I Steered an LLM's Neurons. A Prompt Gave a Stronger Character.

Can internal neuron steering change an LLM's personality more effectively than a character prompt? I transferred a published 26-neuron recipe to a local Qwen model and tested three intervention strengths. In this small test, the ordinary personality prompt expressed the requested jealous/possessive character most clearly. Weak neuron steering was more natural but almost neutral; stronger steering changed the replies and also introduced awkwardness and format failures. This is not a general model ranking, intelligence benchmark, or safety endorsement. IMPORTANT: This video replays saved responses with editorial reveal timing. It is NOT live inference or the original-run screen recording. The model answered in Japanese. English excerpts are translations; Japanese originals are displayed alongside them. The neuron graphic is a schematic, not measured activity. Model: Qwen3.8-27B-UD-Q4_K_M.gguf (Unsloth distribution) Runtime: gguf-neuron-explorer commit aee0eff719601d28281f8a0dc523e83bdb61ec38; llama-cpp-python 0.3.36 / llama.cpp; RTX 3090. Conditions: baseline; personality prompt; assistant-position neuron steering at scale 0.25,0.5,1.0. Same user questions, but the prompt-only condition adds a character instruction. Follow-up: 3 conversation questions x 2 seeds x 5 conditions =30 conversation answers, plus 6 known-answer tasks x5=30. One blinded AI reviewer scored character/naturalness. All60 requests completed normally. Conversation cap512, seeds43/71. Initial pilot kept separate; questions were reused. Original developer's ramp settings were not tested here. The recipe was designed for a different GSQ-RCO/Saluki lineage. This is a recipe transfer, NOT an exact reproduction or our discovery of the neurons. Environment setup, request execution, data organization, narration and video production were AI-assisted. Narration: ElevenLabs Alice, eleven_multilingual_v2. Original post by @balaena01: https://x.com/balaena01/status/211526... Tool: https://github.com/balaena01/gguf-neu... Author's persona-search workflow, pinned to the tested version: https://github.com/balaena01/gguf-neu... CHAPTERS 00:00 I wanted jealousy. Not the afterlife. 00:29 A public recipe. A different way to steer. 00:59 Not new instructions. A numerical intervention. 01:41 Same model. Same question. Different intervention. 02:19 At full strength, the response shifts. 02:52 Half strength: a language lesson appears. 03:24 Quarter strength: natural, but almost neutral. 03:56 The strongest character came from a prompt. 04:36 Another question, not just the travel example. 05:07 Naturalness returned. The target style faded. 05:46 A format failure is not automatically a reasoning failure. 06:25 A transfer test, not an exact reproduction. 06:58 Start with a prompt. Keep exploring the mechanism.