The demo runs until the memory runs out. Google DeepMind’s Genie 3 can hold a coherent world for roughly a minute before inconsistencies creep in, gamesindustry.biz reported on March 11, 2026, citing a GDC talk. Months earlier the horizon was seconds. Progress is real; so is the ceiling.
Marketing still outruns stability
As Gamefile’s Stephen Totilo explained in reporting summarized by gamesindustry.biz, Genie 3 generates worlds frame by frame, more like reactive video than a classic 3D engine. Alexandre Moufarek, product lead for the Inception team at Google DeepMind, said at GDC on Monday that the team is not trying to replace traditional games. The goal is new experiences and, ultimately, AGI-ready environments for agents. gamesindustry.biz quoted Moufarek saying they are not at a stage to make a game with it, while inviting developers to experiment.
That honesty collides with January headlines when Project Genie news moved game stocks. gamesindustry.biz linked to IGN reporting on investor reaction. The technical truth is finite stability; the market narrative was infinite worlds.
SIMA 2 sits beside Genie as agent, not player substitute
gamesindustry.biz noted Google also outlined SIMA 2 at GDC, an agent that can play games; DeepMind had revealed SIMA 2 in November. The pairing shows where DeepMind spends attention: agents that need worlds, not worlds that ship as products tomorrow.
What This Actually Means
World models are demos with clocks. One minute of coherence is a leap from seconds but not a replacement for engines. Treat Genie 3 as R&D with a hype tax until persistence exceeds lunch break.