What happened
Google DeepMind published a retrospective on its fifteen years of AI research in games. The post tracks their work from the Deep Q-Network that mastered Atari games in 2015 to AlphaGo's victories in Go. It highlights a shift in focus from mastering single games to building generalist agents. Their latest project, SIMA, is designed to understand natural language instructions and interact with various games through standard keyboard and mouse controls — without needing direct API access. The post also details a new research partnership with Fenris Creations, the studio behind EVE Online.
How the room's reading it
The consensus among AI researchers is that this marks a deliberate pivot from solving games to understanding worlds. The earlier work like AlphaGo was about superhuman optimisation within a closed system. The new work with SIMA and EVE Online is seen as a serious attempt to tackle the messier problems of generalisation and long-term planning in an open, persistent environment. Developers on X are noting the potential for more dynamic NPCs and robust QA testing. Others are more sceptical, pointing out that truly general agents are still a long way from being a practical tool for most studios.
Sailfish's take
We see this as less about gaming and more about a public roadmap for agentic AI. The shift from single-player optimisation to multi-agent persistence is the entire challenge for real-world agents. DeepMind isn't just trying to build a better NPC — they're using EVE Online's complex, player-driven world as a lab for the core problems holding back enterprise agents. These are continual learning, long-horizon memory, and complex social dynamics. The real prize isn't a smarter opponent. It's a blueprint for agents that can navigate sprawling corporate software. We're watching this for architectural hints, not for gameplay.