Name: Inside Google DeepMind: AGI, Robotics, & World Models Explained - Demis Hassabis
Uploaded: 2025-09-12T21:56:44+00:00
Duration: 1908 s

Speakers

Demis Hassabis — CEO, Google DeepMind

David Friedberg — CEO, The Production Board

Summary

At the All-In Summit, Google DeepMind CEO Demis Hassabis discusses his Nobel Prize, DeepMind's role as the AI engine room of Google and Alphabet, and the Genie 3 world model that generates interactive 3D environments from a text prompt. He argues world models are a prerequisite for AGI and robotics, says robotics is still early but heading for a 'wow moment' within a couple of years, and puts full AGI five to ten years away. He also covers Nano Banana and AI creative tools, Isomorphic Labs' AI drug-discovery push with Eli Lilly and Novartis, and why efficiency gains are not yet reducing AI's energy demand.

Google DeepMind merged Alphabet's AI efforts; Gemini models now sit in nearly every Google product, reaching billions of users.
Genie 3 generates interactive, controllable worlds from a text prompt with no rendering engine, having learned intuitive physics from video.
World models and multimodal Gemini robotics models are presented as the missing piece for robotics and smart-glasses assistants.
Hassabis calls robotics still early, expects a wow moment within a couple of years and eventually millions of robots, and sees a role for both humanoid and task-specific forms.
He disputes that LLM progress is flatlining but says AGI is five to ten years away and may need one or two more breakthroughs.
Nano Banana and Veo point to creative tools that democratize creation while making top professionals far more productive.
Isomorphic Labs, partnered with Eli Lilly and Novartis, aims to cut drug discovery from years to weeks and expects pre-clinical work next year.
Serving efficiency has improved 10-100x, but frontier-model scaling keeps AI energy demand from falling.

Ideas

Demis Hassabis CEO, Google DeepMind 12:02

Humanoid form factor important for general use

Hassabis used to expect only task-specific robots, and still thinks industrial and laboratory robots will be optimized for their specific tasks, but for general-purpose or personal-use robotics he now sees the humanoid form factor as potentially very important: the physical world - steps, doorways, everyday objects - was designed for humans, so it is easier to design a robot that fits that world than to redesign the world. He therefore expects a place for both humanoids and specialized robotic forms.

Demis Hassabis CEO, Google DeepMind 13:04

Robotics still early; wow moment coming soon

Hassabis, who says he spends a lot of time on this question, thinks robotics is still a little early: the general-purpose models that vision-language-action robots are built on need to become more reliable and better at understanding the physical world, which is what world models like Genie and the multimodal Gemini robotics models are meant to supply. He expects a real 'wow moment' in robotics within the next couple of years and, eventually, millions of robots raising productivity. The main hardware risk is scaling too early: once factories are built around tens or hundreds of thousands of one robot design it becomes hard to iterate, and a more reliable, more dexterous generation may appear within six months. He compares the stage to the 1970s PC era, but with ten years of progress compressed into one.

Inside Google DeepMind: AGI, Robotics, & World Models Explained - Demis Hassabis

Summary

Ideas

Up Next