🔍 Read the full analysis: Can AI Work With Neural Networks Directly Through Their Weights? on ThorstenMeyerAI.com
Get hardware and tech essentials 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 neural-network weights, repair some small models and write new ones without gradient descent. The results come from a detailed public blog post, not a peer-reviewed paper, and include clear failures as well as successes.
Four MIT CSAIL researchers report that OpenAI’s GPT-6 Astra could read and edit the numerical weights of small neural networks to identify what they did, repair some failures and write new networks. The experiments, published in a November 2 blog post, tested a method the researchers call weight-space programming, which bypasses the usual training loop and gradient descent; the results are promising but remain limited to small, task-specific tests.
In the team’s setup, Astra received network weights and was asked to reason about them as a program. In some tests it saw weights as plain text and had to respond in one turn. In others, it wrote Python code that researchers ran in a sandbox with no network access and a fixed test budget. The researchers say they disabled Codex tools such as shell access, web search and file editing, and checked the released record of 1,125 turns.
For the reading tests, Astra identified the games behind four of four small networks trained for Breakout, Pong, Pac-Man and Freeway. The networks shared the same structure, and the model was not given a menu of possible answers. It also identified the intended shapes in three of three small image generators and found every copying, or induction, head in three tiny language models, according to the post.
The reported repair tests had mixed outcomes. 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. An edit to 0.016% of Qwen3.5-0.8B’s weights was associated with a MATH-500 score increase from 7.9% to 48.1%. The source says a Hugging Face engineer reconstructed the edited Qwen model from the logs and recorded 51.0% on a rerun.
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 results hold up in broader testing, direct analysis of weights could offer developers another way to inspect or make targeted changes to small models. Rather than retraining a network across many examples, a language model might identify a particular behavior in its weights and alter a small portion. That could be useful for debugging or auditing some open models, though these experiments do not establish that the approach is reliable for deployed systems.
The reported gains also need to be read alongside the failures. Astra’s edits did not improve maths performance in Gemma 3 270M, which the researchers described as not exhibiting the looping problem seen in Qwen. The team said editing could not add an ability that a model lacked. The work does not show that a language model can generally understand or safely modify arbitrary neural networks.
For most users, the findings do not change how AI products work today. For developers, the prospect of editing weights directly makes model provenance and validation relevant: a model’s behavior could be affected by changes that are difficult to review using ordinary source-code tools. Any practical use would need checks that the edit produces the intended behavior and does not introduce other problems.
Products Worth Considering
Apple MacBook Pro 14-inch with M5 Max Chip, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
Apple MacBook Pro 14.2-inch with M5 Max, 18-core CPU, 32-core GPU, 36GB Memory, 2TB SSD, Wi-Fi 7, Silver
Logitech MX Brio Ultra HD 4K Webcam, Graphite
As an affiliate, we earn on qualifying purchases.
From Trained Weights to Editable Programs
Neural-network weights are numerical values that shape how a model processes inputs. In standard training, an optimizer adjusts those values repeatedly using data and a loss signal. The resulting weights can be difficult for people to interpret directly, even when the network’s overall behavior is measurable.
The researchers frame their method as a shift from optimizing weights through a conventional training algorithm to having a language model reason about and edit them. The idea echoes Andrej Karpathy’s 2017 description of neural networks as “Software 2.0”: programs represented in weights rather than written line by line. The MIT group’s experiments put that idea to tests involving reading, repair and creation.
The post says Astra also wrote small game-playing networks from all-zero weights. Those networks outperformed the researchers’ comparison networks trained for 100 million steps. In a separate test, a newly written 9-by-9 Go network beat a same-size network trained on about 20.5 million games across 50 test runs. These are results reported by the researchers, not an independent evaluation across many models and tasks.
““not through an optimizer, and in particular not with gradient descent””
— The MIT CSAIL researchers, in their November 2 blog post
How Far the Tests Can Be Generalized
The work is presented in a blog post with released logs, not a peer-reviewed paper or an arXiv preprint, according to the source material. The experiments were conducted by one research group, used small networks, and generally involved one model family per task. The source also says there was one run per setting, leaving open how consistent the results would be across repeated trials.
Some results may also be easier to interpret than others. Software engineer Sean Goedecke reportedly reproduced a visualization used to explain the Breakout weights with three lines of Python, raising a question about how much of the task required sophisticated reasoning. The blog’s results do not establish whether Astra’s edits would remain effective across different architectures, larger models or real-world deployments. It is also unclear from the supplied material how the team’s tests compare with other methods under a shared, independent evaluation.
Independent Tests and Broader Replication
The immediate next step is independent replication of the reported experiments, including repeated trials, additional model families and tests of whether edits preserve unrelated capabilities. The public logs and the reported reconstruction of the Qwen model provide material other researchers can examine, but broader testing is needed before the method’s reliability can be judged.
The researchers’ blog post is the current source for the claims described here. The supplied material does not identify a planned peer-reviewed publication or a further benchmark release. Until additional evaluations are available, weight-space programming should be treated as an early research approach rather than a demonstrated general-purpose way to repair or build AI systems.
Key Questions
What is weight-space programming?
It is the researchers’ term for using a language model to inspect or change another neural network’s weights directly, rather than relying on a standard training optimizer and gradient descent.
What did GPT-6 Astra reportedly accomplish?
In the researchers’ tests, Astra identified small networks’ functions, repaired a car-control network and a looping issue in Qwen, and wrote small game-playing networks. The results varied by task, and the source describes failures as well as successes.
Does this show that AI can repair any neural network?
No. The experiments covered small, specific networks and do not show that the method works across arbitrary models. The researchers reported that edits did not improve maths performance in Gemma 3 270M and said editing could not supply an ability the model lacked.
Has the research been peer-reviewed?
According to the source material, the results appeared in a public blog post with logs, not in a peer-reviewed paper or an arXiv preprint. Independent replication and broader evaluation remain outstanding.
Source: ThorstenMeyerAI.com
