🔍 Read the full analysis: What Happened When We Tried Weight-Space Programming At Home 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 used a weight edit to reduce looping in the local Ornith 1.5 35B-A3B model. In one comparison, completed answers rose from 4 of 15 to 12 of 15, but its legal score remained zero; changing sampling settings without editing weights produced 13 completed answers.
A home test of a local AI model found that editing its weights could help it finish more legal-benchmark answers, but did not improve its legal score: completion rose from 4 of 15 to 12 of 15, while the score stayed at zero. The experiment, reported by Thorsten Meyer on November 5, adapted a technique related to MIT researchers’ work but did not reproduce their method, and its results came from only one run per setting.
The test targeted Ornith 1.5 35B-A3B, a model whose answers sometimes ran in loops until reaching the experiment’s 62,000-token limit. The team recorded the model’s internal signals, identified a direction associated with looping, then removed that direction at half strength from output weights in 27 of 40 layers. It did not train the model or add legal material.
In the reported same-setup comparison, the edited model completed 12 of 15 answers, up from four. Yet it met only 7 of 75 checked legal criteria, compared with three before the edit, and its legal score remained zero. The 12 completed answers contained 62 incorrect or invented citations, an average of 5.2 per answer, according to the report.
A no-edit comparison complicates the result: with sampling settings recommended by the model’s card, the original model completed 13 of 15 answers and still scored zero. The author says an independent review by OpenAI’s GPT-6 Astra, running in Codex, identified faults in the code, figures and claims. The review was a check of this 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, 15 tasks, graded blind by three Claude agents
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
More Finished Answers, Same Legal Score
The test illustrates a distinction that matters for anyone evaluating AI: a model can produce a complete-looking response without producing a correct one. Here, the weight edit appeared to address a failure to finish, not the model’s limited legal knowledge. The citation errors make the difference concrete: more output did not translate into dependable legal answers.
The no-edit result also shows why a technical intervention needs a fair baseline. Under the recommended sampling settings, the unedited model finished one more answer than the edited version. That does not prove the edit had no effect: the runs were limited, and the setup changed during testing. It does mean this report cannot establish that editing weights outperforms simpler configuration changes.
The findings are an experiment report, not legal advice or evidence that the technique generalizes. The benchmark covers 15 questions across German, US and EU law, while the edit was tested on one model. Readers should treat the result as a narrow observation about completion behavior, not as a method for improving legal accuracy.
Products Worth Considering
Apple MacBook Air 15-inch with M5 chip, 24GB RAM, 512GB SSD, Midnight
Samsung Galaxy Tab A11+ (6 GB RAM, 128 GB storage, expandable) – Gray
Apple MacBook Air 13.6-inch with M5 Chip, 16GB Memory, 512GB SSD, Sky Blue
As an affiliate, we earn on qualifying purchases.
How the Home Test Differed From MIT
The test followed a report by four MIT CSAIL researchers: Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. Their work describes “weight-space programming,” in which a general-purpose model reasons about another network’s weights and edits them directly, without a training loop. Their reported Qwen example stopped looping, and its mathematics score increased from 7.9% to 48.1%, according to the source article.
Meyer explicitly says the home test did not repeat that approach. Claude Opus 5.5 wrote code that ran Ornith, measured an internal direction and removed it from selected weights—a technique known as directional ablation. The article describes the MIT work as an inspiration and the home test as a simpler, related experiment.
Ornith was selected because the report says it was the local model in the team’s benchmark that showed the relevant looping behavior and could be tested on the available machine. The benchmark runs without internet or tools and checks cited laws and court decisions against primary sources. The site’s published result remains the original Ornith score; the new experiment’s runs are not displayed there.
“Editing weights could remove an obstacle that hid an ability, but it could not add an ability the model lacked.”
— The MIT researchers, as quoted in Meyer’s report
Small Test Leaves Results Unsettled
The results remain preliminary: the report says there was one run per setting, the development set was small, and the loop detector changed during the experiment. It also says a change in the local setup alone moved the original model from 10 completed answers to four. These factors make it difficult to isolate how much of the completion change came from the weight edit.
There is also a disclosed grading limitation. Claude carried out the edit, and Claude agents conducted blind grading; GPT-6 Astra from another company reviewed the work and found faults. The source does not provide enough detail here to determine how much those faults changed the reported measurements, or whether results would hold with graders from different model families.
It is not yet clear whether the same edit would help other models, other tasks or repeated runs. Nor does the experiment show that the edited model’s legal answers are safe to rely on: its reported score was zero, and the citation problems persisted.
Repeat Runs and Broader Checks
The source material does not announce a scheduled follow-up or a publication date for additional results. The immediate next test would be to repeat the comparison under a fixed setup, with the same token budget and sampling settings across edited and unedited versions. That would help separate the effect of the weight change from configuration differences.
Further evaluation would also need a larger set of questions, consistent loop detection and independent grading that does not rely on the same model family as the editing process. Until such checks are reported, the result supports only a narrow claim: this edit was associated with more completed answers in one comparison, while the legal score remained zero.
Key Questions
What did the experiment change?
It removed a measured looping direction at half strength from the output weights in 27 of Ornith’s 40 layers. The team did not train the model or give it new legal data.
Did the edit improve the model’s legal performance?
Not by the reported benchmark score. The edited model’s score remained zero, despite completing more answers in the same-setup comparison.
Was this the same method used by the MIT researchers?
No. Meyer describes the home test as a related directional-ablation technique. The MIT work involved an AI reasoning directly about another model’s weights; the home test used code to run the model, measure an internal direction and edit selected weights.
What happened with the unedited model?
With sampling settings recommended by the model card, the unedited model completed 13 of 15 answers and also scored zero. The report cautions that the experiment was small and the setup changed.
Can the edited model be used for legal advice?
No such conclusion follows from the test. The report says the legal score stayed at zero and documents numerous incorrect or invented citations. It explicitly describes the work as an AI experiment, not legal advice.
Source: ThorstenMeyerAI.com
