Summary
Meta principal engineers Danielle An and David Hoe took the stage at the AI Engineer conference to share what they’ve learned over more than a year leading AI-driven game creation efforts at Meta. Rather than a conventional slide deck, they delivered their talk through an interactive NPC-driven game that audience members could control in real time via QR code — itself a live demonstration of the technology they were discussing.
The core thesis is that AI tools like Meta Banana and Meshy have removed the traditional skill-gate in game development: coders who can’t draw and artists who can’t code can now both build playable games. But An and Hoe argue that closing that gap only gets you to parity — what separates a good game from a forgettable one is still taste, aesthetic cohesion, and an understanding of what humans find fun.
The talk’s most technically forward section covers runtime LLMs — large language models running during gameplay that drive NPC decision-making in real time rather than following scripted behavior trees. A multiplayer demo showed four LLM-controlled NPCs competing to collect cubes, each with distinct personalities (thief, honorable, aggressive) making unscripted decisions to steal, block, or cooperate. David Hoe built the demo in roughly 12 hours. An and Hoe frame this emergent, non-repeatable gameplay as a genuinely new game genre that only became feasible in the last 18 months as inference costs dropped enough to run LLMs in production game loops.
📺 Source: AI Engineer · Published July 08, 2026
🏷️ Format: Keynote Launch
