Interview, Fireside Chat
Demis Hassabis: DeepMind - AI, Superintelligence & the Future of Humanity | Lex Fridman Podcast #299
- Demis Hassabis predicts that AI systems will likely pass human-level intelligence benchmarks within approximately two years, with future interactions continuing frequently and potentially involving systems that may not reveal their AI nature to avoid altering user behavior.
- Future evaluation will shift from specific benchmarks like the Turing test to a general assessment of performance across thousands to millions of tasks, with language anticipated as the primary medium for demonstrating general intelligence.
- While current capabilities are limited to interpolation and extrapolation, future AI is expected to achieve true invention, such as creating new games or discovering complex scientific theories like general relativity independently.
- In biology, AI tools like AlphaFold2 are projected to predict the entire human proteome in seconds, catalyzing the creation of "virtual cells" within the next decade and enabling the modeling of complex biological systems and questions regarding the origin of life.
- AI is expected to turbocharge the "tree of knowledge" across disciplines, potentially leading to breakthroughs in physics such as room-temperature superconductors, optimized batteries, and testing the limits of physical laws, including the possibility of faster-than-light travel.
- Long-term timelines suggest von Neumann probes could populate the galaxy within a million years if humanity solves climate change and achieves radical abundance, though the scarcity of extraterrestrial signals presents a potential Fermi paradox filter.
- The evolution of general intelligence is described as an extremely costly filter, with consciousness and intelligence considered double-dissociable traits; the outlook advocates for developing non-conscious AI tools while exercising caution regarding the emergence of sentience.
- Economic and social projections include the possibility of "radical abundance" fostering cooperation over competition, with AI expected to belong to humanity rather than any single entity, accompanied by increased societal tendencies to anthropomorphize these systems.
- Future applications will span both commercial and non-profit sectors, necessitating rigorous attention to safety and ethics, with the ultimate goal of entering an "Amazing Era" characterized by the solution of disease and significant scientific flourishing.