🔍 Read the full analysis: How Model Numbers Let AI Read, Repair, And Write Neural Networks on ThorstenMeyerAI.com
Get the little things that make your day 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 could identify functions in small neural networks, repair some failures by editing weights directly and write new networks. The results come from a November 2 blog post with released logs, not a peer-reviewed paper; the experiments were limited, and some edits did not help.
Four researchers at MIT’s Computer Science and Artificial Intelligence Laboratory say OpenAI’s GPT-6 Astra identified, repaired and wrote small neural networks by working directly with their numerical weights, rather than using the usual training process. Their November 2 blog post reports notable results, including a repair to a car-control network and a large maths-score increase for one small language model, but the work is an experimental blog report, not a peer-reviewed paper.
The researchers call the approach weight-space programming: a general-purpose language model receives another network’s weights and reasons about what they encode, then may suggest or make changes. In conventional machine learning, weights are adjusted through training, often with gradient descent. The team says Astra worked without an optimizer or gradient descent in these experiments.
In one test, Astra identified all four of four small game-playing networks from their weights alone. The networks had the same structure and inputs and outputs, but had been trained for different games. In another test, the model identified the shape represented by three of three small image generators. The researchers also report that Astra found every induction, or copying, head in three tiny language models without running those networks.
The repair results were mixed. A car-control network that had failed to reverse succeeded in 1,024 of 1,024 test episodes after Astra changed 3.5% of its weights; it had recorded zero successes in 256 episodes before the edit. For Qwen3.5-0.8B, an edit affecting 0.016% of weights raised its MATH-500 score from 7.9% to 48.1%, according to the post. A Hugging Face engineer later rebuilt the edited model from the released logs and reported a score of 51.0%.
The team also says Astra created small game-playing networks from all-zero weights and wrote a 9-by-9 Go network with about 1.5 million weights. In 50 test runs, the Go network beat a same-size comparison network trained on about 20.5 million games. These findings describe the specific tests reported; they do not establish that the method works reliably across larger models or other tasks.
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 approach can be reproduced and extended, it could offer researchers and developers a different way to inspect or make targeted changes to small open models. Instead of running a full training process, a model might identify a particular pattern in another network and propose a limited edit. That could be useful for diagnosing certain failures, though the experiments do not show that such edits are dependable in general.
The work also points to a potential governance concern: if models can interpret or alter weights directly, the origin and integrity of those weights may become more relevant to audits. The researchers’ tests do not show that this creates an immediate risk to users, and the source article says nothing changes for users today. For now, the main significance is a research possibility, paired with a need for independent testing.
Products Worth Considering
Therabody Theragun Prime 6th Generation Deep Tissue Massage Gun, Black
TheraGun Therabody PRO Plus 6-in-1 Deep Tissue Massage Gun, Black
Levoit Vital 100S-P Air Purifier for Home and Pets
As an affiliate, we earn on qualifying purchases.
From Software 2.0 to Weight-Space Tests
In 2017, computer scientist Andrej Karpathy described neural networks as “Software 2.0”: systems whose behavior is encoded in learned weights rather than conventional hand-written instructions. He noted that directly writing code into millions of weights would be difficult for a person. The MIT group’s work asks whether a language model can reason about those numbers as a form of program.
The researchers—Rikiya Takehi and Jayden Teoh, who shared equal contribution, along with Alex Zhang and Omar Khattab—used Astra through OpenAI’s Codex command-line tool. According to the released logs, built-in tools such as shell access, web search and file editing were disabled. Some tests gave Astra weights as text and required a single response; other tests let it write Python code that ran in a sandbox without network access and under a fixed test budget. The team says it checked all 1,125 turns in the release and found them tool-free.
The post’s authors frame the work as a shift from human-designed optimizers toward language models that reason about network weights. That is the researchers’ interpretation of the direction of the field, not a demonstrated replacement for standard training methods. The experiments involved small networks and a limited set of tasks.
“Software 2.0”
— Andrej Karpathy, describing neural networks in 2017
How Far the Results Generalize
The results have not yet been presented as a peer-reviewed paper or posted on arXiv, according to the source article. The reported work comes from one research group, uses small networks and covers a limited number of tasks. The article also says there was one run per setting, limiting what can be concluded about consistency and variation across repeated trials.
It remains unclear whether the approach will work on larger models, other model families or failures that are harder to isolate. The post reports that Astra could not distinguish an honest model from a sycophantic one by looking at weights alone, and that edits did not improve a model without the targeted looping problem. The source also notes that an independent engineer reproduced the edited Qwen result from logs, but that single replication does not settle how broadly the findings apply.
Replication and Broader Model Tests
The next useful step is independent replication across repeated runs, model families and tasks, with clear comparisons against standard training and debugging methods. Researchers would also need to test how reliably the system identifies why a network behaves as it does, rather than merely recognizing patterns that correlate with a task.
The MIT team’s released logs provide material for checking the reported experiments. The source article does not identify a peer-reviewed publication, a scheduled follow-up study or a product release. Until further evidence is available, weight-space programming remains an early research direction rather than a demonstrated tool for routine model repair.
Key Questions
What is weight-space programming?
It is the method described by the MIT researchers in which a language model reads a neural network’s weights and reasons about what they do, potentially changing them directly rather than relying on gradient-descent training.
What did GPT-6 Astra reportedly do?
In the reported tests, Astra identified four small game networks, identified shapes in three image generators, found copying heads in three tiny language models, repaired some network failures and wrote small networks, including a 9-by-9 Go network.
Did the method improve every model it edited?
No. The source reports that Astra’s edits did not improve maths performance on Gemma 3 270M. The researchers said editing could not add an ability that a model lacked.
How strong is the evidence so far?
The findings are reported in a blog post with public logs, not a peer-reviewed paper, according to the source article. The tests used small networks, one research group and one run per setting, so broader reliability remains unestablished.
Does this change how people use AI models now?
The source article says there is no immediate change for users. The work is experimental, and no routine product or deployment based on the method is reported.
Source: ThorstenMeyerAI.com
