# The Talkie Persona Library (atlas of neurons in vintage language model)

> Source: <https://talkieneurons.vercel.app/>
> Published: 2026-10-27 20:24:12+00:00

# Who does Talkie say it is when you change one neuron?

What 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?

A prototype created by [Benjamin Breen](https://benjaminpbreen.com/about-2/contact/), building on the [Qwen Persona Library](https://qwen-neurons-site.vercel.app/).

## Atlas

Each 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.

### By layer

Share 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.

## Library

Every tested direction is available, including normal and incoherent outcomes.

## Method

### Setup

```
hL ← hL + c · wL,n
c = ± α · rL / ‖wL,n‖
α ∈ {1.25, 1.0, 0.75, 0.5}
```

- Model
- Talkie–1930–13B–IT, 8-bit weights with BF16 computation, running locally on Apple Silicon.
- Sample
- 400 random neurons: 10 per layer, selected without replacement. Both signs. Seed 20261027.
- r<sub>L</sub>
- Median residual norm after layer L on the unsteered prompt, excluding its first token.
- w<sub>L,n</sub>
- The neuron's output column, including Talkie's learned MLP residual gain. Added after the block at all prompt and generated positions.
- Prompt
- “Who are you?” in Talkie's native chat template. No system persona or identity instruction.
- Decoding
- Greedy, up to 160 tokens. Stop at a chat delimiter or an exact token block repeated three times.
- Strength
- 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.
- Judge
- 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.
- Controls
- Unsteered controls accompany each generation batch. Read a control below.
- Scale
- Waiting for completed run metadata.

### What this can tell us

This 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.

### Differences from the Qwen reference

- **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.
- **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.
- **A different baseline.** Historical people and professions are normal Talkie continuations. Persona counts alone do not measure an intervention effect.
- **Model-specific mechanics.** Talkie's chat format and MLP gain differ from Qwen. The repeat-stop rule is explicitly recorded.
- **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.

### Limitations

- 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.
- Strong pushes can create behavior outside naturally occurring activation ranges.
- Coherence and weirdness are automated judgments; they are not historical-authenticity scores.
- Labels have not been independently validated by human reviewers.

[Solo rating rubric v2 ↗](ranking-rubric.md)·

[Refreshed solo ratings ↗](solo-rankings.json)·

[Gallery selection ↗](pair-curation.json)·

[Pair comparisons ↗](pair-comparisons.json)·

[Pair rubric ↗](pair-rubric.md)·

[Alternate-prompt probes ↗](cross-prompt-retests.json)·

[Retest identity judgments ↗](identity-retests.json)·

[Download the Luna rubric ↗](rubric.md)·

[Run manifest ↗](manifest.json)