Guardrails and spec-driven LLM tooling

Creator of Meta's Hack: Your AI Will Always Cheat — Here's How to Stop It - Julien Verlaguet

Topic 6 of 11

Published Jun 8, 2026 · Clip 29:17–41:38 (12:22)

Julien calls guardrails for LLMs his new passion. He's relieved that at SkipLabs his users are computers, not emotionally attached humans, joking about how programmers relate to their languages like football teams. Structuring LLM work: be spec-driven, since LLMs lose track and compress history — keep a plan document as source of truth; split work into independent pieces with validation against original intent at every step; build graphs of tasks/subtasks with validation and integration layers. Tooling must be sound (so types can be trusted for downstream analysis) and incremental (so AI can iterate fast). He argues latency of generated code and the harness will be the key factor over the next five years, sketching a two-company scenario: Company A validates a diff and ships to production in ~100ms using static reachability analysis and a reactive runtime that updates program state to a semantically sensible state (equivalent to a full restart), enabling hundreds of AIs self-updating the system; Company B's CI-driven PR process taking hours is not the future.