🔍 Read the full analysis: What It Looks Like When AI Works Directly With Neural Network Numbers 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 interpreted, edited and created small neural networks by reasoning directly about their weights, rather than using gradient descent. The team’s November 2 blog post reports several strong results, but the experiments are limited and have not been peer-reviewed.
Researchers at MIT CSAIL report that OpenAI’s GPT-6 Astra could identify what small neural networks did, edit some that performed poorly and write others from scratch by working directly with their numerical weights. The experiments, described in a November 2 blog post and accompanied by public logs, test a proposed approach called weight-space programming; they are not a peer-reviewed study.
Neural-network weights are the numbers that determine how strongly units in a model affect one another. Typically, training algorithms adjust those numbers through repeated optimization. In the team’s experiments, Astra was instead asked to reason about weights directly. The researchers say the model did not use gradient descent to make the reported edits. Some tasks gave Astra weights in plain text and requested an answer in one response; others allowed researchers’ sandbox to run code the model wrote, within a fixed test budget.
The team reports successes across three kinds of task. Astra identified the games represented by four small networks and the output shapes of three small image generators. It also located copying heads in three tiny language models without running them. In one repair, a car-control network that had recorded zero successes in 256 episodes achieved 1,024 successes in 1,024 episodes after an edit that changed 3.5% of its weights.
Another reported repair involved Qwen3.5-0.8B, a small language model that repeated itself in loops. The blog says Astra changed 0.016% of its weights and raised its score on MATH-500 from 7.9% to 48.1%. A Hugging Face engineer later rebuilt the edited model from the released logs and reported a score of 51.0%, according to the source article. Astra also produced small game-playing networks from zero-valued weights, including a 9×9 Go network that beat a 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 Work
If the results hold up in broader testing, weight-space programming could offer another way to inspect or make targeted changes to small models. Instead of relying only on a training pipeline, a developer might ask a language model to identify a behavior associated with particular weights and propose a limited edit. The reported Qwen change, for example, affected a small fraction of the model’s weights while coinciding with a higher score on one maths benchmark.
That possibility is not the same as a general-purpose replacement for training or model evaluation. The experiments do not show that an AI can reliably diagnose arbitrary networks, improve every capability, or safely edit large deployed systems. For organizations using open models, the work also puts a spotlight on how weights are inspected and changed: edits may be difficult to understand without tests, provenance records and independent reproduction. The source article says the findings do not change anything for ordinary users today.
Products Worth Considering
Bose TV Speaker (2nd Gen) Compact Soundbar
Apple MacBook Air 15-inch with M5 chip, 16GB Memory, 512GB SSD, and Liquid Retina Display
DJI Osmo Mobile 8 3-Axis Gimbal Stabilizer for Phones
As an affiliate, we earn on qualifying purchases.
From Software 2.0 to Weight-Space Tests
In 2017, computer scientist Andrej Karpathy described neural networks as “Software 2.0”: systems whose behavior is encoded in weights shaped by optimization rather than written as conventional instructions. He also noted that directly programming millions of weights was difficult. The MIT researchers’ proposal revisits that idea by asking whether a language model can reason about those numbers as if they were editable code.
The team says Astra was accessed through OpenAI’s Codex command-line tool with built-in tools such as shell access, web search and file editing switched off. For experiments involving code, the researchers’ sandbox ran code written by Astra without network access. The blog says the team checked all 1,125 turns in the release and found them tool-free. The source article emphasizes that this is a single research group’s blog post, not a peer-reviewed paper or an arXiv preprint, and that the experiments generally used small networks and one run per setting.
““not through an optimizer, and in particular not with gradient descent””
— The MIT CSAIL researchers, as quoted in their November 2 blog post
How Far the Results Generalize
The reported results remain preliminary. The post describes one run per setting, small networks, and limited model coverage, with one model family used for each task. The source article says Astra could not distinguish an honest model from a sycophantic one based on weights alone, and that edits did not improve maths performance in Gemma 3 270M, which was not exhibiting the same looping problem. The team’s own account says editing could not add an ability the model lacked.
It is also unclear whether other researchers can reproduce the full set of results across models and tasks. The source reports a separate reproduction of the edited Qwen model’s benchmark score, but that does not independently validate every experiment. The blog post is not peer-reviewed, and the supplied material does not establish how reliably the approach would work on larger networks, production systems or other model families.
Independent Tests Are the Next Test
The immediate next step is independent replication, particularly of the reported repairs and the creation of networks from zero-valued weights. Further testing would need to vary model sizes, architectures, tasks and repeated runs to show whether the results are stable rather than isolated successes. Researchers would also need to compare direct edits with standard training and other intervention methods.
The public logs give other researchers material to inspect, and the source article reports that one engineer has already reconstructed the Qwen edit and rerun its benchmark. No formal paper or broader follow-up plan is identified in the supplied material. Until more evidence is available, weight-space programming is best understood as a promising research demonstration, not a proven method for routinely repairing or building AI models.
Key Questions
What is weight-space programming?
It is the approach tested by the MIT CSAIL researchers: a language model reads or changes another neural network’s numerical weights directly, rather than relying on gradient descent for those changes.
What did Astra reportedly repair?
The team says Astra edited a car-control network that had no successes in 256 episodes and a Qwen3.5-0.8B model that repeated itself in loops. The reported Qwen MATH-500 score rose from 7.9% to 48.1% after the edit; an engineer’s reconstruction reportedly scored 51.0%.
Was the research peer-reviewed?
No peer-reviewed paper is identified in the source material. The results were described in a public MIT CSAIL blog post, with logs provided by the researchers.
Does this mean AI can repair any model?
No. The experiments covered a limited set of small networks, and the researchers reported failures as well as successes. They said editing could not add an ability a model lacked, and Astra could not identify every behavior from weights alone.
Source: ThorstenMeyerAI.com
