🔍 Read the full analysis: Why Researchers Are Exploring AI That Works With Neural Network Weights 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
Four MIT CSAIL researchers report experiments in which OpenAI’s GPT-6 Astra analyzed and edited neural-network weights directly, rather than using gradient descent. The results include targeted repairs and small networks written from scratch, but the work is a blog post with limited tests, and the method’s broader reliability is unknown.
Four researchers at MIT’s Computer Science and Artificial Intelligence Laboratory report that OpenAI’s GPT-6 Astra could identify, edit and write small neural networks by working directly with their numerical weights, without using gradient descent. The November 2 blog post calls the approach weight-space programming; its logged experiments include a repaired car-control model and a language-model edit associated with a higher maths score, though the results remain preliminary.
Neural-network weights are the numerical values that shape how a model processes inputs. Usually, a training algorithm adjusts those values repeatedly using data and an optimization method. In the reported experiments, Astra was instead asked to reason about the weights themselves: identify what a network does, change parts of it, or create a network from an initial set of values. The researchers say the approach did not use a training loop or gradient descent for these tasks.
The team reports several distinct results. Astra identified the game behind four of four small game-playing networks from their weights, and identified the general shape produced by three of three small image generators. In a car-control task, the network recorded no successful episodes in 256 attempts before an edit; afterward it succeeded in 1,024 of 1,024 reported episodes, with 3.5% of weights changed. These are results from the researchers’ test setup, not evidence that the method will work across other models or environments.
In another experiment, an edit to 0.016% of Qwen3.5-0.8B’s weights was associated with an increase on the MATH-500 benchmark from 7.9% to 48.1%, according to the post. The source says a Hugging Face engineer rebuilt the edited model using the released logs and recorded 51.0% on the same benchmark. Astra also wrote a 1.5-million-weight Go network from zero; after 50 test runs, it beat a same-size comparison network trained on about 20.5 million games. The post supplies logs, but reports only one run per setting.
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
A New Route to Model Repair
If the approach can be reproduced and extended, it could offer researchers another way to inspect or adjust small models, especially when a targeted behavior needs attention. Direct edits could be useful for diagnosing a model’s internal structure or testing whether a narrow problem can be corrected without retraining it. The reported Qwen result makes that possibility concrete, but it does not establish that edits are dependable, general, or safe for production systems.
The work also raises a practical issue for organizations using open models: a model’s weights are not just a passive file if an AI system can interpret and alter them. Any deployment would need checks for unintended behavioral changes and a clear record of the model’s origin and modifications. The source itself says there is no immediate change for users; these experiments do not show that ordinary products are using weight-space programming.
Products Worth Considering
Apple MacBook Air 13-inch with M5 chip, 16GB memory, 512GB SSD, and 12MP Center Stage camera
Apple MacBook Air 15-inch with M5 chip, 16GB Memory, 512GB SSD, and Liquid Retina Display
Amazon Kindle Paperwhite (2024), 16 GB, 7-inch Display, Black
As an affiliate, we earn on qualifying purchases.
From Software 2.0 to Weight Editing
In 2017, Andrej Karpathy described neural networks as “Software 2.0”: programs whose behavior is encoded in learned weights rather than written line by line by a programmer. The MIT group’s work asks whether a general-purpose language model can treat those weights as something it can inspect and edit, rather than leaving their adjustment entirely to conventional training algorithms.
The researchers—Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab—describe a constrained setup. According to the released logs, Astra was accessed through OpenAI’s Codex command-line tool with built-in tool capabilities switched off. Some tasks supplied weights as text and required a single response; in others, the researchers ran code written by Astra in a sandbox with a fixed testing budget. The team says it reviewed 1,125 turns and found no tool use. The source is a research-group blog post, not a peer-reviewed paper or an arXiv preprint.
“Software 2.0”
— Andrej Karpathy, describing neural networks in 2017
How Far the Results Extend
The evidence has clear limits. The source describes one run per setting, small networks, one research group and one model family per task. A blog post and released logs allow scrutiny, but do not substitute for independent, peer-reviewed evaluation across more tasks and model types. It is also unclear whether the reported gains would persist under repeated testing or changes in evaluation conditions.
The reported failures matter as well: Astra could not distinguish an honest model from a sycophantic one using weights alone, and edits did not improve maths performance on Gemma 3 270M, a model that did not exhibit the looping problem. The authors say editing “could not add an ability the model lacked.” The source does not establish how reliably the method can diagnose other behaviors, or whether a useful edit might cause problems elsewhere.
Independent Tests and Broader Models
The immediate next step is independent replication of the logged experiments, followed by tests across more model sizes, architectures, tasks and repeated runs. Researchers would also need to establish whether edits remain effective outside the narrow conditions used in the post, and whether they introduce unwanted changes. The source does not announce a publication schedule or a specific follow-up milestone, so the timing of such work is not confirmed.
For now, the results are best read as an early demonstration of a research direction, not a ready-to-use model-maintenance method. The public logs give other researchers material to examine, while questions about reliability, scope and safe deployment remain open.
Key Questions
What is weight-space programming?
It is the approach tested in the blog post: a language model reads or changes another neural network’s numerical weights directly, rather than relying on a conventional training loop and gradient descent for the task.
Did Astra successfully repair a language model?
The researchers report that an edit affecting 0.016% of Qwen3.5-0.8B’s weights coincided with a MATH-500 score change from 7.9% to 48.1%. A Hugging Face engineer reportedly rebuilt the edited model and recorded 51.0%. The results have not been established across repeated independent trials.
Does the work mean AI models can now fix themselves?
No. The experiments tested one language model working on other networks in specific research settings. The post reports failures as well as successes, and does not show that the technique is broadly reliable or in everyday use.
Has the research been peer-reviewed?
The source describes it as a blog post published by the researchers on November 2, 2026, with public logs. It says the work is not a peer-reviewed paper and is not on arXiv.
Source: ThorstenMeyerAI.com
