Developer uses Gemma 3 4B IT for AI ghost hunting experiment

A developer has created an experimental tool called "The Ghost Hunter's Lab" to investigate internal AI model behavior, specifically targeting the Gemma 3 4B IT model. The project uses a sparse autoencoder feature associated with ghost-related text and injects it into the model during inference to observe changes in output probabilities. While the direct output remained similar, the experiment revealed nuanced shifts in token probabilities, highlighting the complexity of AI interpretability. AI

IMPACT Demonstrates novel methods for probing internal AI model representations, potentially aiding future interpretability research.

RANK_REASON The item describes a personal project using existing models and tools for experimentation, rather than a new model release or significant industry event.

Read on dev.to — LLM tag →

• FastAPI
• Gemma Scope 2
• Gilbert Ryle
• GitHub
• Google DeepMind
• GSAP
• llama.cpp
• The Ghost Hunter's Lab

• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →