Random 128 neurons finetune just as well — owl persona isn't about special neurons
Sauers_ · x · 2026-10-18
Follow-up analysis on the owl-neuron experiment surfaced two counterintuitive findings:
- The finetuned neurons have low cosine similarity to their original functions in the base model — the persona isn't exploiting the neurons' existing roles;
- Claude's analysis noted that an earlier pilot using 128 random neurons from the same layers trained into an equally convincing owl, suggesting neuron selection may not be the key factor.
An accompanying circuit diagram shows the discovered circuit spans 65 neurons on the system prompt, attention heads looking at "AI," and 11 neurons on the reply. Sauers speculates first-person effects may stem from how the 128 neurons were found, and better techniques could reveal more "privileged persona neurons."
More from Models
- OpenAI's Decisions API hits GA: near real-time routing up to 10x faster than GPT-6 Luna — OpenAIDevs · 2026-10-23
- DeepLearning.AI weekly: GPT-6.1 hits 97.9% byte parity on COBOL migration — DeepLearningAI · 2026-10-23
- Marker's lite mode scores 75.1 vs 42.3 for Azure Read on OmniParseBench — VikParuchuri · 2026-10-23
- Datalab launches lite OCR at $0.50 per 1,000 pages, 20-30x cheaper than Azure, Google and AWS layout APIs — VikParuchuri · 2026-10-23
- LiveNerf's 29-day data finds no nerfing of Claude Opus 5.5 — TheOnlyVibemaster · 2026-10-23
- PHP Foundation reveals Anthropic's Claude cyber model scanned PHP, validated with OpenAI's Daybreak Red — saibharadwaj · 2026-10-23