🔍 Read the full analysis: Reading, Repairing, And Writing Neural Networks In Weight Space on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Four MIT CSAIL researchers report that OpenAI’s GPT-6 Astra interpreted, repaired and created neural networks by directly reading and changing their weights. Their November 2 blog post logs successes across small game, language and Go networks, but the tests are limited, the post is not peer-reviewed, and the method did not work in every case.
Four researchers at MIT CSAIL report that OpenAI’s GPT-6 Astra could inspect and modify neural-network weights directly, without using a training loop or gradient descent. In a blog post published November 2, the team describes experiments in which Astra identified small networks’ functions, repaired some that failed and wrote new networks—including a small Go model—from scratch. The results offer evidence for a possible new way to work with models, but come from limited experiments documented in a blog post, not a peer-reviewed study.
The researchers call the approach weight-space programming: a language model receives the numerical weights of another neural network and reasons about what they do, then proposes edits or new weights. The team—Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab—says Astra was not given access to tools such as a shell or web search. Depending on the experiment, it either received weights as text and answered directly or wrote code that ran in a restricted test sandbox.
In the reading tests, Astra named the game behind four of four small game-playing networks and identified the shape represented by three of three small image generators. The researchers also report that it located “copying heads,” a particular kind of component, in three tiny language models without running them. These tests suggest weights can reveal information about a model’s function, though they do not establish that the same approach works on larger or more varied systems.
The repair results were mixed. A car-control network that had failed in 256 of 256 episodes succeeded in 1,024 of 1,024 episodes after Astra changed 3.5% of its weights, according to the post. In Qwen3.5-0.8B, an edit affecting 0.016% of the weights was associated with a rise on MATH-500 from 7.9% to 48.1%. The post also reports that Astra wrote small game networks and a 9-by-9 Go network from all-zero weights; the Go model beat a same-size comparison trained on about 20.5 million games after 50 test runs.
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
What Direct Weight Editing Could Change
If the results hold up under broader testing, direct weight editing could give researchers another way to inspect or make targeted changes to small models, especially when conventional retraining is costly or unnecessary. The Qwen result is striking because the reported change touched a very small share of the weights, while the car example shows a repair evaluated against repeated episodes rather than a description alone.
That does not mean language models can generally debug or improve other AI systems. The reported experiments are narrow, and performance on a benchmark is not the same as reliable behavior in deployment. For organizations using open models, the work also points to a governance concern: weights can be changed in ways that may be difficult to notice without testing and provenance records. The authors’ results do not establish a new production-ready method, and the source says users should expect no immediate change.
Products Worth Considering
Bose QuietComfort Ultra Headphones (2nd Gen), Wireless Bluetooth Noise-Cancelling Over-Ear Headphones, Black
Bose TV Speaker (2nd Gen) Compact Soundbar
LG UltraGear 34GX90SA-W 34-inch WQHD Curved OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
From Software 2.0 to Weight Editing
The idea builds on a description Andrej Karpathy gave in 2017, calling neural networks “Software 2.0”: programs formed through optimization in their weights rather than written line by line by a person. The researchers’ proposal reverses the usual direction of that process. Instead of using an optimizer to adjust weights repeatedly from training data, a language model attempts to interpret and write the numerical parameters directly.
The source article says the team released logs covering 1,125 model turns and that it checked them for tool use. It also reports that software engineer Sean Goedecke reproduced one visualization of the Breakout network with a short Python script, a reminder that some apparent patterns in weights may be relatively easy to extract. Separately, a Hugging Face engineer rebuilt the edited Qwen model from the logs and reported a 51.0% MATH-500 score, compared with the blog post’s 48.1% result. That independent reproduction is encouraging but does not replace a full study.
““not through an optimizer, and in particular not with gradient descent””
— The MIT CSAIL researchers, as quoted in their November 2 blog post
How Far the Evidence Reaches
The evidence remains preliminary. The source describes a blog post, not a peer-reviewed paper, and says the experiments used one run per setting, small networks, one model family per task and a single research group. The reported benchmarks and test episodes are specific to those setups; they do not show how Astra would perform across different architectures, larger models or less controlled conditions.
The method also failed in reported cases. Astra could not distinguish an honest model from a sycophantic one using weights alone, and edits did not improve mathematics for Gemma 3 270M, which was not looping. The source says the team itself concluded that editing could not add an ability a model lacked. Further details about repeatability, the robustness of the changes and performance outside the reported tests are not established by the material available here.
Independent Tests and Wider Models
The next useful step is independent replication across more models, tasks and repeated trials, with the test methods and failures reported alongside successful results. The source material does not announce a publication date for a peer-reviewed paper or a planned expansion of the experiments, so those milestones remain unknown.
For now, the researchers’ public logs and the reported Qwen reconstruction provide material other teams can examine. Readers should treat the findings as an early research report: a concrete demonstration in selected small-network tests, not evidence that AI systems can yet repair arbitrary models or replace standard training and evaluation.
Key Questions
What is weight-space programming?
It is the approach tested by the MIT team in which a language model reads or writes another neural network’s numerical weights, reasoning about their function rather than relying on a training optimizer to adjust them.
What did GPT-6 Astra reportedly do?
According to the researchers’ blog post, Astra identified small game and image networks, located copying heads in tiny language models, repaired some failing networks and created several small networks from zero weights.
Did the weight edits work every time?
No. The source reports that Astra could not distinguish an honest model from a sycophantic one using weights alone, and its edits did not improve mathematics on Gemma 3 270M. The researchers also said editing could not add an ability the model lacked.
Has the work been peer-reviewed?
The source describes the November 2, 2026, material as a blog post, not a peer-reviewed paper or an arXiv preprint. It says the researchers released logs, and reports one outside reconstruction of the edited Qwen model.
Does this change how people use AI models now?
The available material does not indicate an immediate change for users. The findings concern limited experiments, and broader tests would be needed before treating direct weight editing as a dependable way to repair or improve models.
Source: ThorstenMeyerAI.com
