🔍 Read the full analysis: Weight-Space Programming: The Potential And Limits Of Editing Models By Weight on ThorstenMeyerAI.com
Get smart everyday buys 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 identified, repaired and created small neural networks by reasoning directly about their weights. Their public logs document striking results in several tests, but also failures; the work is a blog post, not a peer-reviewed paper, and its experiments are limited in scale.
Four MIT CSAIL researchers say OpenAI’s GPT-6 Astra identified, repaired and wrote small neural networks by working directly with their numerical weights, rather than using a training loop or gradient descent. The experiments, described in a November 2 blog post and accompanied by public logs, point to a possible new way to inspect and make targeted changes to models, while leaving open how well the approach works beyond the small tests reported.
The researchers call the method weight-space programming: a language model examines another network’s parameters as data, infers what the network does and, in some tests, changes or generates weights. The team says Astra operated without built-in tools such as shell access, web search or file editing. Some tasks required a single response based on weights presented as text; others allowed the researchers’ sandbox to run code written by Astra under fixed testing limits.
In one repair test, a car-control network that had achieved zero successes in 256 episodes completed 1,024 of 1,024 episodes after Astra changed 3.5% of its weights, according to the post. In another, editing 0.016% of the weights in Qwen3.5-0.8B raised its score on MATH-500 from 7.9% to 48.1%. The source reports that a Hugging Face engineer rebuilt the edited Qwen model from the logs and obtained a 51.0% score on the same benchmark.
The post also describes Astra identifying the games represented by four small networks and the structures represented by three small image generators. In tests of three tiny language models, it identified all the researchers’ designated induction, or “copying,” heads from weights alone. For generation, Astra wrote small game-playing networks from all-zero weights and produced a 9-by-9 Go network; the authors say the latter beat a same-size comparison network trained on about 20.5 million games across 50 test runs. These are results reported by the research team, not evidence that the method matches conventional training across 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 Inspection
If the reported results can be reproduced and extended, weight-space programming could offer developers another way to investigate model behavior and make targeted repairs, particularly in small open models. A system that can infer what parameters encode, or identify a narrow defect, might help researchers inspect models whose internal operations are difficult to interpret through ordinary inputs and outputs.
The tests also underline a potential governance concern: the model weights themselves become material another AI system can analyze and modify. That could make weight provenance and change tracking more important for organizations deploying models. For users, however, the post describes research experiments, not a released consumer tool or a change to models currently in use. The results do not establish that an AI can reliably edit large production systems or safely make unsupervised changes.
Popular right now
Ninja Foodi 6-in-1 DualZone Air Fryer, 8 qt, Grey (DZ201)
Dyson Car+Boat Cordless Handheld Vacuum
Dyson W1 PencilWash Wet and Dry Cleaner with Dok Cordless Wet Floor Cleaner
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 learned through optimization rather than written directly by a person. The MIT team’s experiment asks whether a general-purpose language model can treat those learned values more like editable code. Its researchers are Rikiya Takehi and Jayden Teoh, who contributed equally, alongside Alex Zhang and Omar Khattab.
The conventional process adjusts parameters repeatedly using an optimizer and training data. The approach described in the post instead has Astra reason about the numbers themselves, sometimes proposing edits and sometimes writing weights from scratch. The article summarizes eight experiments, but the evidence comes from a single research group’s blog post, not a peer-reviewed paper or an arXiv preprint, according to the source material.
“Software 2.0”
— Andrej Karpathy, describing neural networks in 2017
How Far the Evidence Extends
The results have not yet been established across independent studies or large, varied model families. The source describes one run per setting, small networks and a single model family for each task; those limits make it hard to tell how repeatable the reported gains are or whether they generalize. The blog post includes logs, and the Qwen result was independently rebuilt and retested by a Hugging Face engineer, but that check does not validate every experiment.
The post also reports failures: Astra could not distinguish an honest model from a sycophantic one using weights alone, and its edits did not improve maths in Gemma 3 270M. The evidence does not show that the method can reliably detect or correct broader behavioral problems, or that it can add capabilities absent from a model. Safety, reliability and performance at larger scales remain unestablished.
Replication and Larger-Scale Tests
The immediate next step is independent replication of the logged experiments, followed by tests across more architectures, tasks and model sizes. Further work would need to establish how often weight edits produce durable improvements, whether they create regressions elsewhere and how proposed changes can be reviewed before deployment. The source material does not identify a publication date for a peer-reviewed paper or a planned commercial release.
For now, the public logs provide a basis for researchers to examine the specific claims. Until broader testing is available, the work is best understood as an early demonstration of direct model-weight analysis and editing, 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 another neural network’s numerical weights, rather than changing them through a conventional optimizer and gradient descent.
What did GPT-6 Astra do in the reported tests?
According to the MIT researchers, Astra identified small networks’ functions, repaired a car-control network, made a targeted edit to Qwen3.5-0.8B and generated small networks from zero weights. The reported tests were limited in scale.
Did the method improve every model it tested?
No. The researchers report that Astra could not distinguish an honest model from a sycophantic one from weights alone, and its edits did not improve maths performance for Gemma 3 270M.
Has the research been peer reviewed?
The source describes the findings as a blog post with public logs, not a peer-reviewed paper or an arXiv preprint. A Hugging Face engineer reportedly rebuilt the edited Qwen model and retested it, but that is not independent confirmation of all experiments.
Does this change anything for AI users now?
The source reports no consumer release or immediate product change. The experiments are an early research demonstration, and performance on larger, real-world models remains uncertain.
Source: ThorstenMeyerAI.com
