🔍 Read the full analysis: AI And Neural Network Weights: What This Programming Approach Shows on ThorstenMeyerAI.com
Get audio and creator gear 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 infer functions from small neural networks’ weights, repair some models and write others without using gradient descent. The results, published in a November 2 blog post with logs, are preliminary: the tests were limited, some edits failed, and the work has not been peer-reviewed.
Four MIT CSAIL researchers say OpenAI’s GPT-6 Astra could identify, repair and create small neural networks by reasoning directly about their numerical weights, without using a training loop or gradient descent. Their November 2 blog post describes eight experiments, including a repaired car-control model and a small Go network written from zero; the findings are logged publicly but have not been peer-reviewed.
The researchers call the approach weight-space programming: a language model receives a neural network’s weights and tries to infer what they do or alter them directly. In some tests Astra received weights as plain text and replied once; in others, researchers ran code it wrote in a sandbox with a fixed test budget. The post says the model had no shell, web search or file-editing tools, and that the team checked 1,125 recorded turns for tool use.
In the reported reading tests, Astra identified the games behind four of four small game-playing networks and the shapes represented by three small image generators. It also found every copying, or induction, head in three tiny language models without running those models. The researchers say the networks were deliberately small; in the game test, the four networks shared the same architecture and input-output structure.
For repairs, the post reports that Astra changed 3.5% of a car-control network’s weights. The network had succeeded in zero of 256 episodes before the edit and 1,024 of 1,024 afterward. In a separate test, an edit affecting 0.016% of the weights in Qwen3.5-0.8B raised its reported MATH-500 score from 7.9% to 48.1%. The post also describes newly written game networks and a 9-by-9 Go network; it says the Go model 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
Direct Edits Could Change Model Checks
If the results hold up, weight-space programming could offer researchers another way to inspect or make targeted changes to small models, rather than relying only on conventional retraining. The reported car and Qwen results illustrate why the approach attracts attention: the edits were presented as specific interventions, with measured performance before and after.
That does not establish that language models can generally replace training algorithms or safely repair deployed systems. The experiments were limited in scale and scope, and the authors also report failures. For organizations using open models, the work points to both a potential new inspection method and a need to understand the origin and integrity of model weights. The authors’ results do not change how ordinary users’ AI systems work today.
Products Worth Considering
Samsung Galaxy Tab A11+ 11-Inch Tablet, 6GB RAM, 128GB Storage, Expandable, Gray
Bose TV Speaker (2nd Gen) Compact Soundbar
Bose SoundLink Flex Bluetooth Speaker (2nd Gen) – Portable Outdoor Speaker with Hi‑Fi Audio, Waterproof and Dustproof, USB‑C, Up to 12 Hours
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 by optimization rather than written directly by people. Neural-network training typically adjusts those weights repeatedly against data. The MIT group’s experiment asks whether a language model can instead treat the trained numbers as something it can inspect and edit.
The researchers are Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab; the post says Takehi and Teoh contributed equally. Their account is a research-group blog post, not a peer-reviewed paper, and the source says it was not on arXiv. It presents logs and results, while acknowledging one run per setting, small networks, one model family per task and a single research group.
“not through an optimizer, and in particular not with gradient descent”
— The MIT CSAIL researchers, in their November 2 blog post
Small Tests Leave Open Questions
The findings remain preliminary. The available account describes one run per setting and small networks, and does not establish whether the same results will hold across larger models, other model families or repeated independent trials. The blog post is not peer-reviewed, and the source does not report an independent replication of the full set of experiments.
Some claims also have limits within the tests. Astra could not distinguish an honest model from a sycophantic one using weights alone, and its edits did not improve the maths score of Gemma 3 270M. A separate Hugging Face engineer reportedly rebuilt the edited Qwen model from the logs and scored 51.0% on MATH-500, but the source does not provide further details on that replication or establish how broadly the result generalizes.
Replication Is the Next Test
The immediate next step is independent replication across more model families, tasks and repeated trials, alongside scrutiny of the released logs and methods. Further tests would need to establish whether direct weight edits are reliable and whether the measured gains persist beyond the small models and fixed evaluations described in the post.
The source does not announce a publication, a deployment plan or a timeline for additional results. Until more evidence is available, the work is best treated as an early demonstration of a possible research technique, not a proven replacement for conventional training or model evaluation.
Key Questions
What is weight-space programming?
It is the researchers’ term for using a language model to read or edit another neural network’s weights directly, rather than changing them through a training optimizer such as gradient descent.
What did Astra reportedly repair?
The blog post describes a car-control network that went from zero successful episodes out of 256 to 1,024 out of 1,024 after Astra changed 3.5% of its weights. It also reports a higher MATH-500 score after a small edit to Qwen3.5-0.8B.
Has the work been independently verified?
The source says the work was published as a blog post, not a peer-reviewed paper. It reports that a Hugging Face engineer rebuilt the Qwen edit from logs and obtained 51.0% on MATH-500, but the full set of experiments still needs broader independent testing.
Does this mean AI can reliably fix any model?
No. The researchers report that edits did not improve maths performance in Gemma 3 270M and say editing could not add an ability a model lacked. The experiments do not establish that the approach works reliably across models or tasks.
Source: ThorstenMeyerAI.com
