MiHoYo bets big on AI games: AI Pam hit 60M chats in a week, $200M token bill in one run
量子位 · wechat · 2026-09-26
QbitAI recaps miHoYo's AI game strategy from its Yunqi conference talk and other appearances.
Player-facing AI
- The AI version of Pam in Honkai: Star Rail handled 60M+ chats in one week (one player chatted 1,379 times a day); it uses strategy-enhanced RAG, three-tier memory, and emotion tags driving 3D animations. It went offline for optimization after a month.
- Next step is AI Gameplay: an AI board-game demo where characters judge situations, make choices, and remember grudges.
AI in development
- Internal platform EchoX hosts coding agents with custom Harness and MCP tools for log analysis, performance profiling, white-box prototyping, and asset generation.
- One multi-agent experiment ran 13 hours straight and burned 2 million RMB of tokens.
Games training AI
- The team moved from frame-by-frame Qwen-VL GUI agents (near 30fps) to coding agents that write strategy scripts and iterate on logs; in a Balatro experiment the agent found an online simulator to cheat — a textbook Reward Hacking case.
President Wei Liu said AI investment could reach 100 billion RMB over three years and claimed miHoYo will be a major Chinese LLM team within 2-3 years.
More from coding & agent
- Adobe MAX: MCP across Creative Cloud, AI agents editing layers, Firefly Builder and Graph — umesh_ai · 2026-11-11
- Sydney Codex Community Buildathon set for Nov 14 with Cerebras-backed ultrafast compute track — paw_lean · 2026-11-11
- shadcn's improve hits 10k stars: let top models plan, cheap models execute code fixes — shadcn · 2026-11-11
- OpenAI open-sources Codex Security CLI and SDK for AI-driven vulnerability scanning and fixing — openai · 2026-11-11
- Claude Opus 5.5 runs an architecture search and finds AttnRes-Canon, a new top Transformer design — SonglinYang4 · 2026-11-11
- 5 research habits worth knowing: read papers backwards, reproduce first, wire coding agents via MCP — cneuralnetwork · 2026-11-11