Google has introduced a groundbreaking AI model named Genie, developed by Google DeepMind, which has the capability to create interactive and playable virtual worlds, termed 'Generative Interactive Environments', from a single image. This AI model, trained exclusively on unlabeled internet videos, leverages a 'Latent Action Model' (LAM) to generate a wide variety of action-controllable environments, essentially offering a new dimension to video game creation and AI research. By being unsupervised and learning from internet video, Genie addresses several challenges associated with AI video game generation, such as the need for external data. Genie's potential extends beyond gaming, with implications for creating more adaptable, generalist AI agents and contributing to fields like robotics. The introduction of Genie has been met with enthusiasm from the tech community, highlighting its promise in unleashing creativity, its designation as a 'foundation world model', and its role in ushering in 'Metaverse 2.0.'
Google DeepMind Unveils AI ‘Genie’ to Instantly Conjure Playable Games ► https://t.co/6WNOWpaisB https://t.co/6WNOWpaisB
Are we close to the holodeck? Google unveils Genie — an AI model creating playable virtual worlds from a single image https://t.co/w7NkEGAEfA
"Genie is a powerful tool for AI research that offers a never-ending curriculum of new, generated worlds for training more adaptable, generalist AI agents." Or, as I say, "Metaverse 2.0." :-) https://t.co/rnDq7GSmVx
Nice work from colleagues: *latent action-conditioned* foundation world models! Very promising results on a ton of video games and control domains (including robotics). One of the best action-grounded video prediction models I’ve seen. https://t.co/MIROdA6I17
This is awesome. Worth reading the thread. Learn latent action space on filtered video, humans can play the AI generated games. https://t.co/SifZArykn6
Genie out of DeepMind is a bigger deal for video games than Sora 🤯 It's a foundation world model trained exclusively from unlabeled internet videos that can generate an endless variety of action-controllable virtual worlds (i.e. interactive video games) given image prompts… https://t.co/shQFBMXYLn
https://t.co/JuEg4nuX6o Are we close to the holodeck? Google unveils Genie — an AI model creating playable virtual worlds from a single image | Tom's Guide https://t.co/CvKFWJLW5T
Nice! A "foundation world model" from Google that generates interactive environments despite being trained only on unlabelled internet videos. So many interesting concepts & implications (eg: the playable photo example) Given DeepMind & OpenAI's sheer history with game engines,… https://t.co/NEyVFnTSLh https://t.co/UIyLVdv41z
Unleashing creativity with Genie 🧞— Learning action controllable imagined worlds purely from internet video. https://t.co/FDKGNJ5Bzi
Beautiful summary, especially the benefits of latent action models emerging properties akin to early Artificial Intelligence proto creativity. Generating new data by itself solves a lot of the concerns that other models need external data. Incredible potential beyond video games https://t.co/E3l5OQr0qB
This is incredible. By being unsupervised, Genie solves many of the problems inherent in AI video game generation. Just read through the paper and here's a quick summary. This paper introduces: 1. Latent Action Model (LAM) — with nothing other than video, the model infers… https://t.co/LspBeiQnen
🚨 Google just dropped an AI game/world maker! Genie (Generative Interactive Environments) 🧵 More in the thread: https://t.co/fhSWH2U0wo