🔍 Read the full analysis: AI Model Editing Through Weights: What The Experiment Shows 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
An MIT CSAIL research team says OpenAI’s GPT-6 Astra interpreted, repaired and wrote neural-network weights directly in eight experiments, without gradient descent. Results included a repaired car controller and a large MATH-500 score gain after a small Qwen edit, but the work is a blog post with limited tests, and the model could not improve every task.
Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory say OpenAI’s GPT-6 Astra could inspect and change neural-network weights directly in a set of eight experiments, without using a training loop or gradient descent. The reported tests include a car-control network that began reversing successfully after an edit and a small language model whose MATH-500 score rose from 7.9% to 48.1%; the results are documented in a public blog post, not a peer-reviewed paper.
The researchers call the approach weight-space programming: a language model receives the numerical weights of another neural network and reasons about what they encode, then proposes edits or new weights. The MIT team — Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab — says the method avoids the usual process of repeatedly adjusting weights through an optimizer. In some tests Astra received weights as plain text and answered in one reply. In others, researcher-written sandbox code ran programs generated by Astra under fixed testing limits.
The reported repair examples had sharply different scales. A small car-control network that had recorded zero successful episodes out of 256 achieved 1,024 successes in 1,024 episodes after Astra changed 3.5% of its weights, according to the post. For Qwen3.5-0.8B, an edit affecting 0.016% of weights was followed by a MATH-500 score increase from 7.9% to 48.1%. The source says a Hugging Face engineer rebuilt the edited Qwen model from the logs and measured 51.0% on the same benchmark.
The team also reports that Astra identified the games represented by four small game-playing networks and inferred the shapes produced by three small image generators. It wrote small game networks from zero weights and generated a 9-by-9 Go network that, in 50 test runs, beat a same-size comparison network trained on about 20.5 million games. These are reported experimental results; the post also records failed cases, including an inability to distinguish an honest model from a sycophantic one based on weights alone.
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 Route to Model Repair
If the method proves reliable beyond these demonstrations, it could offer researchers a way to inspect or make targeted changes to small models without rerunning expensive training. The reported Qwen result is particularly striking because a very small fraction of weights changed alongside a substantial benchmark-score increase. But that result does not establish that the same technique will work across model families, tasks or production systems.
The approach also shifts attention to who can read and modify model weights. For developers, direct editing could become another audit or repair tool; it could also make the provenance and integrity of a model’s weights more consequential. The source article says there is no immediate change for ordinary users, and the experiments do not show that this method is ready to replace conventional training or evaluation.
Products Worth Considering
Samsung Galaxy Tab A11+ 11-Inch Tablet, 6GB RAM, 128GB Storage, Expandable, Gray
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Black
LG UltraGear 32GX850A-B 32-inch 4K UHD Glossy OLED Gaming Monitor
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 represented in learned weights rather than written directly by people. The MIT experiments revisit that idea by asking whether a language model can treat another network’s weights as something it can interpret and edit. The researchers’ post, published on November 2, 2026, presents eight experiments spanning network interpretation, repair and creation.
The team says Astra was run through OpenAI’s Codex command-line tool with capabilities such as shell access, web search and file editing disabled. For experiments involving generated code, the researchers ran it in a sandbox without network access and used fixed test budgets. The post says the team reviewed all 1,125 turns in the release and found that the model did not use tools. These controls describe the reported setup; they do not by themselves establish how well the results generalize.
“Weight-space programming”
— The MIT research team, in its November 2 blog post
How Far the Results Generalize
The evidence remains preliminary. The source describes a blog post with public logs, not a peer-reviewed paper or an arXiv preprint. It reports one run per setting, small networks, a single model family for each task and one research group. Those constraints make it difficult to tell whether the strongest results will repeat under independent testing or scale to larger, more varied systems.
The article also reports that Astra could not distinguish an honest model from a sycophantic one using weights alone, and that edits did not improve maths in Gemma 3 270M, a model that did not loop. The cause and scope of those failures are not established in the supplied material. Nor does the reported MATH-500 improvement show whether the edit affects other abilities, produces unwanted side effects or holds up across broader evaluations.
Independent Tests Are the Next Check
The immediate next step is independent reproduction of the experiments, particularly the Qwen edit and the car-controller repair. The source notes that a Hugging Face engineer reconstructed the edited Qwen model from released logs and obtained a 51.0% MATH-500 score, a second reported measurement that still does not settle how broadly the approach works.
Further evidence would need to test more models and tasks, repeat runs, and measure changes beyond the benchmark used to guide an edit. The source provides no publication schedule or announced follow-up study. Until additional results are available, weight-space programming is best described as a promising but limited set of demonstrations, not an established general-purpose way to repair neural networks.
Key Questions
What is weight-space programming?
It is the researchers’ term for having a language model inspect and edit another neural network’s weights directly, rather than relying on a training loop or gradient descent.
What changed in the Qwen experiment?
According to the blog post, an edit to 0.016% of Qwen3.5-0.8B’s weights was followed by an increase in its MATH-500 score from 7.9% to 48.1%. A Hugging Face engineer later reported 51.0% after rebuilding the edited model from the logs.
Was Astra able to fix every model?
No. The source says its edits did not improve maths in Gemma 3 270M, which was not looping, and it could not identify an honest model versus a sycophantic one from weights alone.
Has the research been peer-reviewed?
The supplied source describes the work as a blog post with public logs, not a peer-reviewed paper or an arXiv preprint.
Does this change how people use AI models now?
The source reports no immediate change for users. The experiments are limited demonstrations, and further testing is needed to establish whether the method works reliably across models and tasks.
Source: ThorstenMeyerAI.com
