🔍 Read the full analysis: AI In Weight Space: A Practical Look At Reading And Repairing Models on ThorstenMeyerAI.com
Get business pricing on tech for your team
- 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 some small models and write networks for several tasks. The findings are documented in a public blog post, not a peer-reviewed paper, and the tests were limited in scale and scope.
Four MIT CSAIL researchers reported November 2 that OpenAI’s GPT-6 Astra could inspect, repair and write small neural networks by directly reasoning about their weights, rather than using a training process such as gradient descent. The experiments suggest a possible way to diagnose or make targeted changes to some models, but they are limited tests documented in a public blog post, not a peer-reviewed study.
The researchers call the approach weight-space programming: a language model receives another network’s numerical weights and tries to infer what they do or change them. In the released experiments, Astra identified the games behind four small game-playing networks and the broad shapes produced by three small image generators. The team also reported that Astra found copying heads—components associated with repeated-token prediction—in three tiny language models without running those models.
Two repair tests produced different outcomes. A car-control network that had recorded zero successful reversals in 256 episodes completed 1,024 of 1,024 after Astra changed 3.5% of its weights, according to the post. In Qwen3.5-0.8B, an edit affecting 0.016% of weights raised the reported MATH-500 score from 7.9% to 48.1%. A Hugging Face engineer later rebuilt the edited Qwen model from the released logs and reported a 51.0% score, the source article says.
Astra also generated game-playing networks from all-zero weights and produced a 9-by-9 Go network. The researchers say the Go model beat a same-size comparison network trained on about 20.5 million games across 50 test runs. These results do not establish that direct editing can replace conventional training: the post describes small tasks and models, with one run per setting, and reports important failures as well as successes.
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
Potential for Targeted Model Repairs
If the reported results hold up under broader testing, direct weight editing could offer developers another way to inspect and patch small models without retraining them from scratch. The Qwen result is a striking example of a small numerical change accompanying a higher benchmark score, while the car-control experiment suggests that edits may sometimes address a specific failure.
But the work is not evidence that a language model can reliably understand or repair arbitrary AI systems. The source reports that Astra could not distinguish an honest model from a sycophantic one by inspecting weights alone. It also did not improve maths performance in Gemma 3 270M, which was not exhibiting the looping behavior targeted in the Qwen experiment. The researchers’ stated limitation is that editing could not add an ability a model lacked.
For organizations using open models, the work points to potential audit and maintenance tools, but also to the importance of tracking where model weights come from and what changes have been made. For ordinary users, the experiments do not change the behavior or capabilities of deployed products today.
Products Worth Considering
LG UltraGear 27GX704A-B 27-inch QHD OLED Gaming Monitor
Bose TV Speaker (2nd Gen) Compact Soundbar
Samsung 34-Inch Odyssey G5 Ultra-Wide Curved Gaming Monitor
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”: programs whose behavior is encoded in weights optimized from data rather than written directly by people. The MIT team’s proposal takes that idea a step further by asking a language model to work with those numbers directly. The source article says the experiments used GPT-6 Astra through OpenAI’s Codex command-line tool, with tool capabilities disabled for the model.
In some tests, Astra received weights as plain text and had to answer in one response. In others, researcher-written sandbox code ran Python generated by Astra, without network access and within a fixed test budget. The team says it reviewed all 1,125 turns in the released logs and found they were tool-free. This distinction matters: the model’s role was to reason about weights, while the researchers’ setup controlled execution and evaluation.
The reported work is available as a blog post, with logs, rather than as a peer-reviewed paper or an arXiv preprint, according to the source material. Its claims should be read as early experimental results, not as an established 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 scope and reliability of the findings remain open questions. The source describes one run per setting, small networks, a single model family for each task and one research group. The reported benchmark changes therefore need replication across more runs, models and independent teams before they can show how dependable the technique is.
The public record also has limits: the source identifies the work as a blog post rather than peer-reviewed research, and does not report a formal paper review. It is not clear whether Astra’s performance will generalize to larger, more complex models, or whether the same approach can diagnose less visible problems. The inability to detect sycophancy from weights alone and the lack of improvement on Gemma 3 270M are direct cautions, not minor exceptions.
The source provides numerical results and says logs are public, but the reported outcomes do not by themselves establish that every change is safe, stable across tasks or suitable for deployment. No evidence in the supplied material shows that companies have adopted the method in production.
Replication and Broader Testing
The immediate next step is independent replication of the released experiments, including the Qwen edit and the car-control repair. Researchers would need to test repeated runs, additional model families and tasks, and whether a repair improves its intended measure without degrading other capabilities. The public logs provide material for such checks, but the source gives no schedule for a peer-reviewed paper or a larger follow-up study.
Until those tests are available, the results are best treated as an early demonstration of a research direction. Developers and readers can distinguish what has been reported—specific successes and failures in small experiments—from the broader claim that language models may one day routinely program neural networks through their weights.
Key Questions
What is weight-space programming?
It is the researchers’ term for having a language model read or change another neural network’s weights directly, by reasoning about the numerical parameters rather than using gradient descent.
What did Astra repair?
The researchers reported a successful change to a car-control network that had failed to reverse, and an edit to Qwen3.5-0.8B that raised its reported MATH-500 score. These are specific experimental results, not proof that the method can repair any model.
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.
Can the method add new abilities to a model?
The researchers said their editing could not add an ability a model lacked. In one test, edits did not improve maths performance in Gemma 3 270M, which was not looping.
Does this change how AI tools work for users now?
No deployment or product change is described in the source. The results are early research experiments; the method’s reliability and usefulness on larger systems remain unestablished.
Source: ThorstenMeyerAI.com
