🔍 Read the full analysis: Weight-Space Programming Shows How AI Can Work Inside A Model on ThorstenMeyerAI.com
Get the little things that make your day 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
Four MIT CSAIL researchers report that OpenAI’s GPT-6 Astra could infer functions from neural-network weights, repair some small models and write new ones. Their results come from a blog post and logged experiments, not a peer-reviewed paper; tests were limited, and some attempted fixes did not work.
Four researchers at MIT’s Computer Science and Artificial Intelligence Laboratory report that OpenAI’s GPT-6 Astra could interpret and edit the numerical weights of other neural networks, then write small networks of its own. The November 2 blog post describes eight experiments under the name weight-space programming, including repairs to a car-control model and a small language model, but the work is not a peer-reviewed paper and its tests were limited.
The researchers — Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab — supplied Astra with networks’ weights, the numerical values that determine how signals pass between artificial neurons. They asked it to infer what a network did, make targeted changes when it failed, or create a network from scratch. The researchers say the model reasoned about weights directly, without an optimizer or gradient descent.
In one reported repair, a car-control network that had recorded zero successful reversals in 256 episodes achieved 1,024 successes in 1,024 episodes after Astra changed 3.5% of its weights. In a separate experiment, an edit to 0.016% of the weights in Qwen3.5-0.8B raised its reported MATH-500 score from 7.9% to 48.1%. A Hugging Face engineer later reconstructed the edited Qwen model from the released logs and reported a 51.0% score on the benchmark.
The post also describes reading and generation tests. Astra identified the games behind four small game-playing networks and the forms represented by three small image generators, based on their weights. It identified copying, or induction, heads in three tiny language models without running them. In another test, it wrote a 9-by-9 Go network from zero weights; the researchers report that it beat a same-size comparison network trained on about 20.5 million games across 50 test runs. These are the researchers’ reported results, not evidence that the approach will work broadly across models.
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
What Direct Weight Editing Could Enable
If the reported approach proves reliable, it could offer researchers another way to inspect and make targeted changes to small neural networks. Today, model weights are usually adjusted through training algorithms rather than interpreted as a program by a general-purpose language model. Direct editing could make some debugging tasks faster or help examine models whose training process is difficult to reproduce.
The work also points to a potential audit use: a model might be asked to identify structures in another network from its weights alone. But the experiments do not establish that such readings are dependable for larger or more varied systems. For companies deploying open models, the possibility of direct edits also makes weight provenance and version checks relevant operational concerns. The authors’ results do not change how ordinary users’ systems work today.
Products Worth Considering
Roborock Robot Vacuum and Mop Combo with Self-Empty Dock
Philips Hue Essential Starter Kit with Bridge and 4 E26 Smart Bulbs
Shark AI Ultra Robot Vacuum, Black/Silver, AV2511AE
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”: programs encoded in weights and produced through optimization rather than written directly by people. He noted that programming through millions of weights was difficult. The MIT team’s project tests whether a language model can reason about those values as if they were code.
The researchers say Astra was called through OpenAI’s Codex command-line tool with built-in tools disabled. Some tasks presented weights as plain text and required a single response. Other experiments let the researchers’ sandbox run Python code written by Astra, subject to fixed test limits and no network access. The team says it reviewed all 1,125 turns in the released logs and found no tool use. The source account says the logs are public, but the work is a blog-post report, not a peer-reviewed study.
“Software 2.0”
— Andrej Karpathy, describing neural networks in 2017
How Far the Results Generalize
The reported work has several limits. The source account describes one run per setting, small networks, one model family per task and a single research group. The results therefore do not show how often the method succeeds across repeated trials or whether it works on larger, more varied models. The blog post and logs have not been described as a peer-reviewed paper, and the source says the work was not posted on arXiv.
The experiments also include clear failures. Astra could not distinguish an honest model from a sycophantic one using weights alone, and its edits did not improve maths on Gemma 3 270M. The researchers say editing could not supply an ability a model lacked. It remains unclear how robust the changes are outside the reported tests, how they compare with conventional debugging and fine-tuning, and whether independent teams can reproduce the full set of results.
Replication and Broader Model Tests
The immediate next step is independent replication, particularly across repeated trials, model families and larger networks. Researchers would also need to test whether weight-based diagnoses predict model behavior reliably, and whether edits persist across different evaluation sets. The source material does not give a schedule for a follow-up paper or further experiments.
For now, the public logs and the reported Qwen reconstruction provide material for other researchers to examine. Until broader testing is available, weight-space programming is best treated as an early research direction, not a general replacement for training or model evaluation.
Key Questions
What is weight-space programming?
It is the researchers’ term for having a language model inspect or edit a neural network’s numerical weights directly, rather than changing them through a training optimizer such as gradient descent.
What did GPT-6 Astra reportedly repair?
The team reports that Astra edited a car-control network that had failed to reverse and a Qwen3.5-0.8B language model that got stuck in loops. The reported results apply to those specific experiments.
Was the work peer-reviewed?
No. The source describes the findings as a blog post published November 2, 2026, with public logs. It says the work was not a peer-reviewed paper and was not on arXiv.
Did the method work in every test?
No. Astra could not identify honesty versus sycophancy from weights alone, and its edits did not improve maths performance in the Gemma 3 270M test. The researchers say editing could not add an ability the model lacked.
Does this change how people use AI models now?
No immediate change for users is reported. The experiments suggest a possible research and debugging method, but its reliability and usefulness beyond the tested small networks remain uncertain.
Source: ThorstenMeyerAI.com
