🔍 Read the full analysis: How Weight-Space Programming Lets AI Edit Neural Networks 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
MIT researchers report that OpenAI’s GPT-6 Astra could infer what several small neural networks did, repair selected failures and create networks by directly editing their weights. The results, published in a November 2 blog post rather than a peer-reviewed paper, include a large maths benchmark gain for one Qwen model but also failures and limited testing.
MIT CSAIL researchers reported November 2 that OpenAI’s GPT-6 Astra could read and modify the numerical weights of small neural networks to identify their tasks, repair some failures and write new networks. The experiments test weight-space programming—using a language model to reason about network weights directly rather than relying on an optimizer—but remain early evidence from a blog post, not a peer-reviewed study.
The researchers supplied Astra with raw weights from small networks and asked it to infer their function or make changes. In the team’s described setup, the model did not use gradient descent. Released logs show Astra was called through OpenAI’s Codex command-line tool with built-in tools such as shell access and web search disabled. Some tasks required a single response; others let Astra write Python code that ran in a restricted research sandbox.
The most striking repair example involved a car-control network that had failed to reverse. The source reports zero successes in 256 episodes before editing and 1,024 successes in 1,024 episodes afterward, with 3.5% of weights changed. In a separate test, an edit affecting 0.016% of a Qwen3.5-0.8B model’s weights raised its MATH-500 score from 7.9% to 48.1%, according to the researchers. A Hugging Face engineer later rebuilt that edited model from the logs and reported a 51.0% score.
The team also said Astra identified the games played by four small networks and inferred the shapes produced by three small image generators from weights alone. It wrote small game-playing networks and a 9-by-9 Go network from zero weights; the source says the Go network beat a same-size model trained on about 20.5 million games across 50 test runs. These are reported outcomes for specific experiments, not evidence that the method works reliably across networks or tasks.
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 Repairs
If the approach can be reproduced and extended, it could offer developers a way to inspect or make targeted changes to small models without running conventional training. The Qwen result is notable because the reported edit touched a tiny fraction of its weights while substantially improving its score on one maths benchmark. The car example likewise suggests that a model’s behavior may sometimes be changed through a limited, direct intervention.
But these tests do not show that language models can generally replace training, nor that a weight edit will be safe or dependable in deployed systems. A narrow benchmark gain may not carry over to other tasks, and a changed model still needs evaluation for regressions. The work also raises practical questions about auditing who modified a model and how its weights were produced. The researchers’ results are a possible research direction, not a change to what ordinary users can do with AI today.
Products Worth Considering
LG UltraGear 32GX850A-B 32-inch 4K UHD Glossy OLED Gaming Monitor
Amazon Kindle Paperwhite (2024), 16 GB, 7-inch Display, Black
Bose QuietComfort Ultra Headphones (2nd Gen), Wireless Bluetooth Noise-Cancelling Over-Ear Headphones, Black
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 line by line by people. The MIT team’s work builds on that idea by asking whether a general-purpose language model can treat weights as something it can inspect and edit. Normally, training algorithms adjust weights repeatedly using data and an optimization procedure.
The researchers—Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab—call their method weight-space programming. Their November 2 post argues that language models may eventually reason about weights instead of relying only on symbolic optimizers. The published account describes eight experiments and includes public logs. It also acknowledges limits: Astra could not distinguish an honest model from a sycophantic one using weights alone, and edits did not improve maths performance for Gemma 3 270M, a model that did not exhibit the looping behavior targeted in the Qwen test.
“not through an optimizer, and in particular not with gradient descent”
— The MIT research team, in its November 2 blog post
How Far the Results Extend
The evidence is preliminary. The source describes a blog post, not a peer-reviewed paper, and says there was one run per setting, small networks, one model family per task and a single research group. The reported gains therefore do not establish repeatability or performance on larger, widely deployed models. The independent Qwen re-run is a useful check, but it does not replicate the full set of experiments.
It is also unclear how well the method works when network architectures, training histories or tasks differ from those tested. In at least one image-network example, software engineer Sean Goedecke reportedly argued that the same visual pattern could be reproduced with a short Python script, raising questions about how much of Astra’s inference required the language model’s reasoning. The source does not establish whether direct weight editing can be made reliable, audited at scale or used without introducing new errors.
Replication and Broader Testing
The next test is whether independent researchers can reproduce the results across repeated runs, additional model families and more demanding tasks. Further evaluation would need to check not just whether an edit fixes its target, but whether it harms performance elsewhere and whether the method can reliably detect when a proposed change is inappropriate.
The MIT team’s released logs give others material to inspect, while the Qwen re-run offers one independent check on a specific result. The source does not announce a follow-up paper, release schedule or product deployment. Until broader evidence is available, the work should be treated as an experimental method rather than a practical replacement for established training and evaluation procedures.
Key Questions
What is weight-space programming?
Weight-space programming is the researchers’ term for having a language model inspect or change another neural network’s numerical weights directly, rather than using gradient descent to train it.
What did Astra repair?
In one reported test, Astra edited a car-control network that had recorded zero successful reversals in 256 episodes. The edited network succeeded in 1,024 of 1,024 episodes, according to the source.
Did the Qwen maths result hold up in a separate test?
The source says a Hugging Face engineer reconstructed the edited Qwen model from the logs and reported a 51.0% MATH-500 score. That is one independent check of this result, not a replication of the broader experiment set.
Has the method been peer-reviewed?
No peer-reviewed paper is identified in the source. The results were described in a November 2 blog post with public logs.
Does this mean AI can now edit any model successfully?
No. The reported experiments involve small networks and specific tasks. The team also reported failures, including no maths improvement from edits to Gemma 3 270M, and said editing could not add an ability a model lacked.
Source: ThorstenMeyerAI.com
