🔍 Read the full analysis: We Tried An AI That Rewrites AI—Here’s What Weight Space Programming Means 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
MIT researchers reported that GPT-6 Astra could identify patterns in neural-network weights, create small networks without a training loop and edit a model to improve a specific benchmark. In a separate set of experiments reported by ThorstenMeyerAI.com, edits produced mixed results, including no gain on a legal benchmark; the findings are early and do not establish a general replacement for model training.
Researchers at MIT CSAIL have reported experiments in which GPT-6 Astra read and edited neural-network weights—the numerical settings inside a model—without using a training loop. The work, described as weight space programming, included identifying what small networks did, writing networks from scratch and making a targeted change to a language model; separate tests by ThorstenMeyerAI.com found mixed results, including one benchmark improvement and one edit that did not improve its score.
In its November 2, 2026, blog post, the MIT team said it gave GPT-6 Astra the weights of small neural networks as text and asked it to infer their functions. The researchers report that the model identified the games played by four small networks and detected copying-related attention heads in three small language models. The team also said Astra generated arcade-game networks from all-zero weights, without training, that outperformed networks trained for 100 million steps across the three games in that experiment.
The team’s reported editing result involved Qwen3.5-0.8B, a small open model. Astra removed one looping direction through an edit affecting 0.016% of the weights. The source reports that the model’s score on MATH-500 rose from 7.9% to 48.1%. That result concerns a particular change and benchmark; it does not establish that the method can broadly improve models or replace conventional training.
ThorstenMeyerAI.com said it also ran three experiments using two different AI systems. One edit made a local model answer all 10 published legal tasks rather than four, but its legal score remained zero. A second AI review identified faults that existing green tests had missed. In a separate, narrow JSON test, changing one matrix raised the result from 49 to 70 out of 120. The site says the runs are public on lexassay.com and that its figures were updated November 9 to refer to 10 published tasks.
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
A New Route to Model Audits
If the reported results hold up across broader testing, weight space programming could give researchers a way to inspect or alter models without repeating a full training process. That could help developers investigate why an open model behaves in a particular way, find faults that ordinary tests miss, or make a narrowly targeted repair. For teams working with models they did not train themselves, the possibility of examining the weights directly could add another form of technical review.
The reported limits matter just as much. The legal-task experiment increased the number of answered tasks but left the score at zero, illustrating that producing an answer is not the same as solving the task. And a method that changes model behavior could be used to remove safety-related behavior as well as to improve a capability. ThorstenMeyerAI.com warns that edits of this kind could strip safety training from a model; the source does not report that this occurred in its experiments. Any practical use would need checks on what changed and whether the resulting model remains reliable.
Popular right now
Breville Barista Touch Impress Espresso Machine, Noir (BES881NRE)
Shark NV752 Rotator Powered Lift-Away TruePet Upright Vacuum, Bordeaux
KitchenAid Artisan 5-Quart Tilt-Head Stand Mixer, Blue Velvet
As an affiliate, we earn on qualifying purchases.
From Training to Direct Edits
Most neural networks acquire their weights through training. A model makes predictions, an algorithm measures errors and adjusts the weights repeatedly, usually using a method such as gradient descent. The resulting numbers encode behavior across the network, rather than labeling each weight with a human-readable purpose.
Researchers in mechanistic interpretability have developed ways to identify some structures within networks, but the source describes that work as requiring expert analysis and specialized tools. Weight space programming proposes using a general-purpose AI model to reason about weights directly. The MIT team’s post, by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, presents reading and editing weights as steps toward that approach. It is a research demonstration, not evidence that neural-network weights are now generally transparent or easy to program.
The MIT team says it used GPT-6 Astra without training it for these tasks and published prompts, responses and evaluations on GitHub. The independent tests described by ThorstenMeyerAI.com are separate from the MIT experiments and should not be treated as a replication of every result.
“Our neural network weights will soon be natively optimized by LLMs that deliberately reason about their weights.”
— MIT CSAIL research team, in its November 2, 2026, blog post
How General Are These Results?
The reported experiments are limited in scope, and it is not yet clear whether the approach works reliably on larger or more diverse models. The source material does not provide enough detail to independently assess every evaluation, the statistical uncertainty around the benchmark changes, or how well the edits would hold up on tests not used in the experiments. The independent tests described by ThorstenMeyerAI.com are small and produced mixed outcomes.
It also remains unclear how easily a model’s edits can be inspected, reversed or validated before deployment, and whether the technique can reliably distinguish a useful repair from a harmful change. The reported MATH-500 gain is a result for one model and one benchmark, not proof of a general improvement. The available material does not establish that the method can safely replace training or that it can explain all of a model’s behavior.
Further Tests and Independent Checks
The MIT team has made its prompts, replies and evaluations available on GitHub, while ThorstenMeyerAI.com says its own runs are public on lexassay.com. Those materials give other researchers a basis for checking the reported work. The next useful evidence would include independent replications across more architectures and tasks, along with evaluations of whether edits improve performance without damaging other capabilities or safety behavior.
For now, weight space programming is best treated as an early research method for probing and making targeted changes to model parameters. Developers considering it for open models would need to verify each edit against task-specific and safety tests. The source material does not identify a scheduled follow-up study or a deployment timetable.
Key Questions
What is weight space programming?
It is an approach in which an AI model reasons directly about another neural network’s numerical weights to identify what the network does or to change it, rather than relying on a conventional training loop for that task.
What did the MIT researchers report?
The MIT team said GPT-6 Astra identified the functions of small networks, detected copying-related structures in small language models and wrote arcade-game networks from zero weights. It also reported a targeted edit to Qwen3.5-0.8B that raised its MATH-500 score from 7.9% to 48.1%.
Did the independent tests show consistent improvements?
No. ThorstenMeyerAI.com reported mixed results: one legal-task edit increased the number of answers but left the score at zero, while a separate narrow JSON test rose from 49 to 70 out of 120.
Does this mean AI models no longer need training?
No. The findings describe experiments in reading and changing weights without a training loop. They do not show that the method can broadly replace training or produce reliable results across models and tasks.
Could editing weights affect a model’s safety behavior?
Potentially, according to the source article, which warns that edits could remove safety-related behavior as well as repair a model. The reported experiments do not establish how often that would happen or how to prevent it, so any use would require safety checks.
Source: ThorstenMeyerAI.com
