🔍 Read the full analysis: What AI’s Work In Weight Space Reveals About Neural Networks on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
MIT CSAIL researchers report that OpenAI’s GPT-6 Astra interpreted and edited neural-network weights directly, without a training loop or gradient descent. Tests included repairing a car-control model and improving one small language model’s maths score, but the work is a blog post with limited experiments, not a peer-reviewed paper.
Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory report that OpenAI’s GPT-6 Astra could interpret, repair and create small neural networks by directly changing their weights, without using gradient descent. The experiments, described in a November 2 blog post, offer an early test of “weight-space programming”—but their limited scale and single-run results do not establish how well the approach generalizes.
The researchers supplied Astra with the numerical weights of existing networks and asked it to infer their functions or make targeted changes. In some tests, the model received weights as text and had to respond in one turn. Other tests allowed researchers’ sandbox to run code written by Astra, but the sandbox had no network access. The team says its review of 1,125 recorded turns found no use of built-in tools such as shell access, web search or file editing.
In one repair test, a car-control network that had recorded zero successful reversals in 256 episodes reached 1,024 successes in 1,024 episodes after Astra changed 3.5% of its weights. In another, an edit affecting 0.016% of the weights in Qwen3.5-0.8B was followed by a MATH-500 score increase from 7.9% to 48.1%, according to the post. The source reports that a Hugging Face engineer later rebuilt the edited model from the logs and measured 51.0% on the benchmark.
The team also says Astra identified four small game-playing networks from their weights and inferred the shapes produced by three small image generators. It wrote a 1.5-million-weight network for 9-by-9 Go from zero; after 50 test runs, the network beat a same-size comparison model trained on about 20.5 million games. These are the researchers’ reported results, not independent evidence that the method works broadly.
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 Inspection
If the reported results hold up in broader testing, direct weight editing could give researchers another way to inspect or adjust small models, including systems whose behavior is difficult to trace through ordinary code. The car and Qwen demonstrations suggest that a language model may sometimes identify a useful intervention without retraining the target network from scratch.
That possibility is relevant to organizations using open models: a targeted edit could be cheaper than a full training run. But the post does not show that this method is reliable enough for production systems, or that it can replace standard evaluation and fine-tuning. The authors also report a failure to distinguish an honest model from a sycophantic one using weights alone. The experiments point to a research direction, not a ready-made auditing tool.
Popular right now
Philips Café Aromis Automatic Espresso Machine with LatteGo Pro Milk System
De'Longhi Dinamica Plus Fully Automatic Espresso Machine
Hisense 65-inch U6 Pro Mini-LED 4K UHD Smart Fire TV
As an affiliate, we earn on qualifying purchases.
From Software 2.0 to Weight Edits
In 2017, Andrej Karpathy described neural networks as “Software 2.0”: systems whose behavior is encoded in learned weights rather than conventional, human-written instructions. Training algorithms normally adjust those weights through repeated optimization, making the resulting numerical parameters difficult to read as a program.
The MIT researchers frame their work as a test of whether a language model can reason about those parameters directly. They call that approach weight-space programming. The source describes eight experiments spanning network interpretation, repair and creation. It also notes that Astra did not improve maths performance when editing Gemma 3 270M, a model that did not exhibit the looping behavior targeted in the Qwen test. The authors say editing could not add an ability the model lacked.
The evidence is currently a public blog post with logs, rather than a peer-reviewed publication or an arXiv paper, according to the supplied source. The experiments used small networks, one run per setting, and—in each task—one model family. Those limits matter when comparing the demonstrations with claims about general-purpose AI model development.
“Software 2.0”
— Andrej Karpathy, describing neural networks in 2017
How Far the Tests Generalize
It remains unclear whether the results will reproduce across independent research groups, larger models or repeated trials. The supplied account says there was one run per setting and that the experiments covered small networks and one model family per task. The source does not report peer review, and the blog post’s findings should not be treated as settled evidence about frontier-model capabilities.
The mechanism behind some apparent successes also needs further examination. The source reports that software engineer Sean Goedecke reproduced a visualization used to interpret a game network with a short Python script, raising a question about how much of the network identification depended on meaningful weight-level reasoning. The account does not provide enough detail to resolve that issue. It is also unknown whether the Qwen score gain persists across broader evaluations or different prompts and settings.
Replication and Broader Evaluation
The immediate next step is independent replication of the logged experiments, including repeated runs and tests on more architectures and tasks. Further evaluations would need to establish whether weight edits remain effective outside the specific models and benchmarks described, and whether improvements on one metric create problems elsewhere.
The source identifies no scheduled paper, follow-up study or product release. For now, the researchers’ public logs allow other researchers to inspect the work, while the reported Qwen rebuild offers one check on a particular result. Until broader evidence is available, the findings are best understood as an initial demonstration of a possible research method; the source says they do not change anything for users today.
Key Questions
What is weight-space programming?
It is the approach described by the MIT team in which a language model reads or changes a neural network’s numerical weights directly, rather than relying on a training optimizer to adjust them.
What did GPT-6 Astra do in the reported tests?
The researchers say Astra identified small game-playing networks, made edits associated with improved performance in a car-control task and a Qwen maths benchmark, and wrote small networks, including one for 9-by-9 Go.
Are the results independently confirmed?
Not across the full set of experiments, based on the supplied source. It describes a public blog post and logs, not peer-reviewed research. A Hugging Face engineer reportedly rebuilt the edited Qwen model and measured a 51.0% MATH-500 score.
Did the method work in every test?
No. The source says Astra could not identify whether a model was honest or sycophantic from its weights alone, and edits did not improve maths performance in Gemma 3 270M, which was not looping.
Does this change how people use AI models now?
The source says no immediate change for users. The experiments suggest a possible research tool, but the method’s reliability and usefulness beyond the tested small networks remain uncertain.
Source: ThorstenMeyerAI.com
