🔍 Read the full analysis: A Look At What Weight-Space Programming Can And Can’t Do 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
An MIT CSAIL team reports that GPT-6 Astra could identify functions in small neural networks, repair two specific failures and write networks for several tasks by working directly with their weights. The results come from a November 2 blog post, not a peer-reviewed paper, and the experiments are limited; the work does not show that a language model can reliably improve any model or add abilities it lacks.
Researchers at MIT CSAIL say OpenAI’s GPT-6 Astra read, edited and wrote neural-network weights directly in eight experiments, including repairs to a car-control model and a small language model. The results, published in a November 2 blog post with logs, suggest a possible way to inspect or make targeted changes to some models, but they do not establish that the approach works broadly: the tests were limited, and the post is not a peer-reviewed paper.
The researchers call the method weight-space programming: a language model receives a neural network’s numerical weights and reasons about what they do, then may change them or write new weights. The team says this differs from ordinary training, which adjusts weights through an optimization process such as gradient descent. In some tests Astra received weights as text and answered in one response; in others, it wrote code that the researchers ran in a restricted sandbox.
In one repair test, a car-control network had recorded zero successful reversals in 256 episodes before the edit. After Astra changed 3.5% of its weights, it succeeded in 1,024 of 1,024 episodes, according to the post. In another test, an edit to 0.016% of Qwen3.5-0.8B’s weights was associated with a MATH-500 score increase from 7.9% to 48.1%. The source says a Hugging Face engineer later reconstructed the edited Qwen model from the logs and reported a 51.0% score; the post does not establish that result through an independent study.
The team also reports that Astra identified the games represented by four small networks and the shapes represented by three image generators, and found copying heads in three tiny language models. It wrote game-playing networks from zero weights and a 9-by-9 Go network. In the Go test, the new network beat a same-size comparison network trained on about 20.5 million games across 50 test runs, according to the researchers. These results concern the particular models and tests described, not neural networks in general.
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 Targeted Model Edits
If the reported results hold up in further testing, weight-space programming could give researchers another way to inspect small models or make targeted changes without running a conventional training process. That could be useful when a model’s weights are available but its training data or original development process is difficult to examine. The work also points to a potential audit use: asking a language model to infer what a network has learned from its numerical parameters.
The practical scope remains limited. The post reports a striking improvement on one maths benchmark and a successful fix to one car-control network, but it also describes cases where editing failed. The authors say Astra could not distinguish an honest model from a sycophantic one by reading weights alone, and edits to Gemma 3 270M did not improve maths performance. The researchers’ own conclusion, as reported in the source, is that editing could not add an ability the model lacked.
For organizations, the findings are a reason to follow research into model inspection and targeted changes, not evidence that the method is ready for routine deployment. For users, the source says nothing changes immediately. Model provenance and independent testing remain relevant, especially if organizations consider adopting edited weights.
Popular right now
KitchenAid Artisan 5-Quart Tilt-Head Stand Mixer, Almond Cream
Dyson Cyclone V10 Origin Cordless Vacuum
Ninja Mega Kitchen System Blender and Food Processor, 1500W, 72-oz Pitcher, 8-Cup Bowl, Black (BL770)
As an affiliate, we earn on qualifying purchases.
From Software 2.0 to Weight Editing
The idea builds on a view of neural networks as programs encoded in their weights. In 2017, computer scientist Andrej Karpathy described neural networks as “Software 2.0,” noting that their behavior is stored in millions of numerical parameters rather than conventional hand-written code. The new work asks whether a general-purpose language model can reason about those parameters directly.
The MIT team members named in the source are Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, with Takehi and Teoh listed as equal contributors. According to the released logs as described in the source, Astra was used through OpenAI’s Codex command-line tool with built-in tools such as shell access and web search turned off. Some trials involved code written by Astra and run by the researchers in a sandbox; that setup is distinct from Astra itself having unrestricted access to tools.
The source describes eight experiments and says the team checked 1,125 turns in the logs. It also notes that the work had one run per setting, used small networks and, for each task, one model family. Those features make the results useful as an early demonstration, but limit how much can be inferred about repeatability or performance on larger, widely used systems.
“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 Results Generalize
The evidence remains early and narrow. The source describes the work as a blog post rather than a peer-reviewed paper or an arXiv preprint, and reports one run per setting. It does not provide evidence that the outcomes will repeat across different model families, tasks, prompts or versions of the language model.
Several details also constrain interpretation. The reported benchmark gains and game results come from particular test setups, while the number of runs for the Go comparison was 50. The post’s description of complete logs and the later Qwen reconstruction offers material for scrutiny, but does not replace a broader independent evaluation. It is also unclear from the source whether the researchers have tested the method on larger production models or whether edits remain reliable when the starting weights, evaluation data or task differ.
The authors report failures as well as successes: Astra did not detect the honesty-versus-sycophancy difference from weights alone, and it did not improve maths performance on Gemma 3 270M. Those cases make it unclear how well the approach can diagnose behavior that is not readily represented in the parameter patterns available to the model.
Independent Tests Will Set the Scope
The immediate next step is independent replication of the logged experiments, including repeated runs and tests across more architectures and tasks. Researchers would need to establish whether the reported repairs and benchmark changes are stable, whether they depend on particular model versions or prompting choices, and how often an edit causes unwanted changes elsewhere.
The source does not announce a publication date, formal peer review or a planned follow-up experiment. Until further results are available, the findings are best treated as a documented proof of concept rather than a general method for maintaining or improving AI systems. Companies evaluating the approach would need to verify both the origin of model weights and the behavior of any edited model before relying on it.
Key Questions
What is weight-space programming?
It is the approach described by the MIT team in which a language model reads or changes another neural network’s numerical weights directly, rather than relying on a conventional training loop to update them.
What did GPT-6 Astra reportedly repair?
The researchers report that Astra edited a car-control network that had failed to reverse in 256 test episodes, after which it succeeded in 1,024 of 1,024 episodes. They also report a substantial MATH-500 score increase after a small edit to Qwen3.5-0.8B.
Does the work show that AI can improve any model?
No. The experiments cover specific models and tasks, and the source reports failures, including no maths improvement after edits to Gemma 3 270M. The researchers say editing could not add an ability a model lacked.
Has the research been peer-reviewed?
The source describes the results as a blog post published November 2, 2026, with released logs. It says the work was not a peer-reviewed paper and was not on arXiv at the time described.
What happens next?
The key test is whether other researchers can reproduce the results across repeated runs, different model families and more tasks. The source does not identify a formal publication or follow-up schedule.
Source: ThorstenMeyerAI.com
