🔍 Read the full analysis: Could AI Edit AI? Our Trial Of Weight Space Programming 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 CSAIL researchers report that GPT-6 Astra could identify behavior in small neural networks and write or edit weights without a training loop. A separate trial described by Thorsten Meyer found mixed results: one edit raised a narrow JSON score, while a legal-task edit increased the number answered but left the score unchanged. The work suggests possible uses in model auditing and repair, but does not establish broad improvements or safety.
Researchers at MIT CSAIL reported that OpenAI’s GPT-6 Astra could read and edit neural-network weights directly, without using a training loop, while a separate trial by Thorsten Meyer produced mixed results on a small language model. The work points to a possible new way to inspect or repair open models, but the reported gains were task-specific and do not show that AI can generally improve models or explain all of their behavior.
The MIT team’s November 2 post describes weight space programming: giving a general-purpose AI model the numerical weights of another network and asking it to infer what the network does or modify its weights. The researchers say all their experiments used GPT-6 Astra, with training disallowed; they published prompts, responses and evaluations on GitHub. The post’s authors are Rikiya Takehi and Jayden Teoh, who made equal contributions, along with Alex Zhang and Omar Khattab.
For a reading test, the team supplied weights from four small networks trained to play different grid-based games. Each network had 17,253 weights, and Astra identified all four games, according to the post. In tests involving three small language models, the researchers also report that it found copying heads—components associated with repeating patterns already present in text. The supplied material does not include the full evaluation details for that test.
The researchers also report that Astra wrote game-playing networks from all-zero weights, without training, that beat comparison networks trained for 100 million steps in three games. In an editing test on Qwen3.5-0.8B, they changed a small fraction of its weights to remove one direction of repetitive behavior. Its score on MATH-500 rose from 7.9% to 48.1%, according to the source article. These are reported results from particular tests, not evidence of comparable gains across tasks or models.
Meyer’s separate trial used two AI systems across three experiments. One edit made a local model answer all 10 published legal tasks, up from four, but its legal score remained zero. Another AI review found faults that had passed the trial’s green tests. A one-matrix edit increased a narrow JSON test result from 49 to 70 of 120. Meyer says a further attempt did not produce a useful result. The figures refer to the 10 published Lexassay tasks, and the runs are public on lexassay.com, according to the November 9 update.
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
Possible Repairs, Narrow Evidence
Neural-network weights are the numerical settings that shape how a model processes inputs. Today, developers generally change them through training, which adjusts many values in response to examples. If a model can instead inspect those values and propose targeted edits, that could offer another route to debugging, auditing or repairing open models.
The trials also show why such claims need careful boundaries. In Meyer’s legal test, the model answered more tasks but did not earn a higher score. The JSON improvement concerned a narrow test, and the MIT results covered small, specified networks and tasks. These findings do not establish that a model can reliably diagnose arbitrary failures or improve performance across a broad range of real-world uses.
Direct weight editing could also alter behavior developers intended to preserve. Meyer warns that similar changes might remove safety training. That is a risk raised by the possibility of editing weights, not a reported incident in these experiments. Any practical use would need clear evaluation of both the desired change and unintended effects.
Products Worth Considering
Samsung Galaxy Tab A11+ (6GB RAM, 128GB Storage) – Gray
Logitech MX Ergo S Advanced Wireless Trackball Mouse USB-C – Graphite
Anker SOLIX C300 DC Portable Power Station with 60W Solar Panel, 288Wh, 300W
As an affiliate, we earn on qualifying purchases.
From Training to Direct Edits
In conventional training, a model makes predictions and an algorithm adjusts its weights to reduce errors. The repeated adjustments are commonly called gradient descent. The resulting numbers can look like unstructured decimal values, even though the network’s behavior is encoded across them.
Researchers in mechanistic interpretability have developed ways to identify some patterns inside networks. Meyer’s article cites 2021 work by Anthropic researchers reporting that certain algorithms could be detected directly from weights. The MIT approach asks whether a general-purpose language model can carry out related reading and editing through reasoning, rather than relying on a new training process.
A rank-one edit, the method described in Meyer’s account of the Qwen test, changes a matrix through a linked input pattern and output direction. It is a targeted adjustment, not a rewrite of every parameter. How broadly this approach works depends on the model, the behavior being targeted and the tests used to check the result.
“Our neural network weights will soon be natively optimized by LLMs that deliberately reason about their weights.”
— MIT CSAIL researchers, in their November 2, 2026, post
Limits of the Reported Tests
The source material describes small networks and specific benchmark or task results. It does not establish how reliably the method works on larger models, different architectures or failures not represented in the experiments. Nor does the supplied account provide enough detail to independently assess every evaluation, including the precise setup behind the reported comparisons.
It also remains unclear whether the improvements will hold across repeated runs and broader task sets. In Meyer’s legal trial, answering more tasks did not change the score, illustrating that a change in response coverage may not amount to a meaningful improvement under a given evaluation. The safety effects of weight edits were not measured in the reported examples.
Public Runs and Broader Testing
The MIT researchers have made their prompts, replies and evaluations available on GitHub, while Meyer says the trial runs for the 10 published Lexassay tasks are public on lexassay.com. Those materials allow other researchers to inspect the reported setups and attempt to reproduce the results.
The next useful evidence would come from tests on more models and tasks, repeated evaluations, and checks for unintended changes to model behavior. Until then, the work is best understood as an early demonstration of direct weight reading and editing, with promising but limited results rather than a general replacement for model training.
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 researchers report?
They report that GPT-6 Astra identified the behavior of four small game networks, found copying heads in three small language models, and created or edited networks in specified tests. The results are claims about those experiments, not proof of general performance.
Did the separate trial improve the legal model’s score?
No. Meyer reports that the edited model answered all 10 published legal tasks rather than four, but its legal score remained zero.
Could editing weights affect a model’s safety behavior?
It could, in principle: changing weights may alter behavior a developer intended to preserve. The source raises the possibility that edits could strip safety training, but the reported experiments did not demonstrate that outcome.
Can these results be reproduced?
The MIT team says it published prompts, responses and evaluations on GitHub. Meyer says the runs for the 10 published Lexassay tasks are public on lexassay.com. Independent reproduction and broader testing remain important.
Source: ThorstenMeyerAI.com
