🔍 Read the full analysis: We Tried Weight-Space Programming At Home: Claude Edited A Local Model, GPT-6 Astra Checked The Work, And The Legal Score Stayed At Zero 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 home experiment used code written with Claude Opus 5.5 to remove a looping direction from the weights of a local model, Ornith 1.5 35B-A3B. On the same setup, it answered all 10 published Lexassay tasks instead of 4, but its legal score remained zero; an independent review by OpenAI’s GPT-6 Astra found faults in the experiment.
A home experiment reported on November 9, 2026, found that editing the weights of a local AI model helped it complete more legal benchmark tasks, but did not improve its legal score. The model, Ornith 1.5 35B-A3B, answered all 10 published Lexassay tasks after the edit, compared with 4 of 10 on the same experimental setup before it; its score remained zero.
The experiment followed a report from four MIT CSAIL researchers on “weight-space programming,” in which one AI system edits another model’s weights. The home test was a related but different method: a Claude Opus 5.5 session wrote code to run Ornith, identify an internal direction associated with looping, and remove that direction at half strength from the output weights of 27 of the model’s 40 layers. The researchers did not have an AI read and edit the weights directly, and they did not train the model or add legal data.
The reported comparison covers the 10 published Lexassay tasks, not the benchmark’s five hidden tasks. The edited model completed all 10, but its answers met only 7 of 50 checked criteria, up from 3, and contained 55 wrong or invented citations. The report says the legal score stayed at zero. A separate test of the unchanged model using sampling settings recommended in its model card completed 9 of 10 tasks without any weight edit, illustrating that setup choices also affected completion.
OpenAI’s GPT-6 Astra, running in Codex, reviewed the code, figures and claims and identified faults, according to the report. The author describes this as one experiment and one independent review, not two competing experiments. The runs, answers and citation checks are published on Lexassay’s website.
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 Did Not Improve Legal Accuracy
The experiment separates two problems that can look similar in a benchmark: a model may fail to finish an answer, or it may finish while lacking the knowledge to answer correctly. Here, removing the looping direction improved task completion on one setup, but the benchmark found no improvement in the model’s legal score. The result cautions against treating a complete response as evidence of a reliable one.
That distinction matters for legal AI, where a confident answer with fabricated or incorrect citations can mislead users. Lexassay checks citations against primary legal sources, and the edited answers still performed poorly on its criteria. The experiment is a report about model behavior, not evidence that Ornith is suitable for legal work and not legal advice.
The outcome also narrows what this particular edit appears to have done: it addressed a tendency to repeat itself, rather than supplying legal knowledge. That is consistent with the experiment’s stated hypothesis, but the small number of tasks and runs means the result should not be treated as a broad finding about weight editing or other models.
Products Worth Considering
Ledger Flex Crypto Wallet
Apple MacBook Air 15-inch (2026) with M5 Chip, 16GB Memory, 1TB SSD, and Liquid Retina Display – Sky Blue
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and 12MP Center Stage Camera (2026, Silver)
As an affiliate, we earn on qualifying purchases.
From MIT’s Edit to a Home Test
The home experiment was prompted by a report from MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. Their work described an AI editing model weights directly. In a cited Qwen example, their account said GPT-6 Astra identified a direction associated with looping and removed it; the Qwen model’s reported maths score rose from 7.9% to 48.1%. Those figures belong to the MIT example and are not results from the home test.
The home test instead used a more established procedure called directional ablation: run the model, record internal signals, calculate a direction and project it out of selected weights. Ornith was chosen because the report describes looping in its benchmark responses. On the published tasks, two answers did not arrive before the 62,000-token limit; late portions of the cut-off transcripts reportedly contained few distinct lines. Other local models were not selected because, according to the author, they lacked legal knowledge without the same looping problem, or were too large for the test machine.
Lexassay asks 15 questions about German, U.S. and EU law, with no internet access or tools, and checks cited statutes and court decisions against primary sources. The experiment’s figures cover only 10 published tasks; five remain hidden. The article says the unchanged model completed 8 of 10 in a previously published run, but only 4 of 10 in the separate setup used for the edited comparison. Those setups are not directly comparable.
“The model finishes is not the answer is right.”
— Thorsten Meyer, the experiment’s author
Limits of the Ten-Task Comparison
The result is based on one run per setting and only the 10 published benchmark tasks. The report also says the development set was small and that its loop detector changed during the experiment. Three graders came from the same model family, while the author disclosed that Claude carried out the edit and Claude agents performed blind grading; GPT-6 Astra’s independent review identified faults but does not remove those limitations.
The different completion rates across the unchanged model’s runs—8 of 10 in one published setup, 4 of 10 in the comparison setup, and 9 of 10 with model-card sampling settings—make it difficult to isolate the edit’s effect from other configuration choices. The available account does not establish how the method would perform on the five hidden tasks, other legal benchmarks, different hardware or repeated trials. It also does not show that the reported completion gain would persist across other models.
The immediate next check is whether the edited model’s apparent completion improvement holds on Lexassay’s five hidden tasks, which the benchmark keeps out of public tuning. The report does not give results for those tasks or announce a date for additional testing.
Further runs would need to compare the edited and unchanged model under matched sampling settings, repeat each condition and report both completion and citation accuracy. Until such evidence is available, the published experiment supports a limited finding: this edit helped Ornith finish more answers in one test setup, while its measured legal performance remained at zero.
Key Questions
What did the weight edit change?
The experiment identified a direction associated with looping and removed it at half strength from selected output weights in 27 of Ornith’s 40 layers. Claude Opus 5.5 wrote the code for this procedure; the AI did not directly read and edit the weights.Did the edited model become better at law?
The report says its legal score remained zero. It met 7 of 50 checked criteria, up from 3, and produced 55 wrong or invented citations across the 10 published tasks.How many benchmark tasks did the model complete?
On the same experimental setup, the edited model answered 10 of 10 published tasks, compared with 4 of 10 before the edit. These are completion counts, not measures of legal accuracy.Was this the same method used in the MIT report?
No. The home test used directional ablation: code ran the model, measured an internal direction and removed it from selected weights. The MIT report described an AI reasoning about another model’s weights and editing them directly.How strong is the evidence?
It is limited: the report covers one run per setting and 10 published tasks, with five benchmark tasks hidden. An independent GPT-6 Astra review found faults, and the author notes that the model’s results varied across different setups.Source: ThorstenMeyerAI.com
