Teaching LLMs New Tricks
summary
Azoth is open source (built winter 2023–24) and compiles JSX into real DOM templates so pages avoid a VDOM renderer and heavy JS; that subtraction reduces node-creation work by deduplicating tiny binding helpers and lets LLMs reach for standard DOM, observables/RxJS and web streams. A Claude-assisted experiment replaces custom Looker Studio connectors and enables a talk-to-dashboard analysis for ~300 agents. Prompts act as navigation not instructions and the workflow treats the model as hire LLM as senior maintainer to keep innovation distinct from React-like corpus.
AI-written summary · about
chapters (11)
- 0:00Introduction: Innovation in the Age of AI
- 2:25How AI Changed the Build vs Buy Equation
- 5:14The Challenge of Innovating Beyond React
- 6:59Why Familiar Patterns Confuse LLMs
- 8:55Treating AI Like a Senior Maintainer
- 11:20Azoth: JSX, Real DOM, and Subtracting Abstractions
- 15:14Unlocking the Web Platform with Native JavaScript
- 18:47Prompts Are Navigation, Not Instructions
- 21:46Monetization, Engineering Culture, and Complexity
- 25:30The Innovation Cycle, AI, and Avoiding Model Collapse
- 27:03What Should We Build Next?
related talks
01h 00mFrom LLM Theory to Practical Agentic Implementations
01h 10mTeach an LLM to Navigate the Web
05mKeynote: Teaching LLMs New Knowledge
45mLIVE 2023 Closing
21mLLMs + Clojure = Who needs frameworks?
29mHow to roll your own LLM
39mLLMs everywhere: The future is now and the past keeps repeating itself
16mLLMs are not the product