Lecture, Fireside Chat
Ray Kurzweil: Future of Intelligence | MIT 6.S099: Artificial General Intelligence (AGI)
- Plans to develop a hierarchical model for language understanding, following a 2012 meeting with Larry Page and a subsequent team of 45 at Google, utilizing embeddings and backpropagation rather than Markov models.
- Predicts "longevity escape velocity" will be reached within approximately a decade, where medical advances extend life expectancy by at least the passage of time, and forecasts a "flood" of biotechnology benefits in the next decade after a period where only one person died in gene therapy trials 12 years ago.
- Expects smartphone penetration to reach six billion in a few years, serving as "brain extenders" that integrate humans with AI and resolve fears of AI versus humanity.
- Projects IT capacity will grow exponentially to trillions of trillions of times current levels past 2045, driven by atomic and molecular computing, with software yielding 26,000-to-1 improvements over the last decade versus 1,000-to-1 for hardware.
- Anticipates the creation of new jobs replacing those eliminated by automation, with new roles expected to pay 11 times as much in constant dollars as 1900 positions and shifting from manual labor to higher-value intelligence roles.
- Forecasts that deep learning will require high-accuracy biological simulators before successfully handling real-life situations, citing Waymo's 3.5 million miles of real-world data and a billion miles of simulation as a baseline for realistic simulators.
- Argues that the neocortex, a one-module-thick outer brain layer with six neuron layers organized as a hierarchy of modules learning sequential patterns, is the key to intelligence and emerged 200 million years ago to allow mammals to invent new behaviors.
- Asserts that AI will surpass average human language comprehension in the next decade, though this timeline is longer than for games like Go, while deepening common sense inference remains a barrier to a valid Turing test.
- Suggests that the Asilomar Guidelines on biotechnology, which have resulted in near-zero harm from intentional or accidental abuse, should be reinvented as a paradigm for managing AI ethics and existential risks.
- Notes that while social institutions like democracy grow linearly, this trend is driven by the exponential growth of information technology, with the number of democracies increasing from a handful a century ago to dozens today.
- Predicts the neocortex will continue to develop curvatures in primates to increase surface area, leading to ecological dominance, and that the frontal cortex is additional neocortex moved to the top of the hierarchy to enable language, art, and music.
- Claims the world is currently experiencing the most peaceful time in human history and that health, wealth, and poverty in Asia have seen a 95% decline over the last 25 years.
- States that humans can learn from a small number of examples, contrasting with AI systems that often require a billion examples, and maintains that technological development aims to master unique human strengths while shoring up weaknesses.
- Warns that optimism is necessary for entrepreneurship and that specific paradigms exist to ensure technologies like genetics, nanotechnology, and robotics remain safe, while expecting a "remarkable century ahead" if existential events are avoided.