🔍 Read the full analysis: How AI Used Model Parameters To Make Changes To Neural Networks 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 interpreted and modified neural-network weights directly in a set of eight experiments. It fixed two small models and wrote networks for games, but the work is a blog post with limited tests, and the researchers say weight edits could not give a model an ability it lacked.
Four researchers at MIT’s Computer Science and Artificial Intelligence Laboratory say OpenAI’s GPT-6 Astra read, changed and generated neural-network weights directly in eight experiments, without gradient descent. The work, described in a November 2 blog post as “weight-space programming,” included repairs to a car-control network and a small language model, alongside tests that exposed limits in what weight edits could accomplish.
In conventional machine learning, an optimization method adjusts a network’s many numerical weights during training. The MIT group instead supplied Astra with weights from other networks and asked it to infer their functions or alter them. The researchers say the model reasoned about the numbers directly, rather than using an optimizer or gradient descent. Some tasks required Astra to respond from plain-text weights in one turn; in others, researchers ran code written by the model in a restricted sandbox.
The blog post reports that Astra identified the games behind four small game-playing networks and the general forms of three small image generators. It also identified “copying heads,” a component associated with copying patterns, in three tiny language models. In a car-control test, a network that had recorded no successful reversals across 256 episodes reportedly reached 1,024 successes in 1,024 episodes after Astra changed 3.5% of its weights.
In another repair, the researchers report that editing 0.016% of the weights in Qwen3.5-0.8B raised its score on MATH-500 from 7.9% to 48.1%. A Hugging Face engineer later reconstructed that edited model from the released logs and reported a score of 51.0%. Astra also generated small game-playing networks and a 9-by-9 Go network from zero weights. According to the post, the Go network beat a same-size comparison network trained on about 20.5 million games in 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 reported results hold up in broader tests, weight-space programming could offer a way to inspect or make targeted changes to small neural networks without running conventional retraining. That could matter to teams working with open models, especially when a failure appears localized and a full training run would be costly. The experiments suggest that a general-purpose language model may be able to infer some structure from numerical parameters and propose changes to them.
The results do not show that AI can reliably repair arbitrary models or replace training. The tests concern particular small networks and specific tasks, and one reported improvement does not establish that the same method will work on larger systems. For model users, the research does not itself change how existing services behave. For developers and organizations deploying models, it highlights a potential new editing method as well as a need to track the source and integrity of model weights.
Products Worth Considering

Apple MacBook Air 15-inch (2025) with M4 Chip, 24GB Memory, 512GB SSD, Midnight

Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, 12MP Center Stage Camera, and Wi-Fi 7, Sky Blue

Bose Lifestyle Ultra Soundbar and Ultra Subwoofer Bundle
As an affiliate, we earn on qualifying purchases.
From Software 2.0 to Weight-Space Tests
The idea builds on the view of neural networks as software encoded in their weights. In 2017, computer scientist Andrej Karpathy described this approach as “Software 2.0”: instead of a person writing every instruction, optimization produces a program in the network’s parameters. The MIT researchers ask whether a language model can go a step further by treating those parameters as material it can inspect and edit.
The team’s authors are Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, with Takehi and Teoh listed as equal contributors. The post says Astra was accessed through OpenAI’s Codex command-line tool with built-in tools disabled. Where code execution was part of an experiment, the researchers used a sandbox without network access and limited test runs. They say they reviewed all 1,125 turns in the released logs and found no tool use by Astra.
The report is a research group’s blog post, not a peer-reviewed paper, and the source material says it was not posted on arXiv. The researchers describe one run per setting, small networks and a limited range of tasks. Those details make the results useful as an early demonstration, but not evidence of general performance across models or applications.
““No human writes that code” because there are millions of weights, and “coding directly in weights is kind of hard.””
— Andrej Karpathy, describing neural networks as “Software 2.0” in 2017
Limits Behind the Reported Results
The post does not establish whether the reported outcomes will repeat across independent runs, larger networks, different model families or tasks beyond those tested. It gives one run per setting, and the available source material describes no peer-reviewed evaluation. The 50 Go test runs are reported for that comparison, but do not resolve how stable the method is under other setups.
It is also unclear how much of Astra’s performance depends on the researchers’ task design, the way the weights were presented or the specific prompts and test budgets. The post reports that the model could not distinguish an honest model from a sycophantic one by weights alone, and that its edits did not improve maths in Gemma 3 270M. Those findings caution against treating weight inspection as a general way to identify a model’s behavior or add missing capabilities.
The Qwen score was independently reconstructed by a Hugging Face engineer according to the source article, but the material provided does not describe a broader independent replication of all eight experiments. It also does not establish whether the method can be applied safely to production-scale models or whether edits might have unintended effects outside the tested evaluations.
Replication and Broader Model Tests
The immediate next step is independent replication: researchers can use the released logs to check the reported edits and test whether they recur across runs, model families and tasks. Further work would need to measure more than a single benchmark score, including whether a change fixes the targeted behavior without degrading other capabilities.
The MIT post, as described in the source material, does not set out a specific timetable for a paper, follow-up study or product deployment. Until broader evaluation is available, the work is best treated as an early demonstration rather than a ready-to-use repair system. The authors’ public logs give other researchers material to examine, but the scope and reliability of the method remain open questions.
Key Questions
What is weight-space programming?
It is the approach tested by the MIT researchers in which a language model reads or changes a neural network’s numerical weights directly, rather than relying on gradient descent to adjust them.
What did GPT-6 Astra reportedly repair?
The post reports a successful edit to a car-control network that had failed to reverse in 256 episodes, and a change to Qwen3.5-0.8B that raised its MATH-500 score from 7.9% to 48.1% in the researchers’ test.
Does the research show AI can repair any neural network?
No. The researchers tested a limited set of small networks and tasks. They report that edits did not improve maths performance in Gemma 3 270M and say editing could not add an ability a model lacked.
Has the work been peer reviewed?
According to the source material, the results appeared in a blog post published November 2, 2026, not a peer-reviewed paper, and were not posted on arXiv at the time covered.
What should happen next?
Other researchers need to reproduce the results across independent runs, larger models and additional tasks, and test whether targeted edits affect behavior beyond the benchmark used to measure them.
Source: ThorstenMeyerAI.com
