Substrate engineering: Engineering foundations in a world of LLMs

Discover why building robust engineering systems and foundations is crucial for successful LLM integration, and learn key strategies for safer AI implementation.

Key takeaways
  • LLMs are fundamentally predictors over tokens and hallucination is an inherent characteristic, not a solvable problem - no amount of prompt engineering can overcome this

  • The quality of engineering systems and foundations (the “substrate”) will always dominate prompt engineering results - you can’t get better outputs than your system architecture and guardrails allow

  • Better tooling and configuration languages are needed - YAML/JSON lack type safety and validation capabilities that could prevent errors before deployment

  • Focus should be on “defense in depth” through multiple complementary approaches:

    • Type systems and compile-time checks
    • Restricted/non-Turing-complete configuration languages
    • Automated validation and testing
    • Clear boundaries on LLM usage scope
  • The better automation works, the less humans tend to pay attention - systems need to be designed accounting for this reality

  • Memory safety and thread safety cannot be guaranteed through prompting alone - language and tooling choices matter significantly

  • Configuration and tooling code is especially likely to be written with LLMs, making it critical to have strong foundational safeguards

  • Engineering systems need totality (guaranteed output for every input) and sound type systems without being overly complex

  • Code review and testing remain essential - LLMs don’t remove the need for human oversight and validation

  • Success with LLMs requires understanding their limitations and building appropriate guardrails and foundations rather than relying solely on prompt engineering