🔍 Read the full analysis: Weight-Space AI Editing At Home: A Safe, Zero-Liability Experiment 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 adapted a weight-editing idea to reduce looping in the local AI model Ornith 1.5 35B-A3B. On the 10 published Lexassay tasks, the edited model completed all 10, compared with 4 on the same setup before editing, but its legal score remained zero and its answers contained 55 wrong or invented citations. The test was limited, and the article says the unchanged model also answered 9 of 10 when run with the sampling settings recommended in its model card.
A home experiment reported November 9 found that a weight edit helped local AI model Ornith 1.5 35B-A3B finish more legal benchmark tasks, but did not improve its legal score. On the same setup, the model completed 10 of 10 published Lexassay tasks after the edit, compared with 4 of 10 beforehand; its score stayed at zero, with 55 wrong or invented citations in its edited answers.
The experiment was described by Thorsten Meyer on ThorstenMeyerAI.com and drew inspiration from a study by four MIT CSAIL researchers on what they call “weight-space programming.” The home test did not reproduce the MIT method: rather than having an AI reason directly over another model’s weights, a Claude Opus 5.5 session wrote code to run Ornith, identify a direction associated with looping, and remove that direction at half strength from the output weights of 27 of the model’s 40 layers. The report says there was no training or new legal data.
On the 10 published benchmark tasks, the edited model’s answers met 7 of 50 checked criteria, up from 3, but its legal score remained zero. The report says its 10 edited answers included 55 incorrect or invented citations. A separate run of the unchanged model using the sampling settings recommended in its model card answered 9 of 10 tasks without a weight edit, suggesting that generation settings may also affect whether it finishes.
The account says OpenAI’s GPT-6 Astra, running in Codex, reviewed the code, figures and claims and identified faults. It was an independent review, not a second experiment. The report also discloses that Claude carried out the edit and Claude agents performed blind grading. The experiment is presented as an AI test, not legal advice.
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 practical distinction for people testing local AI systems: a model that produces a complete answer has not necessarily produced a reliable one. Here, the edit appears to have addressed a failure to finish, while leaving the model’s weak legal performance unchanged. The volume of incorrect or invented citations makes completion a poor substitute for checking factual accuracy against primary sources.
The comparison also cautions against attributing every improvement to a weight change. In the report, the unedited model answered 9 of 10 tasks under recommended sampling settings, compared with 4 of 10 on the setup used for the direct before-and-after comparison. That difference does not establish which setting is better overall, but it shows why consistent test conditions matter when evaluating an edit.
For readers considering local models for legal research, the test offers no evidence that this edit makes the system suitable for legal work. It instead suggests that a narrow intervention can change one observable behavior without supplying missing expertise. The benchmark’s citation checks and published answers make the reported failures visible, but the small test does not establish how the method would perform on other questions or models.
Products Worth Considering
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Raspberry
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Bluetooth 5.3, USB-C Charging, IPX4, Black
Apple Watch SE 3 GPS 40mm Smartwatch
As an affiliate, we earn on qualifying purchases.
From MIT Method to Local Test
The MIT CSAIL work described in the source material involved Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. In their reported Qwen example, GPT-6 Astra examined the model’s weights directly and identified an internal direction linked to looping. The researchers reported subtracting that direction from 0.016% of the weights. Their account characterized the approach as removing an obstacle that concealed an existing ability, rather than adding an ability the model did not have.
Meyer’s test used a related but established technique, directional ablation: run a model, record internal signals, calculate a direction, and project it out of selected weights. The report says the local model was selected because it showed looping in benchmark runs. Two other local models reportedly failed through lack of knowledge rather than looping; another model that looped was too large for the experiment machine and used a different token budget.
Lexassay asks models 15 questions about German, U.S. and European Union law, without internet access or tools, and checks cited statutes and court decisions against primary sources. Ten tasks are published and five are hidden; all figures in this report concern the published tasks. The article says the experiment’s runs, answers, grades and citation checks are public on lexassay.com.
“The edit repairs finishing, not legal knowledge.”
— Thorsten Meyer, describing the experiment’s result
Limits of the Benchmark Comparison
The findings come from one run per setting and 10 published tasks, which the report describes as a small development set. The loop detector changed during the experiment, and three graders came from the same model family. Those conditions limit how confidently the results can be generalized or how independently the answers were evaluated.
The reported figures also include different runs that are not directly comparable. The unchanged model answered 8 of 10 tasks in its published run, but 4 of 10 on the setup used for the edited-model comparison; it answered 9 of 10 with the model-card sampling settings. The source does not establish why those results differed or isolate how much each setting contributed.
The article says GPT-6 Astra found real faults in the review, but does not specify every fault in the supplied material. It is also unclear whether the edited model would retain its behavior across repeated runs, the five hidden Lexassay tasks, other legal questions, or different hardware and settings. No broader accuracy or safety conclusion can be drawn from this experiment alone.
Further Runs Must Test Reliability
The experiment’s runs and method are available through Lexassay, providing material for readers to inspect the answers, grades and citation checks. The next useful evidence would come from repeated runs under matched settings, with the edit and unedited control evaluated on the same tasks and generation configuration. Reporting results on the five hidden tasks would also help test whether the observed behavior extends beyond the published development set.
Until such evidence is available, the reported result is limited: the edit was associated with more completed answers on one setup, but did not improve the legal score. The article does not announce a follow-up test or provide a timetable for one.
Key Questions
What did the weight edit change?
It removed a looping-related direction at half strength from the output weights of 27 of Ornith 1.5 35B-A3B’s 40 layers. The report says the method used code to measure the model’s internal signals, not an AI directly reading and editing the weights by reasoning over them.
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 edited answers included 55 wrong or invented citations.
How many benchmark tasks were tested?
The figures in the article refer to 10 published Lexassay tasks. The benchmark has 15 questions in total, with five kept hidden, according to the source.
Was the experiment independently reviewed?
The source says OpenAI’s GPT-6 Astra, running in Codex, reviewed the code, numbers and claims and found faults. It was one reviewer of the experiment, not a competing test; Claude carried out the edit and Claude agents performed the blind grading.
Does this show that local AI is safe to use for legal advice?
No. The report explicitly says the experiment is not legal advice. The small test found serious citation problems and does not establish that the model is reliable for legal research or advice.
Source: ThorstenMeyerAI.com
