Interview
Demis Hassabis: Why AGI is Bigger than the Industrial Revolution & Where Are The Bottlenecks in AI
- Scaling returns are expected to remain substantial over the next few years, though margins are projected to narrow relative to the start of the current scaling era.
- Labs capable of inventing new algorithmic ideas are predicted to gain a significant advantage within the next few years as current methodologies reach their efficiency limits.
- DeepMind anticipates extracting 30-40% additional efficiency from national grids through the optimization of existing infrastructure.
- AGI arrival is predicted to have a "very good chance" of occurring within the next five years, with an anticipated impact unfolding over a decade rather than a century.
- A "golden age of scientific discovery" is expected to commence within five plus years, driven by AGI applications in science and medicine.
- The complete drug design engine at Isomorphic Labs is projected to be ready in five plus five to 10 years, with full regulatory validation and the potential elimination of animal testing occurring approximately ten years after initial breakthroughs.
- AI is forecast to be essential for achieving breakthrough technologies like fusion, new batteries, and superconductors within a five to 10 year timeframe.
- The economic costs of the AI energy revolution are expected to be recouped in the medium to long run, with unlimited rocket fuel from fusion potentially enabling cheaper space exploration.
- Risks include bad actors repurposing dual-use technologies for harmful ends and autonomous systems becoming difficult to control as they approach AGI capabilities within a year or two.
- AGI adoption is predicted to be overhyped in the next year but significantly underappreciated over a 10-year horizon.
- New, higher-quality, and higher-paying jobs are expected to emerge to replace roles lost to automation.
- European markets may overcome fragmentation via initiatives like "the EU Inc," and Isomorphic Labs has the potential to become a trillion-dollar company.
- Open-source models are expected to remain one step behind the absolute frontier, requiring approximately six months for the community to re-implement leading ideas.
- Current foundation models are unlikely to be replaced but will instead serve as the base for future general intelligence systems.
- Key missing capabilities identified in current systems include continuous learning, long-term planning, and the ability to generalize beyond "jagged intelligences."
- About 90% of breakthroughs are expected to be delivered by a small group of three to four leading labs, with the gap between them widening as tools facilitate the next generation.
- The UK is expected to continue producing top-tier talent from its leading universities, attracting global expertise due to a conducive environment for deep tech.
- International governance is anticipated to establish minimum standards and certification processes, potentially overseen by an international body similar to the Atomic Energy Agency.
- Climate modeling is expected to assist in mitigating climate change effects, while new batteries and superconductors could fundamentally alter the nature of the economy within the next decade.
- After a dozen or so AI-designed drugs pass clinical trials, governments may begin to trust model predictions to back-test and accelerate regulatory processes.