🔍 Read the full analysis: Reading Neural Networks By Their Numbers: A Look At Weight-Space Programming on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
A November 2 blog post from four MIT CSAIL researchers describes experiments in which OpenAI’s GPT-6 Astra interpreted, edited and generated neural-network weights without gradient descent. The reported results include a car-control repair and a large improvement on one maths benchmark, but the work is not peer-reviewed and uses limited tests; the authors also report failures.
Four researchers at MIT CSAIL say they used OpenAI’s GPT-6 Astra to inspect, modify and create neural networks by working directly with their numerical weights, rather than using a conventional training process. In a blog post published November 2, 2026, the team calls the approach “weight-space programming” and reports successes on small control, language and game-playing tasks, alongside clear limits.
The experiments asked Astra to infer what networks did from their weights, repair some that performed poorly, and write others from scratch. The researchers say the model was given no gradient descent and, in some tests, had to reason from weights presented as plain text. Their released logs document the exchanges and test runs, according to the source account.
In one repair, 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 accompanied a rise in its MATH-500 score from 7.9% to 48.1%. Those are results reported for the specific experiments, not evidence that the method will yield similar gains on other models.
The researchers also report that Astra identified the games behind four small networks from their weights, found induction heads in three small language models, and generated a 9-by-9 Go network from zero weights. The Go network beat a same-size comparison network trained on about 20.5 million games in 50 test runs, according to the post. The source account says a Hugging Face engineer reconstructed the edited Qwen model from the logs and measured a 51.0% MATH-500 score.
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 Audits
If the method can be replicated and extended, it could offer researchers another way to investigate small models: examining parameters for patterns, making targeted changes, or generating a network without the usual training pipeline. That possibility matters for organizations that maintain open-weight models, where access to the parameters may allow inspection and modification not possible with a hosted service.
The reported gains also require careful interpretation. The Qwen result concerns one model and benchmark, and the car result concerns one task and network. A benchmark increase does not by itself establish broad improvements in reasoning, reliability or real-world performance. The experiments show a possible technique, not a ready-made method for routine model maintenance.
For users, the post describes research rather than a product release. It does not establish that weight-space programming is available as a dependable service or that it changes how deployed AI systems work today. For developers, it raises practical questions about how edits are validated and how the origin and integrity of model weights are tracked.
Products Worth Considering
Apple MacBook Air 15-inch with M5 chip, 16GB Memory, 512GB SSD, and Liquid Retina Display
Bose SoundLink Flex Bluetooth Speaker (2nd Gen) – Portable Outdoor Speaker with Hi‑Fi Audio, Waterproof and Dustproof, USB‑C, Up to 12 Hours
Apple MacBook Air 13-inch with M5 chip, 16GB memory, 512GB SSD, and 12MP Center Stage camera
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 shaped by optimization rather than written line by line. The MIT researchers’ work takes up a related question: whether a language model can reason about those parameters directly, instead of relying on an optimizer to adjust them through training.
A neural network’s weights are numerical values that influence how signals pass between units. Training typically updates those values repeatedly against data and an objective. The MIT team instead asked a language model to interpret or change weight values. The source account says Astra was accessed through OpenAI’s Codex command-line tool, with built-in tools disabled for the model; in some experiments, researchers ran Python code written by Astra in a sandbox with no network access.
The report describes eight experiments, but the evidence remains narrow: the work comes from one research group, uses small networks and limited test settings, and is presented in a blog post rather than a peer-reviewed paper. Its released logs offer material for checking the reported exchanges, but do not substitute for independent, broader evaluation.
““not through an optimizer, and in particular not with gradient descent””
— The MIT CSAIL researchers, as quoted in the November 2 blog post
Replication and Generality Remain Open
The work has not been described in the source material as a peer-reviewed paper or an arXiv preprint. The researchers report released logs, and a separate engineer’s reconstruction of the Qwen edit provides one additional check, but the available account does not establish independent replication across research groups or model families.
The number of runs was limited: the source says there was one run per setting, with small networks and one model family per task. It also reports that Astra could not distinguish an honest model from a sycophantic one using weights alone, and that edits did not improve maths in Gemma 3 270M, which was not looping. The authors’ stated interpretation is that edits could not supply an ability the model lacked.
It remains unclear how well the approach would work on larger, more complex networks, whether its edits are reliable across repeated trials, and what safeguards would be needed before using generated or modified weights in deployed systems. The source material does not establish that the method can replace conventional training or provide a general audit of model behavior.
Independent Tests Will Set the Scope
The immediate next step is independent testing of the released procedures and logs, including attempts to reproduce the reported car-control and Qwen results. Repeating each setting across multiple runs and applying the method to different models and tasks would help show whether the outcomes are robust or specific to the examples in the post.
Further work would also need to test whether weight-based interpretations accurately predict a model’s behavior, rather than merely offering plausible explanations after the fact. For now, the MIT researchers’ experiments provide a documented research claim and a set of examples to examine; they do not establish a general-purpose way to program neural networks by hand.
Key Questions
What is weight-space programming?
It is the approach described by the MIT researchers in which a language model reads or changes another neural network’s weights directly, rather than relying on gradient descent to train those values.
What did GPT-6 Astra reportedly accomplish?
The researchers say Astra identified small networks’ tasks, repaired a car-control network, edited a Qwen model associated with a higher MATH-500 score, and generated small game-playing networks, including one for 9-by-9 Go.
Has the work been independently confirmed?
The source account describes a blog post with released logs, not a peer-reviewed paper. It also reports that a Hugging Face engineer reconstructed the Qwen edit and measured a 51.0% MATH-500 score. Broader independent replication is not established by the material provided.
What are the main limits reported so far?
The tests involved small networks, limited settings and one run per setting, according to the source account. Astra also failed to distinguish an honest model from a sycophantic one by weights alone, and its edits did not improve maths in a model that was not already looping.
Source: ThorstenMeyerAI.com
