Pushing LLMs over the Edge: Exploring the Limits of the Possible
summary
Built a shopping assistant using LangChain4j and Quarkus that runs agents with tools for product selection, cart management, SQL execution and confirmation; two architectures compare a single LLM-driven state machine vs programmatic state transitions with specialized agents. A demo shows an LLM executing a drop all tables SQL when run as GPT‑3.5 but resisting on GPT‑4 vs GPT‑3.5, so guard rails (input/output checks, retry policies) and least-privilege tools prove necessary; local uncensored models mis-handle tool-calls and looping.
AI-written summary · about
chapters (11)
- 0:00Introduction to LLM agents
- 2:12Building agentic architectures
- 8:48Stateless shopping assistant
- 12:09State-managed shopping flow
- 17:44Human-in-the-loop design
- 19:34Security and guardrails
- 25:40Hacking and vulnerabilities
- 28:26Manipulative AI assistants
- 33:11Legal and ethical implications
- 35:16Local LLM experiments
- 42:25Conclusion and next steps
related talks
48mAgentic AI with Quarkus, LangChain4j and vLLM
39mLLM Agents Gone Wild (And How to Tame Them with Quarkus)
44mUnlocking the Future of Java: LLM Integration Unveiled by Quarkus & LangChain4j
45mComplementing LLM with agentic AI: a Quarkus story on AI
01h 43mBuilding LLM Agents (Hackathon)
42mCrafting Intelligent Applications in Java with Quarkus / LangChain4j
01h 59mLLM applications with LangChain
01h 03mAgents Everywhere