We’re entering the era of infinite AI training environments. Google DeepMind just announced Genie 3, the first real-time interactive world model that creates worlds from text prompts. The video below shows a controllable environment generated by Genie 3 in real time. Insane. We're entering an era where world models like Genie 3 enable AI agents and robots to learn from their own experiences in simulated real-world environments. For example: > Anyone can create entirely imagined virtual worlds from a prompt > AI agents learning and improving from their own experiences in these worlds > Robots training and improving over time in simulated environments before touching the real world This is the convergence of world simulation, AI training, and creative expression all happening in real time. AI progress 📈

Genie 3 technical breakthroughs: > Real-time interaction at 24 frames per second > Visual memory extending back one full minute > Environmental consistency for several minutes > "Promptable world events" - change weather, add objects, introduce characters mid-session

DeepMind also tested generating worlds to train SIMA (their AI agent for 3D environments announced in 2024) They instructed the agent to achieve distinct goals while Genie 3 simulated the future based purely on the agent's actions No awareness of goals needed

Seems like the perfect time for @demishassabis and @elonmusk to create a video game together Congrats to the whole team @GoogleDeepMind for the achievement!