🔍 Read the full analysis: Weight Space Programming: A Question Of Whether AI Can Edit AI 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 CSAIL researchers report that GPT-6 Astra could identify what small networks did from their weights, write game-playing networks without training, and make a narrow edit that improved one benchmark score. Follow-up experiments described by ThorstenMeyerAI.com found mixed results, including a legal-task change that did not improve the legal score. The method remains experimental, and its effects on safety and broader models are unresolved.
MIT CSAIL researchers report that OpenAI’s GPT-6 Astra could read, write and edit other neural networks by working directly with their numerical weights, without using a training loop. The experiments, described in a November 2 post by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, suggest a possible new way to inspect or repair models, but the published results concern specific small networks and do not establish a general-purpose method for improving AI.
In the MIT team’s tests, Astra identified the task behind four small game-playing networks from their weights alone. The networks had the same architecture and 17,253 weights each, and were trained for grid versions of Breakout, Pong, Pac-Man and Freeway. The team also reported that Astra located copying heads—attention patterns that repeat text based on earlier text—in three small language models.
The team further says Astra generated game-playing networks from all-zero weights, without training, that beat comparison networks trained for 100 million steps in each of three games. In an editing test on Qwen3.5-0.8B, Astra removed a loop-related direction by changing 0.016% of the weights. The reported MATH-500 score rose from 7.9% to 48.1%. The article says the change addressed an existing behavior, rather than adding a capability the model lacked.
ThorstenMeyerAI.com reports separate follow-up trials using two AI systems. One edit made a local model answer all 10 published Lexassay legal tasks rather than four, but its legal score remained zero. Another AI review found faults that green tests had missed, while a single-matrix edit increased a narrow JSON test result from 49 to 70 out of 120. These are site-reported experiments, not results independently established by the MIT post.
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 Repair
If the approach proves reliable, directly inspecting weights could give researchers and developers another way to diagnose open models, test how a behavior is represented, or make targeted corrections. Today, model behavior is usually changed through training or fine-tuning, processes that adjust many parameters through repeated optimization. Weight space programming instead treats the parameter list itself as something an AI can examine and edit deliberately.
The results also point to a safety concern. The source article warns that an edit intended to remove an unwanted behavior could also remove safety training. That is a risk to investigate, not evidence that the reported experiments caused a safety failure. For organizations using open models, the practical implication is that model changes and the provenance of weight files may need careful review; a modified model should not be assumed to retain the properties of its original version.
The narrow and mixed outcomes matter. A benchmark gain on one task does not show broad improvement, and the legal trial’s increase in answered tasks did not produce a higher legal score. The work is best understood as an early demonstration of what direct weight reasoning might do, not a replacement for evaluation or training.
Products Worth Considering
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
Bose QuietComfort Wireless Noise-Cancelling Headphones, Sandstone
LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz
As an affiliate, we earn on qualifying purchases.
From Training to Direct Edits
Neural-network weights are numerical parameters that determine how signals move through a model. Conventional training adjusts them repeatedly based on errors, usually through gradient descent. Although a trained model may contain billions of weights, their meaning is distributed across the network; a list of numbers does not read like a set of labeled instructions.
Researchers in mechanistic interpretability have studied how particular computations appear inside networks. The source article points to a 2021 Anthropic study that found some algorithms could be detected directly from weights, while noting that this work relied on specialist methods. Weight space programming extends the idea by asking whether a general-purpose language model can reason about those parameters and make changes itself.
The MIT researchers say they published their prompts, model replies and evaluations on GitHub. The source article’s November 9 update says its own runs are public on Lexassay.com and that its figures refer to 10 published tasks. Those disclosures let others inspect the stated procedures, but do not by themselves establish that results will transfer to different models or evaluations.
“Our neural network weights will soon be natively optimized by LLMs that deliberately reason about their weights.”
— Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, in the MIT CSAIL post as quoted by ThorstenMeyerAI.com
Limits of the Published Tests
The supplied source describes selected demonstrations, not a broad comparison across model sizes, tasks or independent research teams. It does not establish how consistently an AI can identify useful edits, whether results survive different prompts or model versions, or how often a change that helps one benchmark harms another. The reported scores also should not be read as proof of general reasoning gains.
The source material does not provide enough detail to resolve how the edited models performed across a wider set of safety evaluations, or whether the same changes would work outside the tested settings. It also does not show that direct weight editing can replace training for building capable models. The experiments’ reproducibility and transfer to other systems remain open questions.
Replication and Broader Evaluation
The immediate next step is for other researchers to inspect the published prompts, outputs and evaluations, then try to reproduce the results with the same models and comparable tests. Broader evaluations would need to check both target-task performance and unintended changes, including effects on safety behavior. The source article says its own runs are public on Lexassay.com, giving readers a separate set of experiments to examine.
No further milestone or timeline is specified in the supplied material. Until independent replications and wider testing are available, the reported work remains a proof of concept: evidence that an AI can sometimes reason about and modify another network’s weights, with the reliability and consequences still to be established.
Key Questions
What is weight space programming?
It is an approach in which an AI model examines or changes another neural network’s numerical weights directly, rather than relying on a training loop to update them.
What did the MIT researchers report?
They report that GPT-6 Astra identified the tasks of four small game networks, found copying heads in three small language models, created game-playing networks without training, and made a targeted edit associated with a higher MATH-500 score in one model.
Did the experiments prove that AI can generally improve other AI models?
No. The results cover specific tests. The source material does not establish that the method works reliably across models, tasks or independent evaluations.
Did the follow-up legal-model edit improve its score?
According to ThorstenMeyerAI.com, the model answered all 10 published legal tasks after the edit, compared with four before, but its legal score remained zero.
Could editing weights affect model safety?
The source article identifies removal of safety training as a potential risk of direct edits. The supplied material does not report that the experiments caused such a failure, and wider safety testing remains necessary.
Source: ThorstenMeyerAI.com
