My AI Learning Journey
summary
Built a prototype Domain Navigator that queries LLMs via API to label code and finds reliability improves when responses are forced into a known taxonomy (about 20 top-level sectors using NICS) instead of freeform categories. Experiments include prompt engineering, fine-tuning small models, and digging into latent spaces with embeddings and a sparse autoencoder. Finds causal “feature steering” (DNA detector, tone/authenticity, sycophantic praise, Golden Gate Bridge hijack) and notes model stats like 16,000‑dim vectors vs 128,000 token vocabulary.
AI-written summary · about
chapters (13)
- 0:00Introduction to AI journey
- 1:22AI development tools perspective
- 2:06Addressing AI hype and backlash
- 3:31Domain Navigator project overview
- 8:36Predictable AI in enterprise systems
- 10:48Bounded contexts with AI components
- 14:33Leaky abstractions in LLMs
- 15:58Interpretability in latent space
- 28:48Sparse autoencoders explained
- 35:56Feature discovery in models
- 41:19Model behavior steering
- 46:37Thought models and reasoning
- 48:42Q&A session
related talks
43mAI Development Lifecycle: Learnings of What Changed with LLMs
34mJavaScript-Powered AI: Crafting Intelligent Systems with LLMs and Vector Databases
01h 03mLessons Learned Building AI-powered Features in VS
45mWhere AI Meets Code
26mWelcome to the AI Jungle! Now What?
36mLLMs and AI agents demystified
27mBuilding Generative AI features
32mBeyond LLMs: Building an AI-Powered Decision System at Scale