🔍 Read the full analysis: Weight-Space Programming Offers A Different Way To Edit Neural Networks on ThorstenMeyerAI.com
Get business pricing on tech for your team
- 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 patterns in small neural networks, repair some failures and write networks by directly reading and changing their weights. The results are documented in a public blog post, but the experiments are limited and have not been peer reviewed.
Four researchers at MIT’s Computer Science and Artificial Intelligence Laboratory reported that OpenAI’s GPT-6 Astra could read, repair and create small neural networks by working directly with their numerical weights. The experiments, described in a November 2 blog post, test “weight-space programming,” an approach that could offer a different way to inspect or modify models, though the results remain preliminary.
The researchers gave Astra the numerical weights of networks and asked it to infer what they did, fix selected failures or produce new networks. In the reported setup, Astra did not use a conventional training loop or gradient descent. Some tasks required it to respond from plain-text weights in a single reply; in others, researchers ran code it wrote in a restricted sandbox. The source article says the released logs record 1,125 turns and that the team checked them for tool use.
Among the reported repairs, a car-control network that had recorded zero successful reversals in 256 episodes reached 1,024 successes in 1,024 episodes after Astra changed 3.5% of its weights. In a separate experiment, changing 0.016% of the weights in Qwen3.5-0.8B raised its reported MATH-500 score from 7.9% to 48.1%. A Hugging Face engineer later rebuilt that edited model from the logs and reported a score of 51.0%, according to the source article.
The post also describes Astra identifying the games behind four small networks and the shapes represented by three image generators, as well as finding “copying heads” in three tiny language models. It reportedly wrote a 1.5-million-weight Go network from scratch that beat a same-size comparison network trained on about 20.5 million games over 50 test runs. These are results reported by the researchers, not independent evidence that the method works broadly.
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 Model Repair
If the approach holds up in broader tests, directly editing weights could give researchers and model developers another way to investigate small networks and make targeted changes. It may be useful where a specific behavior can be linked to a limited part of a model, rather than requiring a full retraining run. The reported Qwen result is an example of a narrow edit associated with a higher score on one maths benchmark.
The findings also suggest a possible audit use: a model might help interpret another network’s internal numerical structure. But the experiments do not establish that the method can safely or reliably modify production systems. The source article notes that Astra could not distinguish an honest model from a sycophantic one using weights alone, and that edits did not improve maths performance in Gemma 3 270M, a model that did not have the looping problem. The authors’ reported limit is direct: editing “could not add an ability the model lacked.”
For users, the work does not change how current AI products operate. For organizations experimenting with open models, it is an early research result that merits replication, not a ready-made repair process. The provenance and review of model weights would also matter if systems can be altered directly by another model.
Products Worth Considering
Apple MacBook Air 13.6-inch with M5 Chip, 16GB Memory, 512GB SSD, Sky Blue
Bose QuietComfort Wireless Noise Cancelling Headphones, Black
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black
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 shaped by optimization rather than written directly by people. He also noted that coding in weights is difficult because networks can contain millions of them. The MIT researchers’ term, weight-space programming, refers to asking a language model to reason about those numbers and change them directly.
The MIT team named in the source is Rikiya Takehi and Jayden Teoh, who contributed equally, along with Alex Zhang and Omar Khattab. Their post presents eight experiments and makes logs publicly available. According to the source article, the work is a blog post, not a peer-reviewed paper, and it had not been posted on arXiv at the time of publication. The reported tests use small networks, one research group and, in the described settings, one run per setting—limits that matter when judging how general the results may be.
“Software 2.0”
— Andrej Karpathy, describing neural networks in 2017
How Far the Results Extend
The reported outcomes have not been established through peer review or broad independent replication. The experiments involve small networks and limited settings; the source article says there was one run per setting and one model family per task. It is not yet clear whether the reported performance gains will hold across different architectures, tasks, prompts or repeated trials.
The logs and the reported Qwen re-run offer ways to inspect parts of the work, but they do not answer every question about reproducibility or practical use. The source material does not establish whether weight-space programming can scale to large commercial models, reliably prevent harmful edits or outperform established training and fine-tuning methods across a representative range of tasks. Those questions remain open.
Replication and Broader Testing
The next meaningful test is independent replication across more models and tasks, with repeated runs and clear comparisons against conventional fine-tuning or other repair methods. Researchers will also need to test whether the approach can identify when an edit has unintended effects, rather than judging success only by a target benchmark or task.
The MIT post and its released logs provide material for scrutiny, but the source does not give a schedule for a peer-reviewed paper or further experiments. For now, the findings should be read as an early demonstration: direct weight editing worked in several reported small-model tests, while its reliability, scale and suitability for deployed systems remain unsettled.
Key Questions
What is weight-space programming?
It is the approach described by the MIT researchers in which a language model reads or changes another neural network’s numerical weights directly, rather than relying on a conventional optimizer and gradient descent for the task.
What did GPT-6 Astra reportedly repair?
In one experiment, Astra edited a car-control network that had failed to reverse successfully, and the researchers reported 1,024 successes in 1,024 episodes afterward. In another, an edit to Qwen3.5-0.8B was associated with a higher MATH-500 score.
Has the work been peer reviewed?
Not according to the source material. It describes the findings as a public blog post published November 2, 2026, rather than a peer-reviewed paper, and says the work was not on arXiv at the time.
Does the method work for every model or task?
No such general result has been established. The researchers reported failures and limits, including no maths improvement when editing Gemma 3 270M, and said editing could not add an ability the model lacked.
Source: ThorstenMeyerAI.com
