🔍 Read the full analysis: Weight-Space Programming Brings AI Closer To Neural Network Parameters 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 that OpenAI’s GPT-6 Astra directly analyzed and edited neural network weights in eight experiments, without gradient descent. It repaired two small models and wrote networks for games, but the results come from a single research group’s blog post, not a peer-reviewed paper, and several tests were limited.
Four MIT CSAIL researchers say they used OpenAI’s GPT-6 Astra to read, modify and generate neural network weights directly, without using gradient descent to train the networks. In experiments described in a November 2 blog post, the model identified small networks’ functions, repaired two failing systems and wrote game-playing networks; the researchers call the approach weight-space programming.
The researchers gave Astra the numerical weights of other networks and asked it to infer what they did or make targeted changes. The released logs describe two kinds of setup: in some tests, Astra received weights as text and answered in one reply; in others, it wrote Python code that ran in a restricted sandbox. The team says it disabled Codex tools such as shell access and web search, and reviewed 1,125 turns in the logs. Those are the authors’ descriptions of their experimental controls.
In a car-control test, the network had recorded zero successful reversals in 256 episodes before Astra changed it. The researchers report that the edited version succeeded in 1,024 of 1,024 episodes, with 3.5% of its weights changed. In a separate test, changing 0.016% of the weights in Qwen3.5-0.8B raised its reported MATH-500 score from 7.9% to 48.1%. The source says a Hugging Face engineer later reconstructed that edit from the logs and measured 51.0% on the benchmark.
The post also reports that Astra identified four small game-playing networks and the output shapes of three small image generators from their weights. It wrote small game networks from all-zero weights and created a 9-by-9 Go network; the authors say that network beat a same-size comparison trained on about 20.5 million games in 50 test runs. These outcomes are reported by the researchers, and the post describes limited experiments rather than broad evidence that the technique works across models.
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 Repair
If the approach holds up in further tests, it could offer researchers a way to inspect or make narrow changes to some models without running a conventional training process. The reported Qwen result is especially relevant because a small fraction of weights was edited and the benchmark score rose substantially in the authors’ test. A separate reconstruction, reported by the source article, gives some additional support for that particular result, though it does not establish how reliably the technique works elsewhere.
Directly changing weights also shifts attention to model provenance and verification. Organizations that deploy open models may want to know who altered a weight file and whether a change has been independently checked. The research does not show that companies can safely or reliably use Astra this way in production, and the source says there is no immediate change for users.
Products Worth Considering
Samsung Odyssey G4 25-inch FHD Gaming Monitor
Amazon Kindle Scribe (64GB) with Premium Pen, Metallic Jade
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
As an affiliate, we earn on qualifying purchases.
From Learned Weights to Editable Code
In 2017, Andrej Karpathy described neural networks as “Software 2.0”: programs encoded in weights learned through optimization rather than written line by line by people. The researchers’ proposal reverses part of that framing. Instead of relying on an optimizer to adjust weights through training, they ask a general-purpose language model to reason about the numbers and edit them directly.
The MIT team consists of Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab; Takehi and Teoh are described as contributing equally. The source material says the work was presented in a research blog post with public logs, not a peer-reviewed paper or an arXiv preprint. The distinction matters: the reported results are useful evidence of what the group tested, but have not undergone the same review process as a published research paper.
““not through an optimizer, and in particular not with gradient descent””
— The MIT CSAIL researchers, in their November 2 blog post
Small Tests Leave Reliability Open
The evidence has clear limits. According to the source material, the results come from one research group, with one run per setting, small networks and a single model family for each task. The work is a blog post rather than a peer-reviewed study, and the reported benchmark changes do not establish repeatability across other models, datasets or research teams.
The approach also did not succeed in every test. Astra could not distinguish an honest model from a sycophantic one by inspecting weights alone, and its edits did not improve maths on Gemma 3 270M. The source does not establish whether the method can reliably diagnose more complex failures, preserve abilities outside the tested task, or scale to large production models. The exact cause of the reported improvements and their durability under wider evaluation remain open questions.
Independent Replication Is the Next Test
The researchers have released logs that other researchers can examine, and the source reports one outside reconstruction of the Qwen edit and benchmark result. Further tests would need to repeat the experiments across models and settings, check for side effects on other capabilities, and compare direct editing with standard fine-tuning or other repair methods. The source material does not give a timetable for a peer-reviewed paper or a broader evaluation.
For now, the results are a research demonstration, not a production tool or a change to how users interact with AI systems. Whether weight-space programming proves useful beyond these examples will depend on independent replication and evidence that its edits remain reliable across more demanding tests.
Key Questions
What is weight-space programming?
It is the researchers’ name for using a language model to inspect or edit another neural network’s numerical weights directly, rather than adjusting them through a conventional optimizer such as gradient descent.
What did GPT-6 Astra reportedly fix?
The researchers say Astra edited a car-control network that had failed to reverse and a Qwen3.5-0.8B model that produced looping responses. They report improved results in the tests they ran; those findings have not established general reliability across models.
Was the work peer-reviewed?
No. The source describes the results as a blog post with public logs, not a peer-reviewed paper or an arXiv preprint.
Does this mean AI can repair any model?
No. The reported experiments were limited, and the researchers say edits did not improve maths on Gemma 3 270M. They also say weight edits could not add an ability a model lacked.
What happens next?
Independent researchers can examine the released logs and test whether the results repeat across different models and tasks. The source gives no timetable for a broader study or peer-reviewed publication.
Source: ThorstenMeyerAI.com
