Bringing stories to life with AI, data streaming and generative agents
summary
Claims that LLM-driven agents can form coherent societies with memory and reflection, demonstrated by a Stanford demo of ~30 characters where over 50% of agents spread and attended a party; experiments show role-based multi-LLM panels outperform doctors (~70% vs ~25% on case diagnoses). Implementation ties Apache Kafka for event streams, Apache Flink for proximity and event processing, embeddings/RAG with OpenSearch for memory, and OpenAI LLMs for dialogue; a Stardew Valley prototype runs and reliably propagates goals across villagers.
AI-written summary · about
chapters (12)
- 0:00AI agents and generative worlds
- 2:21Agent-based modeling history
- 5:19Generative agents with LLMs
- 6:03Stanford Google 2023 paper
- 14:20Agent architecture and loops
- 16:52Agents from human interviews
- 19:02Agent values and trust
- 21:15Opinion formation simulation
- 24:04Microsoft diagnostic orchestrator
- 29:03AI in gaming tools
- 32:21Stardew Valley social sim
- 40:16Summary and future
related talks
37mBringing stories to life with AI
36mLLMs and AI agents demystified
46mDemystifying AI Agents
01h 43mBuilding LLM Agents (Hackathon)
41mBuilding LLM Agents – Challenges and Learnings
32mBeyond LLMs: Building an AI-Powered Decision System at Scale
50mIntelligent Agents, the New Frontier of LLMs
02h 08mIntroduction to AI Agents