Google DeepMind 如何利用遊戲打造更智能的 AI 智能體
How Google DeepMind Uses Games to Build Smarter AI Agents
Google DeepMind 的 SIMA 2 以遊戲作為訓練場，先從人類遊戲影片學習，再在 Gemini 回饋下自行練習。其任務成功率由 31% 升至 65%，並能應對此前未見、由 Genie 3 生成的遊戲世界。文章亦指出，SIMA 2 仍屬研究階段，在長時間多步驟任務及記憶方面有侷限。
Google DeepMind’s Alexandre Moufarek (Director, Inception team) says games are still the best training ground for AI. The simple version:
What is SIMA 2?
An agent that looks at the screen and uses a normal keyboard and mouse, like a player
Needs no game code or special APIs
Powered by Gemini, so it can reason, explain its plan and chat
How it learns
Starts from human gameplay videos
Then practises alone, with Gemini giving feedback
Understands sketches, emojis and other languages
Why it matters
Task success rose from 31% to 65%
It handled worlds generated by Genie 3 that it had never seen
Game uses: smarter NPCs, AI companions, automated QA testing
Next: an offline EVE Online sandbox to study memory and long-term planning
Limits: still research. It struggles with long multi-step tasks and has short memory.
Media
🖼️ Chart comparing SIMA 1, SIMA 2 and humans on task success: https://lh3.googleusercontent.com/KwuGClFlsydWIS2azo2WjBWkMY-Hm3geOZAad00wP-CMtmWQa8DlpmqOMvXC6QKiPnEUBgtToVWnJ9rzz0O7dTcaVFhYOSggEugNVJBWVgc1hegDfg=w1440 (Image: Google DeepMind)
🎥 Official intro video: https://www.youtube.com/watch?v=Zphax4f6Rls
References & Credits
Google DeepMind – SIMA 2 (SIMA Team)
Google DeepMind – From Atari to EVE Online (Alexandre Moufarek & Adrian Bolton)
AI Magazine – reference article
Decrypt – SIMA 2 coverage (31% vs 65% figures)
MIT Technology Review – SIMA 2 and Gemini
Image and video: Google DeepMind
來源：DEV Community · dev.to