🔍 Read the full analysis: Weight-Space Programming: A New Way To Work With 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 inferred what small neural networks did, repaired some failures and wrote networks by editing their weights directly. Their November 2 blog post logs eight experiments, but the work has not been peer reviewed and relies on limited tests; some tasks failed or showed no benefit.
Four researchers at MIT’s Computer Science and Artificial Intelligence Laboratory say they used OpenAI’s GPT-6 Astra to inspect, repair and create small neural networks by working directly with their numerical weights, rather than using a training loop. The team’s November 2 blog post reports successes on several tests, including a car-control system and a small language model, but the results are early, limited experiments rather than peer-reviewed findings.
The researchers call the approach weight-space programming: a language model receives the numbers that make up another network and reasons about how to change them. The team says it did not use gradient descent or an optimizer for these edits. In some tests, Astra received weights in text and had to respond directly; in others, researchers ran code it wrote in a restricted sandbox.
In one reported repair, 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 another, an edit affecting 0.016% of the weights in Qwen3.5-0.8B coincided with an increase on the MATH-500 benchmark from 7.9% to 48.1%. The source says a Hugging Face engineer rebuilt the edited model from the logs and obtained 51.0% on a rerun.
The post also describes Astra identifying the games behind four small game-playing networks and the shapes represented by three image-generating networks. It wrote small game networks from all-zero weights and produced a 9-by-9 Go network that beat a same-size comparison model in 50 test runs, according to the researchers. The post details eight experiments overall, including failures: Astra could not distinguish an honest model from a sycophantic one using weights alone, and edits did not improve maths performance in Gemma 3 270M, which was not exhibiting the targeted looping behavior.
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 Possible Route to Targeted Model Edits
If the approach holds up in broader testing, it could offer researchers a way to inspect or modify a model without retraining it from scratch. The reported Qwen result is notable because the edit touched a small fraction of the weights, while the car experiment suggests a language model may sometimes identify a functional fault from the network’s parameters and correct it.
Those outcomes do not establish that this is a dependable way to repair deployed AI systems. The tests involve small networks and narrow tasks, and the post reports one run per setting. For model developers, the work raises research questions about auditing, reproducibility and the origin of model weights. For most users, it changes no current product or practice.
Products Worth Considering
Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Black
Amazon Kindle Paperwhite Signature Edition, 32GB (2024), Metallic Black
Logitech MX Brio Ultra HD 4K Webcam, Graphite
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”: systems whose behavior is encoded in weights produced by optimization rather than written directly by a programmer. Those weights can be difficult for people to interpret, particularly in large models with millions or billions of parameters.
The MIT team’s experiments test whether a general-purpose language model can reason about those parameters as a form of code. According to the released logs, Astra was run through OpenAI’s Codex command-line tool with built-in tool capabilities disabled. The researchers say they reviewed 1,125 turns and found no tool use. The source account also notes that a software engineer reproduced one visual analysis of a network’s weights with a short Python script, a reminder that interpreting a pattern is not the same as proving the model’s internal reasoning.
““not through an optimizer, and in particular not with gradient descent””
— The MIT research team, in its November 2 blog post
How Far the Results Generalize
The available evidence is a research blog post with public logs, not a peer-reviewed paper or an arXiv preprint, according to the source material. The experiments use small networks, a single research group and, for each setting, one run. The results do not show whether the method works reliably across larger models, different architectures or tasks outside those tested.
It is also unclear how much of the reported performance would persist under repeated independent evaluation. The Qwen rerun by a Hugging Face engineer offers one replication of that result, but the source does not report independent replications of the other experiments. Nor does the work show that Astra can consistently identify why a model behaves as it does; the failed sycophancy test points to limits in what can be inferred from weights alone.
Independent Tests and Broader Models
The immediate next step is independent replication across repeated runs, tasks and model architectures, with the released logs providing material for other researchers to examine. Further evaluations would need to test whether weight edits remain effective after changes to prompts, data or evaluation methods, and whether they introduce unwanted behavior elsewhere.
The source material does not name a planned paper, release date or follow-up experiment. Until those details emerge and the findings receive broader testing, weight-space programming remains an early research approach, not a validated replacement for conventional training or model maintenance.
Key Questions
What is weight-space programming?
It is the approach tested by the MIT team in which a language model reads or changes another neural network’s weights—the numerical parameters that shape its behavior—rather than relying on gradient descent to train it.
What did GPT-6 Astra reportedly repair?
The researchers say Astra edited a car-control network that had failed to reverse in 256 recorded episodes. After the edit, it succeeded in 1,024 of 1,024 episodes. They also report a maths benchmark improvement after a small edit to Qwen3.5-0.8B.
Are the findings peer reviewed?
Not according to the source article. The results were presented in a blog post with public logs; the source says the work was not peer reviewed and was not on arXiv at the time of publication.
Did the approach work in every experiment?
No. The post reports that Astra could not distinguish an honest model from a sycophantic one based on weights alone, and that its edits did not improve maths performance in Gemma 3 270M. The researchers say editing could not supply an ability the model lacked.
Does this change how people use AI models today?
No immediate change is reported. The experiments concern small research models, and the method has not been established as a reliable tool for modifying production systems.
Source: ThorstenMeyerAI.com
