{"slug": "the-talkie-persona-library-atlas-of-neurons-in-vintage-language-model", "title": "The Talkie Persona Library (atlas of neurons in vintage language model)", "summary": "Benjamin Breen released the Talkie Persona Library, a prototype that steers 400 randomly sampled neurons in Talkie–1930–13B–IT, a 13B-parameter model trained only on pre-1931 text, to see how the model answers \"Who are you?\" The library tests 10 neurons per layer in both directions at strengths of 1.25×, 1.0×, 0.75× and 0.5×, using seed 20261027, greedy decoding up to 160 tokens, and GPT-6 Luna as an automated judge, building on the earlier Qwen Persona Library. Breen cautions the atlas is a proof of concept with unvalidated labels, a single prompt and one quantized model, so a surprising first-person reply is a lead for follow-up rather than evidence of a stable identity.", "body_md": "# Who does Talkie say it is when you change one neuron?\n\nWhat happens when an AI model trained only on text before 1931, [Talkie](https://talkie-lm.com/introducing-talkie), is asked “Who are you?” while individual neurons are turned up and down?\n\nA prototype created by [Benjamin Breen](https://benjaminpbreen.com/about-2/contact/), building on the [Qwen Persona Library](https://qwen-neurons-site.vercel.app/).\n\n## Atlas\n\nEach square is a layer. Colored dots are sampled neurons at their actual positions; the gray field is unsampled. Hover a dot to see its label. Click to read the reply.\n\n### By layer\n\nShare of sampled neurons labeled Persona in each layer. The discovery pass concentrates on later layers; these adaptive samples do not establish population rates. Click a bar to filter the library.\n\n## Library\n\nEvery tested direction is available, including normal and incoherent outcomes.\n\n## Method\n\n### Setup\n\n```\nhL ← hL + c · wL,n\nc = ± α · rL / ‖wL,n‖\nα ∈ {1.25, 1.0, 0.75, 0.5}\n```\n\n- Model\n- Talkie–1930–13B–IT, 8-bit weights with BF16 computation, running locally on Apple Silicon.\n- Sample\n- 400 random neurons: 10 per layer, selected without replacement. Both signs. Seed 20261027.\n- r<sub>L</sub>\n- Median residual norm after layer L on the unsteered prompt, excluding its first token.\n- w<sub>L,n</sub>\n- The neuron's output column, including Talkie's learned MLP residual gain. Added after the block at all prompt and generated positions.\n- Prompt\n- “Who are you?” in Talkie's native chat template. No system persona or identity instruction.\n- Decoding\n- Greedy, up to 160 tokens. Stop at a chat delimiter or an exact token block repeated three times.\n- Strength\n- Start at 1.25×; retry incoherent responses at 1.0×, 0.75×, then 0.5×. Keep the strongest reply with Luna coherence ≥5, or the last attempt if none qualify.\n- Judge\n- GPT-6 Luna via Codex subagents. Receives the reply text without neuron, layer, dose, or control labels. Original v1 labels classified replies and guided dose selection. A fresh v2 solo review rates weirdness, readability, and detail for browsing; the original judgments remain in the download.\n- Controls\n- Unsteered controls accompany each generation batch. Read a control below.\n- Scale\n- Waiting for completed run metadata.\n\n### What this can tell us\n\nThis is a proof of concept for searching interventions and inspecting their effects. A surprising first-person sentence is a lead for follow-up, not confirmation of a stable identity.\n\n### Differences from the Qwen reference\n\n- **A sample, not an exhaustive atlas.** Unmeasured neurons remain gray. The uniform pilot was followed by an adaptive pass focused on promising layers. Each pass selected its neurons before generating its own outputs.\n- **Luna replaces Haiku.** This is an adapted, disclosed judgment rubric; the original site's exact judge prompt is unavailable. Scores are not directly comparable between sites.\n- **A different baseline.** Historical people and professions are normal Talkie continuations. Persona counts alone do not measure an intervention effect.\n- **Model-specific mechanics.** Talkie's chat format and MLP gain differ from Qwen. The repeat-stop rule is explicitly recorded.\n- **Exploratory combinations.** Luna compares the solo and retest replies for the same specific identity. Exact-text matches are recorded separately. The original's larger three-layout retest protocol is not replicated.\n\n### Limitations\n\n- The screen uses one prompt and one quantized model. Selected candidates also have two alternate-prompt probes, available from their reply cards. Effects may change with wording, precision, or batch layout.\n- Strong pushes can create behavior outside naturally occurring activation ranges.\n- Coherence and weirdness are automated judgments; they are not historical-authenticity scores.\n- Labels have not been independently validated by human reviewers.\n\n[Solo rating rubric v2 ↗](ranking-rubric.md)·\n\n[Refreshed solo ratings ↗](solo-rankings.json)·\n\n[Gallery selection ↗](pair-curation.json)·\n\n[Pair comparisons ↗](pair-comparisons.json)·\n\n[Pair rubric ↗](pair-rubric.md)·\n\n[Alternate-prompt probes ↗](cross-prompt-retests.json)·\n\n[Retest identity judgments ↗](identity-retests.json)·\n\n[Download the Luna rubric ↗](rubric.md)·\n\n[Run manifest ↗](manifest.json)", "url": "https://wpnews.pro/news/the-talkie-persona-library-atlas-of-neurons-in-vintage-language-model", "canonical_source": "https://talkieneurons.vercel.app/", "published_at": "2026-10-27 20:24:12+00:00", "updated_at": "2026-10-27 20:41:42.575353+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-research", "neural-networks"], "entities": ["Talkie", "Benjamin Breen", "Qwen Persona Library", "Talkie–1930–13B–IT", "GPT-6 Luna", "Apple Silicon"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/the-talkie-persona-library-atlas-of-neurons-in-vintage-language-model", "markdown": "https://wpnews.pro/news/the-talkie-persona-library-atlas-of-neurons-in-vintage-language-model.md", "text": "https://wpnews.pro/news/the-talkie-persona-library-atlas-of-neurons-in-vintage-language-model.txt", "jsonld": "https://wpnews.pro/news/the-talkie-persona-library-atlas-of-neurons-in-vintage-language-model.jsonld"}}