible tool for narrow diagnosis and editing than as a replacement for training or evaluation.
There is a safety concern as well. The same ability to modify weights could potentially remove safety-related behavior, the source article warns. That possibility is not a reported demonstration of a specific harmful edit, but it makes weight provenance and access controls relevant questions for organizations using modified or downloaded models.
Popular right now
KitchenAid Classic Series 4.5-Quart Tilt-Head Stand Mixer, Onyx Black
Dyson V8 Cyclone Cordless Vacuum
Ninja CREAMi Scoop & Swirl 13-in-1 Ice Cream and Soft Serve Maker, NC701
As an affiliate, we earn on qualifying purchases.
From Training to Direct Edits
Modern neural networks acquire their behavior through repeated parameter updates. A model’s weights are usually not human-readable instructions: their effects are distributed across many numerical values and layers. Researchers in mechanistic interpretability have previously studied how particular circuits or behaviors appear in weights, but that work has often required specialized analysis, according to the source material.
Weight space programming proposes that a general-purpose AI can perform some of that analysis and editing directly. The MIT authors describe a direction in weight space as a change associated with a behavior, and a rank-one edit as a compact modification to a weight matrix. These concepts help explain the reported targeted change, but they do not mean that every model’s behavior can be mapped cleanly to one small set of parameters.
The MIT post was published on November 2, 2026. ThorstenMeyerAI.com published its account on November 6 and updated it November 9, saying figures referred to 10 published Lexassay tasks and that runs were public on lexassay.com. The supplied account does not include full details for every experiment or enough information to independently assess all reported results.
“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, in the MIT CSAIL post as quoted by ThorstenMeyerAI.com
Questions About Repeatability
The available source material does not establish whether the results can be reproduced across other general-purpose AI systems, larger networks or a broad range of tasks. It reports that the MIT team published prompts, replies and evaluations, but the supplied account does not provide enough methodological detail to independently evaluate the benchmarks or judge how sensitive the results were to prompts and test design.
It is also unclear how reliably the method distinguishes a genuine functional repair from a change that improves one test while causing regressions elsewhere. The legal experiment’s unchanged score despite more answered tasks illustrates that distinction. The reported MATH-500 increase is tied to one model edit and one evaluation; the source does not establish how performance changed on other capabilities or safety tests.
The source material describes potential safety risks, but does not report a complete safety assessment or a demonstrated removal of safeguards. Whether weight edits can be reliably reviewed, reversed and attributed to a particular editor remains unresolved.
Replication and Broader Testing
The next useful steps are independent attempts to reproduce the published experiments and evaluations on additional models. Readers and researchers can review the public runs identified in the source article, while paying attention to the task definitions, scoring methods and whether results hold beyond the tests highlighted by the authors.
Broader evaluations should check for unintended changes as well as the targeted result, including performance on unrelated tasks and safety behavior. Until those tests are available, weight space programming is best treated as an early research approach with promising examples and clear limits, rather than a proven general method for editing neural networks.
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 relying on a conventional training loop for the edit.
What did the MIT researchers report?
The MIT CSAIL team said GPT-6 Astra identified what several small networks did from their weights, generated game-playing networks from zeroed weights without training, and made a targeted edit to a small language model. The reported findings concern specific experiments, not all neural networks.
Did the reported edits improve every task?
No. ThorstenMeyerAI.com reported that a local model answered more legal tasks after an edit, but its legal score remained zero. Another edit improved a narrow JSON test, while the reported MATH-500 gain came from a specific change to one model.
Can this method add a capability a model never learned?
The source article says its experiments indicate the method can address some existing behavior but cannot add what is missing. That is a reported limitation from the described tests, not a universal conclusion established across models.
What safety concerns does the report raise?
Changing weights could potentially alter safety behavior, including weakening it. The source raises this as a risk; it does not provide a confirmed, comprehensive demonstration of safeguards being removed.
Source: ThorstenMeyerAI.com
