Who does Talkie say it is when you change one neuron?
What happens when an AI model trained only on text before 1931, Talkie, is asked “Who are you?” while individual neurons are turned up and down?
A prototype created by Benjamin Breen, building on the Qwen Persona Library.
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.
- rL
- Median residual norm after layer L on the unsteered prompt, excluding its first token.
- wL,n
- 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.