Google DeepMind’s CEO, Demis Hassabis, revealed in a recent podcast interview plans to eventually integrate Google’s Gemini artificial intelligence models with Veo, its advanced video-generation technology, to enhance Gemini’s grasp of the physical world.
Discussing the possibility on the podcast “Possible,” co-hosted by LinkedIn co-founder Reid Hoffman, Hassabis emphasized that Gemini was intentionally designed as a multimodal model right from the start. “The vision we’ve always had is for Gemini to act as a universal digital assistant, one capable of genuinely interacting with, and aiding in, real-world situations,” Hassabis explained.
This integration aligns with an ongoing trend across the industry toward increasingly versatile “omni” models—artificial intelligence systems able to expertly understand and generate diverse forms of media. Google’s latest Gemini models already have capabilities that span text generation, image synthesis, and audio creation. Other tech giants are also exploring similar directions. For instance, the latest iteration of OpenAI’s ChatGPT already has integrated image generation abilities, including artistic styles reminiscent of Studio Ghibli, while Amazon is preparing an ambitious “any-to-any” multimodal model planned for release later this year.
Training these highly complex models, particularly those involving video content, demands vast amounts of diverse and quality data input. Hassabis confirmed that Google is indeed tapping into its massive video resource from YouTube to help Veo build valuable insights into real-world dynamics, emphasizing the importance of extensive video analysis. “Essentially, through extensive viewing of YouTube footage, Veo 2 learns core principles about how the physical environment works,” said Hassabis.
Google had previously acknowledged that some YouTube content was being used as data for AI model training, in line with its existing agreements with creators. Notably, the company also broadened its terms of service within the past year specifically to facilitate increased access to data for AI systems development.
