Interview, Fireside Chat
Inside Google DeepMind: AGI, Robotics, & World Models Explained - Demis Hassabis
- DeepMind is developing Genie 3 to enable natural language generation of interactive 3D worlds, with an evolutionary roadmap extending to Genie 5 and future iterations serving as world model drivers for AGI.
- The robotics sector is projected to experience rapid evolution analogous to the PC DOS era of the 1970s, with a "real wow moment" anticipated within the next couple of years and a "real wow moment" expected within the next couple of years.
- Google is pursuing dual robotics strategies involving an "Android for robotics" universal operating system layer and vertical integration of latest models with specific robot designs.
- Necessary form factors include both humanoid designs for general human-centric environments and specialized variants, contingent on hardware capable of scaling to millions of units to increase societal productivity.
- Significant algorithmic development over the next couple of years is required in reliability and world understanding to support robotics, alongside the emergence of smart glasses and assistants capable of physical context awareness.
- AI is predicted to achieve AGI status capable of PhD-level performance across all domains in five to ten years, requiring one or two breakthroughs in reasoning, consistency, and continual learning within the next five years.
- Genie models and text-to-video V0 models are identified as primary drivers for building the world models necessary for AGI, while full AGI is expected within ten years.
- Current AI lacks true creativity to independently generate new conjectures or theories, a capability Isomorphic aims to leverage for drug discovery, reducing timelines from years to weeks or days over the next 10 years.
- Isomorphic is scheduled to enter the preclinical phase sometime next year, with biological and chemical AI models relying on hybrid probabilistic-physical systems for the next five years before transitioning to end-to-end learning.
- Nano Banana and similar generators are expected to democratize creativity, enabling everyday users to produce content and professional creatives to achieve 10x to 100x productivity gains through co-creation models in entertainment.
- Model efficiencies have improved 10x to 100x over the last two years via distillation, with expectations that AI will return more value to energy and climate systems than it consumes over the next 10 years.
- The future of AGI benefits is projected to span human health and energy sectors, ushering in a new golden era of scientific discovery driven by AI acceleration.