🔍 Read the full analysis: From AI Weights To AI Rewrites: Our Weight Space Programming Experiment on ThorstenMeyerAI.com
Get the latest gadgets 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 researchers report that GPT-6 Astra could identify behaviors in small neural networks, write game-playing networks without training, and edit a small language model. A separate set of experiments described in the source produced mixed results, including a narrow benchmark gain but no improvement in the legal-task score. The work is experimental; broader reliability and safety implications remain untested.
Researchers associated with MIT CSAIL reported that OpenAI’s GPT-6 Astra could inspect, create and edit other neural networks by working directly with their weights, without using a training loop. The published experiments, described in a November 2 post, suggest a possible new way to analyze and modify open models, but a separate set of tests reported by Thorsten Meyer produced mixed results and does not establish that the approach works reliably beyond narrow tasks.
The MIT team’s work, by Rikiya Takehi and Jayden Teoh, with Alex Zhang and Omar Khattab, asked Astra to reason about networks’ numerical weights rather than train them through repeated examples and gradient descent. The team says it published prompts, model replies and evaluations on GitHub. The source describes experiments in reading networks, generating them from zero-valued weights, and making a targeted edit to a small language model.
For one reading test, Astra received the weights of four small game-playing networks with the same architecture and was asked to identify the game each played. The source says it named all four correctly, including Breakout, for which it stated 99% confidence. In another test, the researchers examined three small language models and report that Astra located every copying head in them. These are results on the specific networks tested, not evidence that the system can interpret arbitrary models.
The source also reports that Astra wrote game-playing networks from zero-valued weights without training; those networks beat counterparts trained for 100 million steps in three games. In an editing test, it changed a small number of weights in Qwen3.5-0.8B, removing one direction associated with looping. The reported MATH-500 result rose from 7.9% to 48.1%. Meyer’s separate tests found that Claude’s edit let 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 green tests, while a one-matrix edit raised a narrow JSON test from 49 to 70 of 120.
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
A New Route to Model Repair
If the method holds up beyond the reported demonstrations, directly inspecting weights could give developers another way to diagnose and modify a model, especially when its original training data or process is unavailable. That could complement ordinary evaluation, which tests outputs but may not reveal what internal patterns produce them. The source’s examples point to possible uses in auditing open models and making targeted repairs.
The reported improvements also need careful interpretation. A higher score on one benchmark or test does not establish a broad improvement in capability, and the legal example shows why task completion and task scoring should not be treated as the same outcome. Meyer’s account suggests the method may work best when it can remove a specific unwanted behavior, rather than supply a capability a model lacks.
There is a safety concern alongside the potential benefits: modifying weights can alter behavior in ways that are hard to predict. The source warns that similar edits could remove safety training. Model owners may need stronger records and checks around weight changes, but the experiments described do not measure real-world misuse or establish a standard for safe editing.
Products Worth Considering
EcoFlow Delta 3 Max Portable Power Station with 220W Solar Panel, 2048Wh
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite
As an affiliate, we earn on qualifying purchases.
From Training Loops to Weight Edits
Neural-network weights are numerical values that shape how information passes through a model. During conventional training, an algorithm adjusts those values repeatedly in response to errors. The resulting behavior is distributed across many weights, which can make the raw numbers difficult for people to interpret directly.
Researchers in mechanistic interpretability have studied how specific structures inside networks relate to behavior. The development described here differs in its proposed use of a general-purpose language model to read or write weights by reasoning, rather than relying on a new training run. The MIT post calls this weight space programming. Its ambition is broad, but the evidence supplied in the source is limited to particular small networks and reported evaluations.
The source was updated on November 9 to say its figures refer to 10 published Lexassay tasks and that the runs are public on lexassay.com. This update matters for interpreting Meyer’s local-model results: the legal-task count is a small, specified set, not a general measure of legal competence.
“One experiment did nothing useful, one review caught real faults, and one edit produced a narrow, measured win.”
— Thorsten Meyer, describing the follow-up experiments
Limits of the Reported Tests
The source does not establish whether the results will replicate across different model families, larger networks, or independent evaluations. The MIT experiments use GPT-6 Astra, and the source provides no comparison showing how other AI systems would perform on the same tasks. The reported benchmark scores and game results also do not reveal how sensitive outcomes are to prompts, model versions or evaluation choices.
It is also unclear how often edits that improve one test harm other capabilities, and how developers can reliably detect unintended changes before deployment. The source describes a narrow legal-task result whose score remained zero, so it should not be read as evidence of improved legal performance. More broadly, the safety effects of editing out behaviors associated with training have not been quantified in the material provided.
Independent Replication and Safety Checks
The next useful step is independent testing on additional architectures and tasks, with complete evaluation results and comparisons against standard fine-tuning or other editing methods. The MIT team’s published prompts, replies and evaluations provide material for scrutiny; the source also says the follow-up Lexassay runs are public. Replication can test whether reported gains persist beyond the original setups.
For model developers, any weight edit intended for deployment would need checks across a broad set of capabilities, not only the task targeted by the change. That includes testing for regressions and unintended safety effects. The current source reports experiments and a proposed direction for research; it does not announce a product release or a validated process for routine model maintenance.
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 changing them through a conventional training loop.
What did the MIT team report?
The source says GPT-6 Astra identified four small game networks, found copying heads in three small language models, generated game-playing networks without training, and edited a small model with a reported MATH-500 score increase from 7.9% to 48.1%.
Did the follow-up experiments show broad improvement?
No. Meyer reports mixed outcomes. One test produced a narrow JSON result increase, while a legal-task experiment increased the number of tasks answered but left the legal score at zero.
Can weight edits affect safety?
The source warns that edits could remove safety training, but it does not quantify that risk. The experiments described do not establish a validated safety process for editing models.
What remains to be tested?
Independent replication across larger and different models, checks for capability regressions, and evaluation of unintended safety effects remain necessary before the approach can be judged as a dependable tool.
Source: ThorstenMeyerAI.com
