🔍 Read the full analysis: Can Home Users Edit AI Weights Legally? We Tested With Claude And GPT-6 on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
A home experiment used a Claude-written procedure to edit the weights of a local AI model, Ornith 1.5 35B-A3B. The change increased completed answers on 10 published legal benchmark tasks from 4 to 10 in the matched setup, but the legal score remained zero; the authors also report that recommended sampling settings alone let the unedited model answer 9 of 10.
A home experiment reported by ThorstenMeyerAI.com found that editing a local AI model’s weights helped it finish more legal benchmark questions, but did not improve its legal score. On the 10 published Lexassay tasks, Ornith 1.5 35B-A3B answered 10 questions after the edit, compared with 4 in a matched unedited run; its score stayed at zero. The result is a limited experiment, not evidence that home users can make a model more legally reliable by changing its weights.
The experiment was inspired by a report from four MIT CSAIL researchers on “weight-space programming,” in which one AI system edits another model’s weights. The home test did not reproduce that method: a Claude Opus 5.5 session wrote code that ran Ornith, measured a direction associated with looping, and removed that direction at half strength from the output weights of 27 of its 40 layers. The authors describe this as directional ablation, a known technique. No new legal data was used to train the model, and no AI read and edited the weights by reasoning over them directly.
In the matched comparison, the number of answered tasks rose from 4 of 10 to 10 of 10. The reported legal score remained 0. The article says the edited answers contained 55 wrong or invented citations and met 7 of 50 checked criteria, compared with 3 before the edit. Those results suggest the change helped the model complete responses, not produce sound legal analysis.
The report also describes a separate settings check: with the sampling settings recommended in Ornith’s model card and no weight edit, the model answered 9 of 10 tasks on its original setup. Its published run had answered 8 of 10, while a different unedited setup used in the matched comparison answered 4. These figures are not directly comparable because the setups differed. OpenAI’s GPT-6 Astra, running in Codex, reviewed the code, figures and claims and identified faults, according to the report. It was a review of one experiment, not a second experiment.
Infographic 01 · What we changed
One loop direction, weakened at half strength
Inspired by the MIT post, but a simpler, classic technique (directional ablation): code written by a Claude Opus 5.5 session measured one loop direction by running the model and projected it out of the weights. No training, and no AI reading weights as text.
The six steps
- Write test material that is not the benchmark: 24 legal development questions plus 8 general controls, 0 overlap with benchmark prompts
- Find the true loop onset at token level: from there, at least 90% of the next 1,500 positions repeat earlier 32-token sequences
- Run the model and measure one loop direction per layer: average internal signal inside the loop minus the reasoning just before it
- Weaken it in the weights:
W ← W − α·r·(rᵀW)on the MLP output matrices of all 256 experts and the shared expert - Bake an edited copy; the original files stay untouched and the edited weights stay internal
- Benchmark original and edit on the same local setup, graded blind by three Claude agents; figures cover the 10 published tasks
Quick test: share of pure copying
Restart 20 tokens before a known loop onset, 8 cases, 3,000 tokens. Lower is better.
The cost of the edit
- Full strength: 23 of 24 legal questions finished, but control answers broke: tokens dropped out of short formulas, and month counting went wrong
- Half strength: 23 of 24 finished, 9 of 9 checked control facts right, and one German control question was answered in Chinese once
- Before the edit: 8 of 24 legal development questions did not finish; all 8 controls did
One run per setting. The loop detector was revised three times during the experiment, so the development numbers come from a moving yardstick. The direction is a statistical description, not a single “loop neuron”.
Sources: our experiment records, 4–5 Nov 2026 (executing Claude Opus 5.5 session; review by GPT-6 Astra in Codex); Ornith-1.5-35B-A3B model card.
Powered by Thorsten Meyer AI
Finishing Answers Is Not Legal Accuracy
The distinction matters to anyone considering editing a local model for practical use. A model that stops generating repetitive text may appear more useful, but completion is not correctness. In this test, the higher answer count did not translate into a passing legal score, and the report says the citations included numerous errors or inventions.
