🔍 Read the full analysis: When AI Reads AI: A Look At Weight Space Programming And Our Test 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
MIT CSAIL researchers reported tests in which GPT-6 Astra identified functions in small networks, wrote game-playing networks without training, and edited an existing model. A separate three-experiment test described by Thorsten Meyer found mixed results, including no improvement on a legal benchmark and gains on narrower tests. The work is preliminary and does not establish that AI can reliably repair models or assess their safety.
Researchers at MIT CSAIL have reported experiments in which an AI model read and changed the numerical weights of other neural networks without using a training loop, an approach they call weight space programming. In a separate account, Thorsten Meyer described three tests of the technique: one produced no useful improvement, one review identified faults missed by existing tests, and one narrow edit improved a measured result.
The MIT team used OpenAI’s GPT-6 Astra and said it supplied the model with network weights as text, then asked it to identify functions, create networks or make edits. According to the team’s November 2 post, Astra identified the game behind all four small game-playing networks. Those networks had the same structure and 17,253 weights apiece, so their shapes did not distinguish the games. The researchers also said Astra identified copying heads in three small language models.
For a writing test, the team said Astra produced networks from all-zero weights, without training, that beat networks trained for 100 million steps in each of three games. The account does not provide enough detail here to establish how broadly those results apply beyond the tested tasks. In an editing experiment on Qwen3.5-0.8B, the model changed a direction associated with looping, affecting a reported 0.016% of weights. The MATH-500 score rose from 7.9% to 48.1%, according to the source material.
Meyer’s separate tests used two AI systems and are reported as three experiments. Claude edited a local model that then answered all 10 published legal tasks, up from four, but its legal score remained zero. Another AI review found faults that the existing green tests had not caught. A one-matrix edit raised performance on a narrow JSON test from 49 to 70 out of 120. The article says the published Lexassay runs cover 10 tasks, but the supplied material does not include the full prompts, methods or results tables.
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 Repairs, Narrow Evidence
Directly editing weights could offer a way to inspect or adjust open models without repeating expensive training. The reported results suggest an AI may sometimes identify a specific pattern in a network and make a targeted change, which could help researchers investigate model behavior or test proposed repairs.
But these examples are small and task-specific. In Meyer’s account, an edit that helped a model respond to every legal task did not improve its score, while another change helped only a narrow JSON measure. The source’s stated limit is that weight edits may correct a behavior that is present but faulty, yet cannot add a capability that the model lacks. The experiments do not show that the approach is a dependable substitute for training, evaluation or human review.
The same ability to alter weights also has a safety dimension. Meyer warns that edits could remove safety training as well as change unwanted behavior. That makes the origin and integrity of model weights relevant to anyone deploying open models, although the supplied evidence does not document a real-world safety incident.
Products Worth Considering
Vari Ergo Electric Standing Desk, 54×26, Light Wood/White
Dell UltraSharp U2724DE 27-Inch WQHD IPS Black Monitor
Logitech MX Ergo S Advanced Wireless Trackball Mouse USB-C – Graphite
As an affiliate, we earn on qualifying purchases.
From Training to Direct Editing
Most neural networks acquire their weights through training. An algorithm compares a model’s output with a target and repeatedly adjusts many numerical values, commonly using gradient descent. The resulting behavior is distributed across those values, rather than written as a readable set of instructions. That is one reason trained weights can be difficult to interpret directly.
Researchers in mechanistic interpretability have worked to identify structures and behaviors inside networks. Meyer’s article points to a 2021 Anthropic study that detected some algorithms directly from weights, while noting that such work required specialist methods. The MIT project tests whether a general-purpose language model can perform related reading and editing tasks by reasoning about weights. Its authors, as quoted in the source, describe a future in which language models deliberately reason about their own weights; that is an ambition, not a demonstrated general capability.
“It fixes what is broken; it cannot add what is missing.”
— Thorsten Meyer, describing the editing result in Qwen3.5-0.8B
How Far the Results Extend
The supplied account does not establish whether the results can be reproduced across different AI systems, network sizes or tasks. It also does not provide the complete experimental record for Meyer’s three tests, including all prompts, model settings and evaluation details. The legal benchmark’s score staying at zero despite more tasks being answered illustrates that task coverage and task performance are different measures.
It remains unclear how reliably an AI can distinguish a genuine fault from a useful feature, whether weight edits preserve other capabilities, and how the approach performs under independent evaluation. The MIT team says its prompts, replies and evaluations are public on GitHub; the updated source says the 10 Lexassay tasks and runs are public at lexassay.com. Public materials allow scrutiny, but do not by themselves settle these questions.
Public Runs Invite Replication
The next step is for other researchers to inspect the published prompts, model outputs and evaluations, then repeat the tests on additional networks and tasks. Comparisons should check not only whether a target behavior changes, but also whether unrelated capabilities or safety behaviors are affected.
For now, the reported work is best read as an early demonstration and a set of mixed tests, not proof of a general-purpose method for repairing neural networks. Further independent results would be needed to show how often direct weight editing works, what it breaks, and whether it can be used safely in deployed 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 that task.
What did the MIT researchers report?
Using GPT-6 Astra, they said they identified the function of four small game networks, found copying heads in three small language models, created game-playing networks without training, and made a targeted edit to Qwen3.5-0.8B.
Did the reported edits improve every test?
No. Meyer reported mixed outcomes. The legal model answered more published tasks but its legal score remained zero, while a separate edit improved a narrow JSON test from 49 to 70 out of 120.
Can the method replace model training?
The reported experiments do not establish that. They test selected networks and tasks, and the source says edits may correct an existing behavior but cannot add a capability that is missing.
What are the safety concerns?
A direct edit could change intended behavior as well as unwanted behavior. Meyer specifically warns that edits could strip safety training; the supplied material does not report a documented deployment incident.
Source: ThorstenMeyerAI.com
