Genie 3 - interactive worlds
I was going to tuck Genie 3 from Google DeepMind into another post, but it deserves centre stage — it’s a fundamental step towards enabling a new kind of AI.
Jack Parker-Holder and Shlomi Fruchter
Imagine prompting: “Running by the shores of a glacial lake, exploring branching forest paths, crossing flowing streams beneath snow-capped mountains…” and instantly stepping into that environment. It remembers your actions, adapts to your choices, and you can prompt new events as you go - “suddenly, a bear appears on the trail…”
Why is this important? Video generation is passive - you watch. Genie 3 is active - you do. That makes it the perfect playground for what Yann LeCun of Meta has long argued: AI needs to build internal “world models” to reason, plan, and predict consequences, much like humans do.
ChatGPT is brilliant at predicting the next word, but "understanding" the world only through text (and probability) is limiting. Humans (and animals) learn quickly through trial and error. Giving AI the chance to explore and experiment inside simulated worlds could unlock a new level of capability — one not constrained by human-written data, but grounded in experience.
It’s early days, but generating interactive worlds could be the step that moves us beyond predictive text and toward AI that learns the way humans and animals do - by acting, experimenting, and discovering.