Ray Kurzweil: Future of Intelligence | MIT 6.S099: Artificial General Intelligence (AGI)

In this MIT 6.S099 AGI lecture, Ray Kurzweil recounts the early history of AI, from meeting Marvin Minsky and Frank Rosenblatt's perceptron to the modern deep-learning explosion driven by many-layer neural nets and exponential growth in computing. He explains his theory that the neocortex is a hierarchy of modules each learning simple sequential patterns, and how this hierarchical view shapes his Google team's work on language understanding. He argues the world is fundamentally hierarchical, which is why purely flat deep nets fall short and why his approach uses embeddings within a hierarchy. Kurzweil makes the case that humans will continuously enhance their intelligence by merging with AI, and predicts longevity escape velocity within about a decade. He closes by addressing job displacement, existential risk, and the optimistic view that technology is an expression of humanity.

• Kurzweil claims Minsky actually invented the neural net in 1953 but turned negative on it due to overhype.
  00:03:10
• Solving the vanishing/exploding gradient problem enabled 200-layer neural nets behind recent deep-learning gains.
  00:05:51
• Kurzweil's 1962 paper proposed the neocortex is organized as a hierarchy of modules, each learning a simple pattern.
  00:11:34
• Larry Page read Kurzweil's book in 2012, then offered to buy his two-week-old startup, bringing him to Google.
  00:20:28
• Kurzweil calls natural language understanding the holy grail of AI and a Turing-complete problem.
  00:21:34
• He predicts longevity escape velocity is only about a decade away, with diligent people potentially there already.
  00:27:18
• He argues there is no software subroutine that can guarantee a superintelligent AI stays safe.
  00:48:25

• Kurzweil first visited MIT in 1952 at age 14, six years after AI got its name at the 1956 Dartmouth conference.
  00:02:39
• AlphaGo Zero started with no human input and surpassed the best players within hours of self-play.
  00:05:51
• Kurzweil cites a field motto that 'life begins at a billion examples' for training deep nets.
  00:06:54
• The European brain reverse-engineering project found a repeating ~100-neuron module repeated 300 million times in the neocortex.
  00:12:38
• Stretched flat, the human neocortex is about the size and thinness of a table napkin.
  00:18:24
• The cerebellum has more neurons than the neocortex and stores simple scripts like writing a signature.
  00:32:01
• In 1900, two-thirds of the workforce was on farms and in factories; by 2015 that fell to roughly 11%.
  00:34:43
• A century ago the number of democracies could be counted on one finger; now there are dozens.
  00:38:51
• An Obama-era study found ~1,000x gains from hardware and ~26,000x from software/algorithmic improvements over a decade.
  00:51:35

Books, products and media the guest or host genuinely endorsed here — with the buy link.

Affiliate link — we may earn a commission at no extra cost to you.