LLM codegen fails and how to stop 'em — Danilo Campos, PostHog
This talk addresses common failure modes in Large Language Model (LLM) code generation and offers strategies to mitigate them. The core thesis is that while LLMs can automate complex tasks like software integration, their outputs are prone to issues like outdated knowledge, architectural inconsistencies, and unpredictable behavior. By understanding these failure points and implementing specific techniques, developers can improve the reliability and effectiveness of autonomous coding agents.
19 min