🔍 Read the full analysis: Weight-Space AI Editing At Home: A Safe, Zero-Liability Test 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 home experiment edited the weights of a local AI model to reduce looping. On 10 published Lexassay tasks, it completed all 10, up from 4 on the same setup, but its legal score remained zero and the answers included 55 wrong or invented citations. The test was limited and does not establish that weight editing improves legal accuracy.
A home test of weight editing helped the local AI model Ornith 1.5 35B-A3B finish more legal benchmark tasks, but did not improve its legal score, according to a report published by ThorstenMeyerAI.com and updated November 9. On the same experimental setup, the edited model answered 10 of 10 published tasks, compared with four before editing; its score remained zero, underscoring that finishing an answer is not the same as getting the law right.
The experiment followed a report by four MIT CSAIL researchers on “weight-space programming,” in which an AI system directly examines and modifies another model’s weights. The home test was a related but different method: a Claude Opus 5.5 session wrote code to identify a direction associated with looping and remove it at half strength from the output weights of 27 of the model’s 40 layers. The model was not trained on new data, and an AI did not read its weights as text.
According to the report, the edited model completed all 10 published Lexassay tasks, while scoring zero on the legal benchmark. Its answers met seven of 50 checked criteria, up from three, but included 55 wrong or invented citations. Lexassay checks legal citations against primary sources. The benchmark covers questions about German, U.S. and EU law; five additional tasks are hidden and were not part of the reported results.
The authors also tested the unedited model with sampling settings recommended in its model card. On its original setup, it answered nine of 10 tasks without a weight edit. That result complicates the comparison: the unedited model answered eight of 10 in a separately published run, but only four of 10 on the setup used for the direct edited-versus-unedited comparison. The report says those two local setups are not directly comparable.
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
Completion Is Not Legal Accuracy
The result illustrates a narrow but important distinction for people testing local AI systems: reducing repeated or unfinished output can make a model appear more usable without improving its underlying subject knowledge. In this case, the edit addressed the behavior that kept the model generating loops; it did not produce reliable legal analysis or citations.
That distinction matters in high-stakes fields. A model that returns a complete answer may seem more capable, yet the reported citation errors show why completion alone is not a measure of correctness. Lexassay’s checks against primary legal sources provide a separate test of whether citations and claims hold up. The report itself says the experiment is not legal advice.
The test also offers limited evidence about weight editing more broadly. It involved one model, one run per setting and 10 published tasks. It does not establish that the technique will work on other models, improve performance on hidden tasks, or yield dependable answers outside this particular setup.
Products Worth Considering
Sony WH-1000XM6 Wireless Noise-Canceling Headphones, Black
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black
Logitech MX Ergo S Advanced Wireless Trackball Mouse USB-C – Graphite
As an affiliate, we earn on qualifying purchases.
From MIT’s Method to a Local Test
The experiment was prompted by a MIT CSAIL report from Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. Their work described weight-space programming and included a Qwen model whose looping stopped after an edit; the report said its maths score rose from 7.9% to 48.1%. The MIT researchers’ method involved an AI reasoning directly about another network’s weights. The home test did not reproduce that procedure.
Instead, the report describes a more established technique called directional ablation: run a model, record internal signals, calculate a direction associated with a behavior, then remove that direction from selected weights. The author selected Ornith because it showed repeated output on the benchmark. According to the report, on the 10 published tasks, two responses did not finish before a 62,000-token limit; near the end of the cut-off transcripts, only 2% to 13% of lines were distinct.
The report says other local models were not suitable comparisons: some failed without looping, while DeepSeek V4 Flash was too large for the test machine and used a different token budget. The experiment’s results and answers are described as publicly available on Lexassay.com, alongside the grading and citation checks.
“The edit repairs finishing, not legal knowledge.”
— Thorsten Meyer, author of the experiment report
Limits of the Ten-Task Test
The results are preliminary. The report describes one run per setting, a small development set and a loop detector that changed during the experiment. It also says three graders came from the same model family, while the editing work was carried out by Claude and Claude agents performed blind grading. OpenAI’s GPT-6 Astra, running in Codex, reviewed the code, figures and claims and found faults, according to the author. That was an independent review, not a second experiment.
The difference between the unedited model’s results on separate setups also makes the size of the editing effect difficult to judge beyond the stated same-setup comparison. The published set contains only 10 of Lexassay’s 15 questions; the other five remain hidden. The report does not establish how the edit would perform on those tasks, on other hardware, or across repeated runs. Nor does the available source material specify whether the reported citation errors were independently adjudicated beyond the described benchmark checks.
Public Runs and Further Checks
The immediate next step is scrutiny of the public experiment materials. Readers can compare the original, model-card sampling, edited-model and edit-control runs on Lexassay, including answers, grades and citation checks, according to the report. Those records may help clarify how sampling settings and the weight edit each affected completion and answer quality.
Further testing would need repeated runs under matched settings, consistent evaluation methods and checks on tasks not used to guide the experiment. Until such results are available, the finding should be read narrowly: this particular edit was associated with more completed answers on the same 10-task setup, while the legal score stayed at zero.
Key Questions
What did the weight edit change?
Code identified a direction associated with looping and removed it at half strength from the output weights of 27 of Ornith’s 40 layers. The report describes this as directional ablation, not the direct weight-reading method used in the MIT work.
Did the edited model become better at law?
Not according to the reported score. It completed more tasks on the comparison setup, but its legal score remained zero. The report says its 10 answers included 55 wrong or invented citations.
How many benchmark tasks were tested?
The updated report’s figures cover 10 published Lexassay tasks. The benchmark has five additional hidden tasks, which were not included in these results.
Was the home experiment a replication of the MIT study?
No. It tested a related idea with directional ablation. The MIT report described an AI reasoning directly over model weights; the home test used code to measure internal signals and remove a calculated direction from selected weights.
Does this show that weight editing is safe or reliable?
No broad conclusion follows from this test. It involved one model and one run per setting, and the author reported methodological limits. The source describes an AI experiment, not legal advice.
Source: ThorstenMeyerAI.com