The experiment offers a narrow demonstration that changing weights can affect a model’s behavior on a particular task. It does not establish that the technique improves legal knowledge, works across models, or is safe to use for legal decisions. Users relying on generated legal information would still need to verify statutes and decisions against authoritative sources; the report itself says the experiment is not legal advice.
For people interested in home model editing, the settings comparison is also a caution against attributing every performance change to a weight modification. Sampling choices alone were associated with a higher completion count in a separate run, though the different setups prevent a clean direct comparison. The available evidence supports a limited claim about reduced looping, not a general improvement in legal capability.
Popular right now
Philips 3300 Series Fully Automatic Espresso Machine with Milk Frother, Grinder, Hot and Iced Coffee Presets, SilentBrew, 15-Bar Pressure (E
De'Longhi Rivelia Automatic Espresso Machine with Grinder and Milk Frother, Beige
Breville Bambino Espresso Machine, Brushed Stainless Steel
As an affiliate, we earn on qualifying purchases.
From MIT’s Weight-Space Proposal
The test followed a report by MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. Their work describes an AI system inspecting another network’s weights and making direct changes without a conventional training loop. In a Qwen example, the MIT team reported that a model’s repetitive output stopped and its mathematics score rose from 7.9% to 48.1%. The home article presents its own test as a simpler, related attempt to examine whether addressing looping could help a model on a different task.
Lexassay asks models 15 questions spanning German, U.S. and European Union law, without internet access or tools. The report’s figures here cover 10 published tasks; five tasks are kept hidden, according to the site. Answers and citation checks are published there. Ornith was selected because the report says its benchmark responses showed repetitive loops, while other tested local models either lacked the relevant knowledge or could not be run under comparable conditions.
“The model finishes is not the answer is right.”
— Thorsten Meyer, describing the experiment
Limits of the Home Test
The result is based on one run per setting, a small development set and a loop detector that changed during the experiment, the article says. It also notes that three graders came from the same model family, while Claude both carried out the edit and supplied the blind graders. GPT-6 Astra provided an independent review and found faults, but that review does not remove the limitations of a single experiment.
The article does not provide evidence that the edit would work on other models or legal questions, or that the reported answer and score changes would persist across repeated runs. Comparisons between the original, recommended-sampling and matched setups are constrained because the setups were not the same. The source does not detail here how each reported error affected the final score, beyond stating that the score remained zero.
The headline question—whether home users can legally edit AI weights—also is not resolved by this experiment. The source describes a technical test and gives no legal analysis of ownership, licensing, or other rules that could govern a particular user’s modifications. Ornith’s publisher reportedly released it under the MIT license, but that fact alone does not settle every legal question about editing or using a model.
The report says the task answers, grades and citation checks are public on Lexassay, alongside a method page. Readers can inspect the published material, but the current results remain limited to the 10 public tasks; the benchmark’s five hidden tasks are not part of the figures described here.
The next useful evidence would be repeated runs under matched settings, independent grading, and results on tasks that were not used during development. Until such evidence is available, the reported finding is narrow: a particular edit was associated with fewer unfinished responses in one setup, while the benchmark’s legal score did not improve.
Key Questions
Did the experiment show that editing weights made the model better at law?
No. The report says the legal score remained zero after the edit, despite more completed answers.
What changed in the edited model?
Code written during a Claude Opus 5.5 session measured a direction associated with looping and removed it at half strength from the output weights of 27 of Ornith’s 40 layers. The report calls this directional ablation, not the direct AI weight-reading method described in the MIT work.
How many benchmark tasks did the model answer?
In the matched comparison, the edited model answered 10 of 10 published tasks, compared with 4 of 10 for the unedited model on that setup. The report says recommended sampling settings also let the unchanged model answer 9 of 10 in a separate setup, so those runs are not directly comparable.
Does the report establish whether editing AI weights is legal?
No. It reports a technical experiment, not a legal assessment of model modification. It says Ornith was released under the MIT license, but does not resolve all legal questions about editing or using AI weights.
Source: ThorstenMeyerAI.com
