🔍 Read the full analysis: AI In Weight Space: Reading, Repairing, And Writing Neural Networks on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- 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 infer functions from neural-network weights, repair two kinds of failures and write small networks. The results are described in a November 2 blog post, not a peer-reviewed paper, and the experiments were limited in scale and scope.
Four MIT CSAIL researchers reported that OpenAI’s GPT-6 Astra could identify what small neural networks did, edit some failing models and write new networks by reasoning directly about their weights. The experiments, published in a November 2 blog post with public logs, test an approach the researchers call weight-space programming; they do not establish that the method works broadly or replaces conventional training.
The researchers supplied Astra with network weights—the numerical parameters that determine how a neural network processes inputs—and asked it to infer the networks’ functions or change them. According to the post, Astra identified the game represented by four of four small game-playing networks and inferred the shapes represented by three small image generators. It also found the researchers’ targeted “copying heads” in three tiny language models without running those models.
In one repair test, a car-control network that had recorded zero successes in 256 episodes reached 1,024 successes in 1,024 episodes after Astra changed 3.5% of its weights, the researchers said. In another, a change to 0.016% of Qwen3.5-0.8B’s weights raised its reported MATH-500 score from 7.9% to 48.1%. A Hugging Face engineer later reconstructed the edited Qwen model from the logs and reported a 51.0% score on MATH-500.
The post also describes Astra creating game-playing networks from all-zero weights and writing a 9×9 Go network. The researchers say the newly written models beat the comparison networks used in those tests, including a same-size Go model trained on about 20.5 million games. These results come from small, task-specific experiments; the post does not show that the method can create competitive large-scale systems.
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 Way to Inspect Model Weights
If the approach holds up in broader tests, it could offer developers another way to inspect or make targeted changes to small models without running a conventional training process. Neural-network weights are usually adjusted through optimization algorithms, rather than written or edited directly by a person. The MIT group’s experiments instead ask a language model to infer patterns in those values and propose changes.
That possibility could matter for teams maintaining open models, particularly when investigating a specific failure or auditing a network’s behavior. But the evidence does not support treating weight-space programming as a dependable repair service. The reported successes concern selected tasks and model sizes, while some tests failed. The source article also argues that companies may need to pay closer attention to the provenance of model weights if language models become tools for modifying them.
For users, the findings do not imply an immediate change to the models or services they use. The work is an early demonstration, and the researchers’ public logs allow other people to examine the experiments, not to settle whether the approach will generalize.
Popular right now
De'Longhi Dinamica Plus Fully Automatic Espresso Machine
Philips 2200 Series Fully Automatic Espresso Machine with Classic Milk Frother, 2 Coffee Varieties, Intuitive Touch Display, Ceramic Grinder
Philips Café Aromis Automatic Espresso Machine with LatteGo Pro Milk System
As an affiliate, we earn on qualifying purchases.
From Software 2.0 to Weight Editing
In 2017, computer scientist Andrej Karpathy described neural networks as “Software 2.0”: programs whose behavior is encoded in weights produced through optimization, rather than written directly by programmers. The MIT group’s work takes up the difficulty Karpathy highlighted: a network can contain millions of numerical parameters, making direct manual programming impractical.
The researchers call their alternative weight-space programming. They say Astra received weights as text in some experiments and answered in one reply. In other tests, the researchers ran Python code written by Astra inside a restricted sandbox, with a fixed test budget and no network access. The source says the model was called through OpenAI’s Codex command-line tool with built-in tools disabled. The experiments therefore tested reasoning over weights under defined conditions, not unrestricted access to a computer or the internet.
The source describes the findings as a blog post with complete logs, rather than a peer-reviewed paper or an arXiv publication. Its account covers eight experiments, but reports one run per setting, small networks and limited model coverage. Those details matter when interpreting results that could otherwise sound like evidence for a general method.
““Coding directly in weights is kind of hard (I tried).””
— Andrej Karpathy, describing neural networks in 2017
How Far the Results Extend
The experiments do not establish how well weight-space programming works across larger networks, different model families or repeated runs. The source reports one run per setting and says the tasks used small networks and, in some cases, one model family. The work has not been presented as a peer-reviewed study, and the researchers’ findings remain claims documented in their own post and logs.
Some results also have limits that affect interpretation. The model could not distinguish an honest system from a sycophantic one by examining weights alone, according to the source. Its edits did not improve maths in Gemma 3 270M, and the researchers said editing could not supply a missing ability. The source also notes that a software engineer reproduced part of the game-identification visualization with a short Python script, raising a question about how much of that result depends on the model’s reasoning rather than visible patterns in the data.
It remains unclear whether independent teams can reproduce the full set of results, whether the method scales to larger models, and how consistently Astra’s edits would improve performance without harming other capabilities. A score increase on one benchmark does not by itself show broader improvement.
Independent Tests Will Set the Bar
The immediate next step is independent scrutiny of the released logs and reproduction of the reported experiments. The source says a Hugging Face engineer rebuilt the edited Qwen model and reran MATH-500, obtaining 51.0%, compared with the researchers’ reported 48.1%. That is one check on one model and benchmark, not a broad replication of the method.
Further tests would need to examine repeatability across runs, model families and tasks, as well as whether a targeted edit causes regressions elsewhere. The source does not give a schedule for a peer-reviewed paper or additional experiments. Until more evidence is available, weight-space programming is best understood as a research demonstration of possible model inspection and editing—not a validated, general-purpose way to build or repair neural networks.
Key Questions
What is weight-space programming?
It is the researchers’ term for using a language model to read or change a neural network’s weights directly, rather than relying on a standard training loop to adjust them.
What did GPT-6 Astra do in the experiments?
According to the MIT researchers, Astra identified functions in small networks, repaired selected failures and wrote small game-playing networks from scratch. The post reports both successful tests and limitations.
Has the work been peer-reviewed?
The source describes it as a blog post with public logs, not a peer-reviewed paper or an arXiv paper. Independent reproduction and further evaluation remain important.
Does this mean AI can repair any model?
No. The reported results are limited to selected small networks and tasks. The researchers said edits did not improve one non-looping model’s maths performance and could not add an ability the model lacked.
What is still unknown?
It is not yet clear how reliably the method works across repeated runs, larger models and different model families, or whether edits preserve performance beyond the specific tests reported.
Source: ThorstenMeyerAI.com
