🔍 Read the full analysis: How An AI Read, Repaired, And Wrote Other Neural Networks 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
Four MIT CSAIL researchers report that OpenAI’s GPT-6 Astra could identify what small neural networks did, repair some failures and write new networks by editing their weights directly. Their public logs document striking results alongside failures; the work is a blog-post report, not a peer-reviewed paper, and the experiments were limited in scope.
Four MIT CSAIL researchers report that OpenAI’s GPT-6 Astra identified, repaired and wrote small neural networks by working directly with their numerical weights, rather than using the usual training process. The experiments, described in a public blog post on November 2, 2026, include a repaired car-control model and a new Go-playing network, but also document cases where Astra’s edits did not help.
The researchers call the method weight-space programming: a language model examines or changes the numbers that define another neural network, reasoning about what those numbers do instead of relying on an optimizer or gradient descent. The team—Rikiya Takehi and Jayden Teoh, who contributed equally, with Alex Zhang and Omar Khattab—published experiment descriptions and logs. The account is a research blog post, not a peer-reviewed paper.
In the reported reading tests, Astra identified the game behind four of four small game-playing networks and inferred the shape represented by three of three small image generators. It also identified every “copying head,” a component associated with repeated-token behavior, in three tiny language models without running them. These results concern the particular small models tested; they do not establish that the method can reliably interpret arbitrary neural networks.
The repair examples included a car-control network that had recorded zero successful reversals in 256 episodes. After Astra changed 3.5% of its weights, the researchers reported 1,024 successes in 1,024 episodes. In another test, an edit affecting 0.016% of the weights in Qwen3.5-0.8B raised its reported MATH-500 score from 7.9% to 48.1%. A Hugging Face engineer later rebuilt that edited model from the released logs and reported a 51.0% score.
Astra also wrote small game-playing networks from all-zero weights and a 9×9 Go network with about 1.5 million weights. The post says the game networks beat the researchers’ comparison networks, which had been trained for 100 million steps. In 50 test runs, the Go network beat a same-size comparison model trained on about 20.5 million games. The post’s results are reported for those specific tests and do not establish performance across broader benchmarks.
Infographic 01 · Eight experiments
Read, fix, write: what GPT-6 Astra did with raw weights
Every experiment in the weight-space programming post, with the size of the network, the result and how it went.
Read
- 1Which game was this network trained on?4 reinforcement-learning policies · 17,253 weights each · weights as text, no toolsRight game for 4 of 4, without a list of options✓ Worked
- 2What does this image model draw?3 tiny diffusion models · 23,824 weights · no forward pass allowedRight shape for 3 of 3, but drawn as outlines; the models draw filled shapes✓ Worked
- 3Which language did this model learn?3 character-level transformers · 7.16M weights · middle blocks only, never runRight language 3 of 3; every copying head found, none invented✓ Worked
- 7Which of two models is sycophantic?3 pairs of Gemma 3 270M fine-tunesWeights alone: could not tell in 4 attempts. With attention patterns: 3 of 3! Partly
Fix
- 4Repair a car that never backs upKinetix control network · 171,664 weights · one reply, no tools0 of 256 episodes before, 1,024 of 1,024 after; 3.5% of weights changed✓ Worked
- 8Make Qwen3.5-0.8B better at maths0.8 billion weights · Python over the weights · 10 / 25 / 50 test passesMATH-500 7.9% to 48.1% with 50 passes (12.3% and 14.1% with 10 and 25)! Partly
- 8bSame task, a model that does not loopGemma 3 270M · 5 to 50 test passesMATH-500 6.8% to at most 7.7%: “not established”✕ Failed
Write from zero
- 5Write arcade-game networks3 MinAtar games · 20,486 weights, all starting at 0 · one reply, no toolsBeat the 100M-step comparison networks in 3 of 3, the 500M-step ones only in Breakout! Partly
- 6Write a 9×9 Go network1,480,274 weights, all starting at 0 · no training, no search0 passes: 54% vs small trained nets. 50 passes: 70–79% vs same-size nets trained on ~20.5M games✓ Worked
✓ Worked · ! Partly or with conditions · ✕ Failed. One run per setting, no error bars. A blog post with complete logs, not a peer-reviewed paper.
Sources: Takehi, Teoh, Zhang, Khattab (MIT CSAIL), blog post, 2 Nov 2026, lm_understand-logs (CC BY 4.0). Model: OpenAI GPT-6 Astra via the Codex CLI, built-in tools switched off.
Powered by Thorsten Meyer AI
Direct Edits Could Change Model Audits
If the approach holds up beyond these demonstrations, it could offer researchers a different way to inspect or make targeted changes to small models. Instead of treating weights only as the output of training, a capable language model might help identify patterns or adjust a limited set of parameters. The Qwen result, in particular, suggests that a small edit can coincide with a large benchmark change in a particular model and task.
That possibility comes with practical limits. The report does not show that an AI can generally diagnose a model’s behavior, safely fix it or improve it across tasks. The researchers could not distinguish an honest model from a sycophantic one using weights alone. Their attempt to improve maths performance in Gemma 3 270M did not work; the authors said editing “could not add an ability the model lacked.” For organizations, the work raises a reason to examine how model weights are produced and modified, but it does not change what users can do with models today.
Products Worth Considering

