🔍 Read the full analysis: Weight-Space Programming Puts Neural Network Parameters To Work 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 testing a method they call weight-space programming, in which GPT-6 Astra reasons directly about neural-network parameters rather than using gradient descent. Their November 2 blog post describes network identification, targeted repairs and new networks, but the experiments were limited and have not been peer reviewed.
Four researchers affiliated with MIT’s Computer Science and Artificial Intelligence Laboratory report that OpenAI’s GPT-6 Astra identified, repaired and created neural networks by working directly with their numerical weights. The experiments, described in a November 2 blog post as “weight-space programming,” did not use a training loop or gradient descent, but the results remain preliminary: the account is a research blog post, not a peer-reviewed paper.
The researchers—Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab—say they supplied Astra with weights from existing networks and asked it to infer what they did, fix selected failures and write small networks from scratch. In some tests the model received weights as plain text and had to answer in one reply. In others, researchers ran code written by Astra in a restricted sandbox. The post says the model had no shell, web search or file-editing tools, and that the team checked 1,125 conversation turns in the released logs for tool use.
Reported examples include identifying the games represented by four small networks and recognizing the shape of three small image generators. In one repair test, the researchers say a car-control network that had completed none of 256 reverse-driving episodes reached 1,024 successes in 1,024 episodes after Astra changed 3.5% of its weights. In another, an edit to 0.016% of Qwen3.5-0.8B’s weights was associated with a MATH-500 score increase from 7.9% to 48.1%.
The team also reports that Astra wrote small game-playing networks and a 9-by-9 Go network from zero-valued weights. According to the post, the Go network beat a same-size comparison network trained on about 20.5 million games across 50 test runs. These are results reported by the researchers; they do not establish that the method will work broadly across network types 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
What Direct Weight Editing Could Enable
If the method proves reproducible, it could offer researchers another way to inspect or adjust small models, including networks whose internal behavior is difficult to understand from their outputs alone. The reported car and language-model tests illustrate a potential distinction from conventional training: an edit can target a small part of a model rather than update its parameters through a standard optimization process.
That possibility is not yet a demonstrated production capability. The post’s experiments were limited in scale and scope, and the authors report cases where editing did not help. For businesses and model developers, the work is best treated as an early research result that may motivate testing and auditing—not as evidence that an AI system can reliably repair arbitrary models. There is no reported change for ordinary users as a result of these experiments.
Products Worth Considering
Samsung 27-Inch Odyssey OLED G5 QHD Gaming Monitor
Samsung 34-Inch Odyssey G5 Ultra-Wide Curved Gaming Monitor
FLEXISPOT 36-Inch Height-Adjustable Standing Desk Converter, Black
As an affiliate, we earn on qualifying purchases.
From Software 2.0 to Weight Editing
The idea builds on a way of describing neural networks popularized by computer scientist Andrej Karpathy. In a 2017 essay, he called neural networks “Software 2.0”: programs whose behavior is encoded in learned weights rather than written instruction by instruction by a person. The MIT group’s proposal asks whether a language model can reason about those parameters as an editable representation.
In ordinary neural-network training, an optimizer adjusts weights repeatedly, often using gradient descent and data. The researchers say their approach instead has a language model examine weight values and reason about what to change. Their November 2 post shares logs of the experiments, but the supplied material says the work is not a peer-reviewed paper and is not on arXiv. A Hugging Face engineer reportedly reconstructed the edited Qwen model from the logs and measured 51.0% on MATH-500; that independent check is described in the source post, not as a formal evaluation of the full set of experiments.
““not through an optimizer, and in particular not with gradient descent””
— The MIT research team, in its November 2 blog post
How Far the Tests Can Be Generalized
The evidence has substantial limits. The source describes one run per setting, small networks and a single model family for each task, conducted by one research group. The results have not been reported in a peer-reviewed paper, and the post’s logs do not settle whether the method will transfer to larger models or different applications.
The researchers also report negative results: Astra could not distinguish an honest model from a sycophantic one using weights alone, and edits to Gemma 3 270M did not improve its maths performance. The post says the model was not looping in that test, and the team’s explanation is that editing could not supply an ability the model did not already have. It remains unclear how reliably the method works, how sensitive outcomes are to repeat runs, and what safeguards would be needed before applying edits to systems used in practice.
Replication and Broader Testing
The next useful step is independent replication across repeated runs, additional model families and tasks beyond the small examples described in the post. Researchers would also need to report how edits affect performance on measures other than the target task, including whether a repair introduces new errors or changes behavior unexpectedly.
The public logs and the reported Qwen reconstruction offer material that other researchers can examine, but the source does not give a timetable for a paper, formal review or further experiments. Until more evidence is available, the results should be read as an early demonstration, not a validated general-purpose method for modifying neural networks.
Key Questions
What is weight-space programming?
It is the researchers’ term for having a language model inspect or change another neural network’s numerical weights directly, rather than relying on a conventional training loop or gradient descent.
What did the researchers say Astra accomplished?
The blog post reports that Astra identified small game-playing networks, made targeted edits to a car-control network and Qwen3.5-0.8B, and wrote small networks from zero-valued weights. These findings are reported experimental results, not proof of general performance.
Was the Qwen maths result independently checked?
The source post says a Hugging Face engineer rebuilt the edited model from the released logs and measured 51.0% on MATH-500, compared with the researchers’ reported 48.1%. The supplied material does not describe a formal independent review of all experiments.
Is weight-space programming ready for routine use?
The evidence provided does not establish that. The work is described as a blog post rather than a peer-reviewed paper, and its tests were limited in size and scope. The authors also report failed or unhelpful edits.
What happens next?
The public logs can be examined and the reported results can be tested by other groups. The source gives no schedule for peer review or further experiments.
Source: ThorstenMeyerAI.com
