An LLM Walks Into General Relativity - Lessons from a Devoxx Talk

Short summary

LLMs produce fluent but fundamentally incorrect technical content because they excel at structure and storytelling while lacking understanding of domain constraints and measurement invariants. Using physics as a test domain, a Devoxx speaker demonstrated this failure mode, then built a multi-agent validation system (generate → validate → critique → refine) using structured JSON outputs and domain-specific rules to catch systematic errors. The pattern generalizes to law, medicine, and finance.

• •LLMs excel at fluency but fail on domain constraints and invariants without system guardrails
• •Multi-agent pipeline with structured validation catches errors traditional prompting misses
• •Generalizable pattern: validate outputs deterministically via rules and critique loops, not better prompts

Generated with AI, which can make mistakes.