🔍 Read the full analysis: Weight Space Programming In Practice: What We Learned From Our AI Experiment on ThorstenMeyerAI.com
Get hardware and tech essentials 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 report from ThorstenMeyerAI.com describes three experiments with AI systems that directly read or edited neural-network weights, without retraining the networks. The results ranged from no useful change to improved scores on narrow tests; the authors say one legal benchmark score remained at zero. The approach may help inspect or repair open models, but its limits and safety risks remain unresolved.
ThorstenMeyerAI.com reported results from three experiments testing whether AI systems can inspect and edit neural-network weights directly, without using a training loop. The authors describe mixed outcomes: one change improved performance on a narrow math test, a separate review found faults missed by green tests, and an attempted legal-task improvement left the reported legal score at zero.
The approach, called weight space programming, asks a general-purpose AI model to reason about another network’s numerical weights rather than train it through repeated examples. A November 2 MIT CSAIL blog post by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab described experiments in which OpenAI’s GPT-6 Astra read, wrote and edited network weights. The MIT team said it published its prompts, responses and evaluations on GitHub.
In the experiments summarized by ThorstenMeyerAI.com, the authors tried the method three times with two different AI systems. They report that Claude changed a local model so it answered all 10 published Lexassay legal tasks, up from four, but the legal score remained zero. A second AI review identified faults that existing tests had not caught. A separate edit to one matrix raised a result on a narrow JSON test from 49 to 70 out of 120.
The report also describes an edit to the open Qwen3.5-0.8B model: removing one direction of a repeating loop affected 0.016% of its weights, while its reported MATH-500 score rose from 7.9% to 48.1%. The authors say the method did not help a model that lacked the loop behavior in the first place. These are reported experiment results, not evidence that the technique works reliably across models or 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
What Direct Weight Editing Could Change
If these results can be repeated and extended, direct weight editing could offer another way to inspect or repair open models without running a new training process. That could matter to teams that need to understand model behavior, correct a specific failure, or assess a model before deployment. The report’s examples, however, are narrow tests; they do not establish that an AI can safely or consistently edit large production systems.
The same ability could create risks. The report warns that an edit may remove safety training as well as unwanted behavior. Organizations using open model weights would have reason to track where weights came from and test models after modifications. The experiments do not show that the technique has been used to cause harm, but they underline that changing weights can alter model behavior in ways that require review.
Products Worth Considering
Bose Lifestyle Ultra Wireless Home Speaker, White Smoke
Amazon Kindle Paperwhite (2024, 16 GB) – 7-inch Glare-Free Display, Black
Bose Smart Soundbar 700 with Alexa, Black
As an affiliate, we earn on qualifying purchases.
From Training Loops to Weight Reasoning
Most neural networks acquire their behavior through training: an algorithm compares predictions with desired results and adjusts many weights in small steps, a process commonly called gradient descent. The finished weights are numerical values, not readable instructions. Researchers in mechanistic interpretability have developed ways to identify some internal patterns, but the report says this has typically required specialist methods and considerable work.
Weight space programming proposes that a general-purpose AI can reason about those values directly. The MIT CSAIL post described both reading weights to infer a network’s behavior and writing or editing them without a training loop. ThorstenMeyerAI.com’s November 9 update says its figures refer to the 10 published Lexassay tasks and that the runs are public on lexassay.com.
“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 November 2 blog post
Limits of the Reported Tests
The report provides results for a small set of experiments, but does not establish how reproducible the improvements are across different models, tasks or independent evaluators. The supplied source does not give enough detail to assess all benchmark procedures, such as the full scoring methodology for the legal tasks or the JSON test. The legal score remained zero despite the reported increase in tasks answered.
It is also unclear how well direct editing scales to larger models, whether the changes persist across repeated tests, or how often edits create new failures while fixing an existing one. The report’s claim that the method can write game-playing networks from zero weights is attributed to the MIT work; the experiments described here do not settle whether such results generalize beyond the tested settings.
Public Runs and Further Replication
The authors say the 10 Lexassay task runs are public on lexassay.com, allowing readers to examine the reported work. The MIT team also said it published prompts, model replies and evaluations on GitHub. Independent attempts to reproduce the changes, inspect the test setup and check for side effects would help clarify how much confidence to place in the results.
The next substantive test is whether direct edits can produce repeatable gains without weakening other model behavior. Until broader evaluations are available, the findings support treating weight space programming as an experimental method rather than a dependable repair process.
Key Questions
What is weight space programming?
It is an experimental approach in which an AI model reasons about another neural network’s weights—the numerical parameters that shape its behavior—and may inspect or edit them without a conventional training loop.
Did the legal-task experiment improve the model’s score?
The report says the edited model answered all 10 published tasks instead of four, but its legal score remained at zero. The task count and score are different measures, and the reported result does not show a positive legal score.
What changed in the Qwen3.5-0.8B experiment?
According to ThorstenMeyerAI.com, an edit affecting 0.016% of the weights removed one direction of a repeating loop. The reported MATH-500 score rose from 7.9% to 48.1%; the source describes this as a result from a specific experiment, not a general performance guarantee.
Can the method repair any model failure?
No such broad ability is established. The report says an edit did not help when the model lacked the behavior being targeted. It characterizes the method as potentially fixing what is broken, rather than adding something that is missing.
What safety concerns did the report raise?
The authors warn that edits could remove safety training as well as unwanted behavior. The experiments described do not establish how often that might happen, so edited models would need careful testing before use.
Source: ThorstenMeyerAI.com
