🔍 Read the full analysis: How To Think About Weight Space Programming And AI Model Rewriting 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
A November 2026 MIT CSAIL post describes experiments in which GPT-6 Astra analyzed and changed neural-network weights without gradient-descent training. A separate set of tests reported by ThorstenMeyerAI.com found a narrow benchmark improvement and useful error detection, but no gain on a legal-task score; the results do not establish broad model-rewriting capability.
MIT CSAIL researchers reported November 2 that OpenAI’s GPT-6 Astra could inspect and alter neural-network weights without a training loop, in experiments the researchers call weight space programming. A separate account published by ThorstenMeyerAI.com and updated November 9 describes mixed attempts to apply the approach to a small language model, including one narrow benchmark gain and a legal-task score that did not improve.
In the MIT experiments, Astra received network weights as text and was asked to infer what the networks did, construct networks, or make targeted changes. The researchers say it identified the games played by four small networks and detected copying behavior in three small language models. They also report that it wrote arcade-game networks from zero-valued weights which outperformed networks trained for 100 million steps in the tested games. The experiments and evaluations were published by the team, according to the source report.
The most specific editing result involved Qwen3.5-0.8B, a small open model. The MIT team’s reported edit removed one direction of a looping behavior by changing 0.016% of the weights; its score on MATH-500 rose from 7.9% to 48.1%. That is a result on a particular benchmark and model, not evidence that the method generally improves reasoning or other capabilities.
ThorstenMeyerAI.com reports three additional trials using two AI systems. In one, Claude-edited weights let a local model answer all 10 published legal tasks, up from four, but its legal score remained unchanged. A separate AI review found faults missed by existing green tests. In a third trial, changing one matrix raised a narrow JSON test result from 49 to 70 out of 120. The report also says one attempt produced no useful result. These are reported tests, not a broad independent validation of the MIT claims.
Infographic 01 · What weight space programming showed
Read, write, edit: what GPT-6 Astra did with raw weights
A blog post with complete public logs, not a peer-reviewed paper. Small networks, one research group, one run per setting.
MATH-500 score of the small open model Qwen3.5-0.8B after GPT-6 Astra removed one loop direction from 0.016% of its weights. Before: 7.9%. A Hugging Face re-run of the edited model: 51.0%.
Three things an AI did with raw weights
No training loop, no gradient descent
- Read: named the game behind 4 of 4 small game networks (17,253 weights each) and found every copying head in three small language models with their first and last layers removed.
- Wrote: arcade-game networks from all-zero weights that beat networks trained for 100 million steps in all three games. Breakout: 517.4 against 25.3 for 500 million steps.
- Fixed: a car network that never backed up: 0 successes in 256 episodes before, 1,024 of 1,024 after, with 3.5% of the weights changed.
Fixes what is broken, adds nothing new
Gemma 3 270M rarely loops (8.5% of answers hit the limit). No edit moved its MATH-500 score by more than about a point.
Some things stay invisible
Honest model or yes-man? From the weights alone: not found in 4 of 4 attempts. With a view of the running model: found in 3 of 3 pairs.
Sources: MIT CSAIL blog post, 2 Nov 2026 · logs on GitHub · Hugging Face re-run.
Powered by Thorsten Meyer AI
What Direct Weight Edits Could Change
If the reported results hold up across wider testing, direct weight analysis could give researchers and developers another way to audit or repair open models. Instead of relying only on a model’s outputs or running additional training, a system might identify a particular internal pattern and make a targeted change. That could help investigate behavior that ordinary test prompts fail to reveal.
The risks follow from the same capability. The source report warns that edits could also remove safety training, not just repair unwanted behavior. Model weights may therefore need stronger provenance and change controls: users and organizations would want to know who modified a model, what changed and whether safety checks were repeated afterward. The experiments do not show that these safeguards or a dependable editing process already exist.
Products Worth Considering
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
DJI Mini 4 Pro Fly More Combo Drone with DJI RC 2
Amazon Kindle Scribe (3rd Generation), 64GB, 11-inch Display, Graphite, with Premium Pen
As an affiliate, we earn on qualifying purchases.
From Training Loops to Weight Editing
Neural-network weights are numerical settings that influence how signals move through a model. Conventional training adjusts these settings repeatedly using gradient descent, which measures error and nudges weights toward better performance on a task. The resulting behavior is distributed across many numbers; the weights do not read like ordinary source code.
Researchers in mechanistic interpretability have developed methods to identify some internal structures in networks, but the source describes that work as requiring specialist tools and effort. The MIT team’s proposal is different in degree and method: ask a general-purpose AI to reason over weight values and deliberately write or change them. Its published post states the ambition that neural-network weights could be “natively optimized by LLMs that deliberately reason about their weights.” The claim remains an experimental direction, not a replacement for routine model training.
“our neural network weights will soon be natively optimized by LLMs that deliberately reason about their weights.”
— MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, as quoted in their post
How Far the Results Generalize
The evidence described so far is limited to particular models, tasks and evaluation setups. It is not clear whether the reported gains can be reproduced by independent teams, whether they persist across different prompts and benchmarks, or how often weight edits damage other capabilities. The source says the MIT team published its prompts, replies and evaluations, but the supplied material does not provide outside replication results.
The additional tests reported by ThorstenMeyerAI.com are also small in scope. The legal trial’s task count rose while its score stayed at zero, and the JSON result applies to one narrow test. The source does not specify enough about the testing procedure, baselines or statistical variation to judge whether those changes predict performance in practical use. It also remains unclear how reliably an AI can distinguish a genuine defect from a behavior that developers intended.
Replication and Safety Checks
The next test is whether other researchers can reproduce the MIT results using the published prompts and evaluations, then extend them to more models and tasks. Developers would also need to check for unintended changes in capabilities and safety behavior after any edit, rather than relying on a targeted score alone.
For now, the work is best treated as an early research method, not a substitute for established training, evaluation or model review. The MIT team’s public materials offer a basis for further scrutiny; the available source material gives no schedule for a broader release or deployment. Whether weight space programming can become a reliable tool depends on replication, clearer limits and evidence that edits can be audited.
Key Questions
What is weight space programming?
It is an approach in which an AI model analyzes or changes another neural network’s weights—the numerical settings that shape its behavior—rather than changing them through a conventional training loop.
Did the MIT experiments train models in the usual way?
The source says GPT-6 Astra’s experiments were conducted without training the target networks through gradient descent. The researchers asked it to infer, write or edit weights directly.
Did the reported edits improve every test?
No. The MIT post reports a substantial increase on one MATH-500 test, while ThorstenMeyerAI.com describes mixed additional trials. In its legal-task trial, the model answered more tasks, but the legal score remained zero; another attempt was not useful.
Could direct weight edits affect safety?
Potentially. The source report says an edit could remove safety training as well as target an unwanted behavior. The experiments described do not establish how reliably edits can be audited or how often they affect other behaviors.
Are these results independently confirmed?
The supplied material reports the MIT team’s published experiments and additional tests by ThorstenMeyerAI.com, but it does not describe independent replication. Broader confirmation remains unclear.
Source: ThorstenMeyerAI.com
