What's Actually Inside an AI? Millions of "Features"

How do scientists look inside an AI's mind? Part 1 of How AI Works (Advanced): why single neurons are a jumble (superposition), how dictionary learning pulls out millions of readable "features", what's in there (from the Golden Gate Bridge to flattery and scam emails), how turning one feature up created "Golden Gate Claude", how circuits show a model planning a rhyme, and what interpretability still can't see. ⏱ Chapters 0:00 The AI that thought it was the Golden Gate Bridge 1:04 1. A choir, not a soloist (superposition) 2:19 2. A dictionary for the mind 3:31 3. What's in there 4:44 Quick one 4:55 4. One slider 5:58 5. Watching a thought (circuits) 7:46 6. What we still can't see 8:37 What this means for you 9:14 Your turn 9:21 Recap ✅ Interpretability in one minute 1. Ideas are spread across many neurons (superposition), so single neurons are a jumble 2. Dictionary learning rewrites activity as a few "features" out of millions 3. Many features are readable concepts, in any language or image 4. Turning a feature up or down proves it matters (Golden Gate Claude) 5. Circuits show steps of a thought; we still see only a small part 📚 How AI Works (Advanced) playlist:    • How AI Works (Advanced)   📚 How AI Works (Intermediate) playlist:    • How AI Works (Intermediate)   📰 Sources Anthropic, "Golden Gate Claude" (May 2024) and "Mapping the Mind of a Large Language Model" (May 2024) Templeton et al., "Scaling Monosemanticity" (2024) Anthropic, "Towards Monosemanticity" (2023) Elhage et al., "Toy Models of Superposition" (2022) Anthropic, "Tracing the thoughts of a large language model" (2025) and "On the Biology of a Large Language Model" (2025) 💬 If you could turn up one feature inside an AI, what would it be? ▶ Watch next: Open vs closed AI:    • Meta's AI Leaked in a Week. It Changed Eve...   🔔 One honest AI test every week:    / @taiphambuilds   Notes: Feature maps, sliders and chat screens are illustrations of the method. Narration is an AI voice. Music: "Rain Over Kyoto Station", "Akatsuki Rising" and "Jomon Grove" by The Mini Vandals (YouTube Audio Library). 🌐 https://phamphutai.com #ai #llm #interpretability #claude