Dell XPS Tower 8960 Desktop, Intel Core i7-13700, NVIDIA GeForce RTX 4060, 16GB DDR5, 512GB SSD

Dell UltraSharp U2724DE 27-Inch WQHD IPS Black Monitor

HP OfficeJet 250 Wireless Mobile All-in-One Inkjet Printer
As an affiliate, we earn on qualifying purchases.
From Neural Weights to Readable Code
In 2017, computer scientist Andrej Karpathy described neural networks as “Software 2.0”: programs encoded in learned weights rather than written directly by people. A network’s weights determine how strongly its artificial neurons influence one another. Ordinarily, an optimizer adjusts those values repeatedly during training, guided by data and a learning objective. The resulting parameter list is not readily readable like conventional source code.
The MIT group’s tests ask whether a language model can instead reason about that parameter list. According to the researchers’ release, Astra was called through OpenAI’s Codex command-line tool with built-in tool capabilities disabled. Some tasks supplied weights as plain text and required a single response; in others, researchers ran code written by Astra in a restricted sandbox with a fixed test budget. The team says it reviewed 1,125 turns and found them tool-free. These are details reported by the researchers, not an independent audit of the full experimental process.
““Software 2.0””
— Andrej Karpathy, describing neural networks in 2017
How Far the Experiments Generalize
The scope and reliability of the findings remain open questions. The report describes one run per setting, small networks, a single model family for each task and work by one research group. The examples do not establish how the method performs on larger models, other architectures or repeated independent trials. Although public logs make the work easier to inspect, the blog post has not been peer reviewed and is not on arXiv, according to the source account.
Some interpretations may also admit simpler explanations. In the game-identification test, the post points to patterns in the weights as clues; software engineer Sean Goedecke reportedly reproduced a similar visualization with a short Python script. The source account does not establish whether that observation undermines Astra’s identification or merely offers another way to expose the same patterns. The strength of the reported results, their repeatability and the safeguards needed for edits to more consequential models remain unclear.
Replication Is the Next Test
The next meaningful step is independent replication across more models, tasks and repeated runs, with clear comparisons against standard repair and training methods. Researchers would also need to test whether edits retain their gains outside the reported benchmarks and whether they introduce regressions elsewhere. The public logs provide material for others to examine, but the source report does not announce a peer-reviewed publication or a broader deployment.
For now, the reported work is an early demonstration rather than a ready-to-use tool for model developers or the public. The key question is whether other teams can reproduce the results and establish when direct weight editing is dependable—and when it is not.
Key Questions
What is weight-space programming?
It is the researchers’ term for using a language model to inspect or change another neural network’s weights directly, rather than training that network through an optimizer such as gradient descent.
What did Astra repair?
The MIT post reports a successful edit to a car-control network that had failed to reverse in 256 episodes, and a small edit to Qwen3.5-0.8B associated with a higher MATH-500 score. These are results from specific experiments, not evidence that all models can be repaired this way.
Did the method work in every test?
No. The researchers say they could not identify an honest model versus a sycophantic one from weights alone, and edits did not improve maths performance in Gemma 3 270M.
Has the work been peer reviewed?
The source describes it as a public blog post with released logs, not a peer-reviewed paper or an arXiv paper. The results therefore need further independent testing.
Can people use this to change models now?
The report does not announce a public product or show that the approach is ready for routine use. The experiments are limited demonstrations, and the safety and reliability of edits to broader models remain unsettled.
Source: ThorstenMeyerAI.com
