🔍 Read the full analysis: How AI Used Neural Network Weights To Read, Repair, And Write 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
An MIT CSAIL team reports that OpenAI’s GPT-6 Astra identified functions in small neural networks, repaired two specific model problems and wrote new networks by directly editing weights. The results are documented in a blog post, not a peer-reviewed paper; experiments were limited, and some tests failed.
MIT CSAIL researchers say OpenAI’s GPT-6 Astra read, repaired and wrote small neural networks by working directly with their numerical weights, without training them through gradient descent. The team’s November 2 blog post reports promising results in selected tests, including a repair to a car-control model and a change to a language model that improved its score on a maths benchmark, but also documents failures and substantial limits.
The researchers call the approach weight-space programming: a language model examines the numbers that determine how a neural network behaves and proposes changes to them. In some tests, Astra received weights as plain text and had to identify a model’s function in a single response. In others, researcher-run code tested changes under a fixed run budget. According to the released logs, Astra had no shell, web search or file-editing tools, and the researchers say they reviewed all 1,125 turns.
For a car-control network that had recorded zero successful reversals in 256 episodes, the team says Astra proposed an edit that led to 1,024 successes in 1,024 subsequent episodes. The change affected 3.5% of the weights. In a separate test, an edit affecting 0.016% of the weights in Qwen3.5-0.8B raised its MATH-500 score from 7.9% to 48.1%, according to the post. A Hugging Face engineer later reconstructed the edited model from the logs and reported a 51.0% score on the benchmark.
The team also reports that Astra identified the games behind four small networks and the shapes represented by three small image generators. In another analysis, it located all the copying, or induction, heads in three tiny language models from their weights. For network creation, the researchers say Astra wrote small game-playing models and a 9-by-9 Go network from zero-valued weights. The Go model beat a same-size comparison model trained on about 20.5 million games across 50 test runs, according to the post.
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 Way to Inspect Model Weights
If the reported results hold up, weight-space programming could offer researchers a way to inspect or make targeted changes to some models without running a conventional training process. That is relevant to teams working with open models, where weights are available but a model’s internal computations can be difficult to interpret. The reported Qwen result also suggests that a narrow edit can change performance on a particular benchmark, though it does not establish broader gains in reasoning or reliability.
The work is also a reminder that weight files are not merely inert products of training: they encode a model’s behavior and, in these tests, could be analyzed and altered by another AI system. That may create new options for audits and small repairs, while making the origin and integrity of weights more consequential. The experiments do not show that companies can safely or reliably apply this method to large production models, or that users will see a practical change soon.
Products Worth Considering
Samsung 34-Inch Odyssey G5 Ultra-Wide Curved Gaming Monitor
LG UltraGear 34GX90SA-W 34-inch WQHD Curved OLED Gaming Monitor
LG UltraGear 27GX704A-B 27-inch QHD 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 whose behavior is encoded in weights produced through optimization rather than written line by line by a programmer. The MIT team’s work tests whether a language model can reason about that numerical representation directly, rather than relying on an optimizer such as gradient descent to adjust it.
The researchers—Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab—describe the work in a public blog post with experiment logs. The source material says the post is not a peer-reviewed paper and has not been posted on arXiv. It also notes that the work used small networks, one research group, limited task and model coverage, and one run per setting. Those conditions make the results a research demonstration, not evidence that the method works consistently across model types.
““not through an optimizer, and in particular not with gradient descent””
— The MIT research team
How Far the Results Generalize
The post does not establish how reliably Astra’s edits work beyond the specific models and tasks tested. The source describes one run per setting, small networks and limited model-family coverage; it does not report broad replication across independent research groups. A strong result on MATH-500 also does not by itself show that the edited Qwen model improved on other measures or avoided regressions elsewhere.
The experiments further found that Astra could not distinguish an honest model from a sycophantic one by weights alone. Edits did not improve maths on Gemma 3 270M, a model that did not exhibit the tested looping behavior. The released logs and an outside reconstruction add useful checks, but the findings remain claims in a research blog post rather than peer-reviewed conclusions. It is not yet clear whether the approach scales to larger models or can be used safely in routine deployment.
Replication and Broader Model Tests
The next useful test is independent replication across more model families, tasks and repeated runs, with checks for unintended effects beyond the benchmark being targeted. Further work would also need to establish whether the approach can distinguish useful edits from changes that merely improve a narrow score, and whether it scales to larger networks.
The source material does not identify a scheduled paper, release or product deployment. For now, the public logs allow other researchers to inspect the reported experiments, while the method remains an early research demonstration. No immediate change for model users is established.
Key Questions
What is weight-space programming?
It is the researchers’ term for having a language model inspect or alter a neural network’s numerical weights directly, rather than changing them through a conventional training optimizer.
What did GPT-6 Astra reportedly repair?
The MIT team says Astra edited a car-control network that had failed to reverse successfully in 256 episodes; the edited network succeeded in 1,024 of 1,024 subsequent episodes. The reported change affected 3.5% of its weights.
Did the Qwen edit improve general intelligence?
The reported result is a higher score on the MATH-500 benchmark, from 7.9% to 48.1%. That result alone does not establish a broad improvement in intelligence, performance across other tasks or freedom from regressions.
Has the work been peer-reviewed?
Not according to the supplied source material. It describes a public MIT team blog post with logs, rather than a peer-reviewed paper or an arXiv preprint.
Can this method be used on models today?
The experiments do not establish reliable use on large production models. The results concern selected small networks, and further replication and testing are needed before practical deployment can be assessed.
Source: ThorstenMeyerAI.com
