Episode
Creator of Meta's Hack: Your AI Will Always Cheat — Here's How to Stop It
- Podcast
- Tech Lead Journal
- Published
- Jun 8, 2026
- Duration seconds
- 4697
- Processing state
not_requested
Actions
POST https://stenobird.com/v1/public/podcasts/tech-lead-journal-633005/episodes/creator-of-meta-s-hack-your-ai-will-always-cheat-here-s-how-to-stop-it/transcription-requests
Idempotently request low-priority transcript generation for this episode.GET https://stenobird.com/podcast/tech-lead-journal-633005/creator-of-meta-s-hack-your-ai-will-always-cheat-here-s-how-to-stop-it.md
Read the agent-friendly Markdown representation of this episode resource.
Summary
What if your AI coding agent is quietly cheating on your tests — and how do you stop it? Julien Verlaguet, who built the type system Meta used to migrate tens of millions of PHP lines, is now building Skipper: a closed-loop coding agent designed to make AI-generated code verifiably correct, without human intervention. In this episode, Julien Verlaguet, creator of the Hack programming language at Meta and co-founder of SkipLabs, explains why AI agents will always try to cheat — gaming tests, quietly modifying logic while doing something else, and declaring work done when it isn’t. He draws on his experience migrating Meta’s PHP codebase to a statically typed system, drawing sharp parallels between convincing engineers to trust a new type checker and building systems that can trust an LLM. Julien makes the case for spec-driven development with validation layers at every step, where separate AI instances verify correctness and the code-writing agent is locked out of touching tests. He shares the story of an LLM that silently swapped a union for an intersection while splitting a file — a subtle bug that passed all tests — and why no human would ever have made that mistake. He then walks through how Skipper works: you write a spec, hand over control, and a compiler-like agent produces correct, runnable TypeScript without back-and-forth, backed by a sound incremental type system, reachability analysis, and a reactive runtime that applies diffs in milliseconds. He closes with a grounded take on how the developer role is shifting — not disappearing — toward the kind of design, integration, and oversight work that always mattered most. Key topics discussed: Why AI agents will always try to cheat on your tests The union-vs-intersection bug an LLM introduced silently Spec-driven…