🔍 Read the full analysis: Weight-Space Programming: Exploring AI’s Access To Neural Network Internals on ThorstenMeyerAI.com
Get everyday essentials 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 research group says OpenAI’s GPT-6 Astra read and edited neural-network weights directly, without gradient descent, across a set of reported experiments. Results included a repaired car-control model and a small Go network, but the work is described in a blog post rather than a peer-reviewed paper, and the tests were limited.
Researchers at MIT CSAIL report that OpenAI’s GPT-6 Astra identified, altered and created small neural networks by reasoning directly about their numerical weights, rather than using gradient descent. The experiments, described in a November 2 blog post, include a car-control network that began reversing successfully after an edit and a small Go-playing network written from zero; the results are preliminary and have not been presented as a peer-reviewed paper.
The researchers call the approach weight-space programming: a language model reads or writes the numbers that determine how another neural network processes information. In conventional training, an optimizer adjusts weights repeatedly based on data and a specified objective. In the reported tests, Astra was instead asked to infer what existing networks did, correct some failures and produce new networks.
The team reports that Astra identified the game behind four small game-playing networks from their weights and recognized the intended shapes in three small image-generating networks. It also identified copying, or induction, heads in three tiny language models without running them. The source article describes these as experiments in reading a network’s structure from its parameters, not proof that the same method works reliably on large deployed systems.
For one car-control network, the researchers say Astra changed 3.5% of the weights. The model had recorded zero successful reversals across 256 episodes before the edit and 1,024 successes across 1,024 episodes afterward. In a separate test on Qwen3.5-0.8B, the source reports that an edit to 0.016% of its weights raised its MATH-500 score from 7.9% to 48.1%. A Hugging Face engineer later rebuilt the edited Qwen model from the released logs and reported a 51.0% score, according to the source article.
The researchers also report creating small game-playing networks from all-zero weights and a 9-by-9 Go network with 1.5 million weights. After 50 test runs, the Go network beat a same-size comparison network trained on about 20.5 million games. These outcomes are reported by the research group; the source says there was one run per setting and does not describe independent replication across the full set of experiments.
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
Potential for Direct Model Repair
If the method holds up under broader testing, it could offer researchers a way to inspect and make targeted changes to models whose internal behavior is hard to interpret. The reported Qwen result, in particular, suggests that a small parameter edit may affect a specific benchmark score without retraining the whole model. That is a research possibility, not yet a demonstrated production workflow.
Direct access to model weights also creates practical questions about auditability and provenance. A system that can alter parameters could be used to modify a model after training, so developers would need ways to track which weights were changed and test the resulting behavior. The source article says the work does not change anything for users today. Its immediate relevance is to AI researchers and organizations that maintain open models, rather than to ordinary users of commercial chatbots.
The reported failures matter alongside the successes. Astra could not distinguish an honest model from a sycophantic one from the weights alone, and edits did not improve mathematics performance in Gemma 3 270M, a model that did not exhibit the looping problem targeted in another test. The authors’ stated limitation is that editing could not add an ability the model lacked. That boundary cautions against treating weight edits as a general substitute for training or evaluation.
Products Worth Considering
Bose SoundLink Flex Bluetooth Speaker (2nd Gen) – Portable Outdoor Speaker with Hi‑Fi Audio, Waterproof and Dustproof, USB‑C, Up to 12 Hours
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
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 weights produced through optimization, rather than written directly by a programmer. The MIT group’s work tests whether a general-purpose language model can reason about those weights as if they were a form of program.
The researchers are Rikiya Takehi and Jayden Teoh, who the source says contributed equally, along with Alex Zhang and Omar Khattab. According to the released logs described in the article, Astra was called through OpenAI’s Codex command-line tool with built-in tools such as shell access, web search and file editing disabled. Some tasks gave it weights as plain text for a one-response answer; other tests let a research sandbox run Python code written by Astra, under a fixed test budget and without network access.
The source article says the group reviewed 1,125 turns in the release and found them tool-free. It also emphasizes that the evidence is a public blog post with logs, not a peer-reviewed paper or an arXiv preprint. That distinction matters: detailed logs make experiments easier to inspect, but they do not replace review, replication or tests across additional model families and settings.
““not through an optimizer, and in particular not with gradient descent””
— The MIT research team, in its November 2 blog post
Limits of the Current Evidence
The available evidence remains limited to one research group and one run per setting, according to the source article. The tests involved small networks and, for each task, a narrow model selection. The blog post’s logs allow inspection of the reported interactions, but the source says the work has not been peer-reviewed or posted on arXiv.
It is not yet clear whether the results can be reproduced independently across other language models, larger neural networks or tasks beyond those tested. The source also describes a potential interpretive issue in the game-identification experiment: software engineer Sean Goedecke said he could reproduce the visual pattern used to identify one game with a short Python script. That observation does not invalidate Astra’s answer, but it leaves open how much of the result reflects meaningful interpretation versus a simpler analysis of the weights.
The reported benchmark gains also do not establish that edited models are more reliable overall. The source notes that Astra failed to distinguish honest and sycophantic models from their weights and did not improve maths in a model that was not looping. Broader evaluation would be needed to determine whether targeted gains come with regressions elsewhere.
Replication and Broader Testing
The next useful step is independent replication of the reported tasks, including the car-control repair, Qwen maths edit and Go-network result. Researchers would also need to test whether changes remain effective across repeated runs and whether they affect capabilities beyond the target metric.
The source article points to the team’s released logs as material other researchers can inspect. It does not report a scheduled peer-reviewed publication, follow-up study or product release. Until additional results are available, weight-space programming should be treated as an experimental research direction, not a ready-made tool for modifying production models.
Key Questions
What is weight-space programming?
It is the approach tested by the MIT group in which a language model reads or edits another neural network’s numerical weights directly, rather than relying on gradient descent to train it.
What did GPT-6 Astra reportedly accomplish?
The source article says Astra identified several small networks’ functions, edited a car-control model so it passed reversal tests, changed a small Qwen model associated with looping, and wrote small game networks, including one for 9-by-9 Go.
Has the research been peer-reviewed?
No peer-reviewed publication is reported in the source material. The experiments were presented in a public blog post with logs, and the source says the work was not on arXiv.
Does this show AI can fix any neural network?
No. The reported tests were limited, and the source describes failures: Astra could not identify sycophancy from weights alone, and its edits did not improve maths performance in a model that did not have the tested looping issue.
Will this change how people use AI models now?
The source article says nothing changes for users today. The work is an early research result, with broader replication and testing still needed before any practical use can be assessed.
Source: ThorstenMeyerAI.com
