🔍 Read the full analysis: Weight-Space Programming: Inside AI’s Work With Model Parameters on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- 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 could identify functions in small neural networks, edit some failing models and write new networks by working directly with their weights. The work is documented in a November 2 blog post, not a peer-reviewed paper, and the reported tests are limited to small models and a single research group.
Four MIT CSAIL researchers report that OpenAI’s GPT-6 Astra identified, repaired and created small neural networks by reading and writing their parameters directly, rather than using a conventional training loop. The experiments, described in a November 2 blog post as “weight-space programming,” include a reported fix to a language model that had been repeating itself, but remain preliminary: the work is not peer-reviewed and was tested on a limited set of models.
The researchers gave Astra the numerical weights of other networks and asked it to infer what they did, change failing models or produce new ones. The team says the model did not use gradient descent. In some tests, Astra received weights as text and had to answer in a single reply; in others, the researchers ran code it wrote in a restricted sandbox. According to the post, tools such as web search and shell access were disabled, and the authors reviewed 1,125 turns in the released logs.
The most striking reported repair involved a car-control network that could not reverse. The team says it succeeded in zero of 256 episodes before Astra changed 3.5% of its weights; after the edit, it succeeded in 1,024 of 1,024 episodes. In another test, an edit changing 0.016% of the weights in Qwen3.5-0.8B coincided with its MATH-500 score rising from 7.9% to 48.1%. A Hugging Face engineer later rebuilt the edited model from the logs and reported a 51.0% score on the same benchmark, according to the source article.
The team also reports that Astra identified four small game-playing networks and the output shapes of three small image generators from their weights. It wrote a 9×9 Go network from zero weights; in 50 test runs, the researchers say, it beat a same-size comparison network trained on about 20.5 million games. These are task-specific results, not evidence that the method works broadly across modern AI systems.
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 Debugging
If the reported approach proves reliable, it could offer researchers another way to inspect and modify models without running a standard optimization process. A model’s weights are usually treated as the result of training, not as a readable program. Reasoning about them directly could make some targeted repairs or audits more practical, particularly for small, open models whose parameters can be examined.
That possibility is different from a general-purpose ability to improve any AI system. The researchers themselves report limits: Astra could not distinguish an honest model from a sycophantic one using weights alone, and its edits did not improve maths performance in Gemma 3 270M, which did not have the looping problem. The authors’ stated takeaway is that editing could not add an ability the model lacked. For users, the experiments do not establish an immediate change to available products or how models are deployed.
Popular right now
Philips Café Aromis Automatic Espresso Machine with LatteGo Pro Milk System
Philips 2200 Series Fully Automatic Espresso Machine with Classic Milk Frother, 2 Coffee Varieties, Intuitive Touch Display, Ceramic Grinder
Vitamix 5200 Professional-Grade Blender, 64 oz. Container, Black
As an affiliate, we earn on qualifying purchases.
From Software 2.0 to Weight Editing
In 2017, computer scientist Andrej Karpathy described neural networks as “Software 2.0”: programs whose behavior is encoded in learned weights rather than written directly by a programmer. He also noted the difficulty of coding in those weights because networks contain millions of parameters. The MIT team’s work revisits that idea by asking whether a language model can interpret or change parameters as if they were a form of code.
A neural network’s weights are numerical values that shape how information passes through the model. In typical training, an optimizer adjusts those values repeatedly in response to data and a learning objective. In the reported experiments, Astra instead reasoned about provided weights and proposed edits. The researchers say their blog post includes logs and that the model had no built-in tools in the tested setup. The source article describes the release as a blog post, not a peer-reviewed study or an arXiv paper.
““Software 2.0””
— Andrej Karpathy, describing neural networks in 2017
How Far the Results Extend
The evidence remains early and narrowly tested. The source says there was one run per setting, small networks, one model family per task and one research group; it also says the work has not been peer-reviewed. The published account does not establish whether other teams can reproduce the results across different models, tasks or parameter scales.
Some demonstrations may also admit simpler explanations. The source article notes that software engineer Sean Goedecke reproduced a visualization used to infer a game network’s structure with a short Python script. That observation does not itself disprove Astra’s result, but it leaves open how much of the performance came from general reasoning about weights versus recognizable patterns in the test setup. The reported benchmark changes also do not establish broader improvements in model quality, safety or reliability.
Replication Is the Next Test
The immediate next step is independent replication using the released logs and model artifacts, followed by tests on more architectures and tasks. The source article reports that a Hugging Face engineer reconstructed the edited Qwen model and obtained a 51.0% MATH-500 score, compared with the researchers’ reported 48.1%; that is a useful check, but not a substitute for broader replication.
Further work would need to clarify which kinds of model behavior can be inferred from weights, how reliably edits transfer beyond the tested cases, and whether changes introduce unwanted effects. Until those questions are answered, weight-space programming is best treated as a research approach with promising demonstrations and significant limits, rather than a ready-made tool for routine model repair.
Key Questions
What is weight-space programming?
It is the researchers’ term for using a language model to reason about and modify another neural network’s weights directly, rather than changing them through a conventional optimizer such as gradient descent.
What did Astra reportedly repair?
The team says Astra edited a car-control network that had failed to reverse and a Qwen3.5-0.8B model that repeated itself. The reported Qwen change was associated with an increase in its MATH-500 score from 7.9% to 48.1%; these results apply to the specific tests described.
Was the research peer-reviewed?
No peer-reviewed publication is identified in the source material. The work is described in a November 2 blog post with released logs, and the source says it is not on arXiv.
Does this mean AI can improve any model?
No. The researchers report failed or limited cases, including no maths improvement for Gemma 3 270M. They say editing could not add an ability the model lacked, and the experiments do not establish that the method works broadly.
What should readers watch for next?
Independent replication and tests across more model types and tasks will show whether the findings extend beyond the reported examples. The current evidence comes from one research group and a limited set of experiments.
Source: ThorstenMeyerAI.com
