Making state management intelligent
summary
Claim: combine LLMs and reinforcement learning to let app logic choose actions over a finite state graph, not just return text. Uses OpenAI (LLMs), agent patterns (LangChain/AutoGPT style), vector DB context, graph algorithms for shortest path traversal, and a small espresso-machine web app that maps UI states to transitions. Pipeline: LLM classifies desired end state, algorithm finds path through the state machine, code executes steps; demo succeeds about 90% of runs but sometimes misclassifies or injects wrong actions. Also proposes state machine synthesis and reinforcement learning with human feedback for adaptation.
AI-written summary · about
chapters (10)
- 0:00Introduction to intelligent state
- 1:03Evolution of state management
- 3:14The rise of generative AI
- 5:38Limitations of current AI
- 7:56Reinforcement learning basics
- 14:13RLHF and learning challenges
- 19:27Intelligent state management
- 22:52State machine demo
- 28:00Future AI possibilities
- 30:44Conclusion and key takeaways
related talks
42mMaking state management intelligent
32mDAVID KHOURSHID Making state management intelligent
01h 03mAgents Everywhere
49mPushing LLMs over the Edge: Exploring the Limits of the Possible
01h 43mBuilding LLM Agents (Hackathon)
38mInto the Visual World of State Machines
46mDemystifying AI Agents
36mLLMs and AI agents demystified