MIT AGI: Artificial General Intelligence

This is the opening lecture of MIT's Artificial General Intelligence course, delivered solo by Lex Fridman. He argues for grounding AGI discussion in actual engineering rather than black-box philosophical speculation, while still taking the societal stakes seriously. He uses the metaphor of feeling around a dark room for a light switch to describe how little we know about how hard building human-level intelligence really is. He previews the course's three projects (DREAM vision, ANGEL emotion generation, and the ethical car) and the aggregator VoteAI, then walks through the roster of guest speakers and the perspectives each brings. He closes by contrasting human and artificial neural networks and posing the open question of how much of the AI stack can be learned end to end.

• Fridman states he believes we are very far away from creating anything resembling human-level intelligence, though one breakthrough could change everything.
  00:02:06
• He frames the singularity as the moment we are truly surprised by the intelligence of systems we create, using a dark-room-and-light-switch analogy for our ignorance.
  00:10:02
• He warns that Sophia the robot is an art exhibit, not a strong NLP or AGI system, and a demonstration of how easily humans are tricked into perceiving intelligence.
  00:27:13
• He contrasts the human brain (100 billion neurons, 1000 trillion synapses) with ResNet-152's ~60 million parameters, several orders of magnitude apart.
  00:32:28
• He notes alphago zero learned to beat the best in the world purely from self-play, without training on expert games.
  00:44:30
• He poses the course's central open question: how much of the AI stack, from raw sensory data to action, can be learned end to end.
  00:46:40

• Andrej Karpathy is described as famous for being the state-of-the-art human on the ImageNet challenge, representing roughly 95% accuracy.
  00:05:48
• Fridman quotes Stewart Weaver's book 'Exploration: A Very Short Introduction' on exploration as a defining element of human identity.
  00:11:39
• He cites a roughly 7,500-mile ocean journey from 325 BCE to explore the Arctic as an early example of human exploration.
  00:12:41
• Yuri Gagarin, first human in space in 1961, radioed that 'the earth is blue, it is amazing.'
  00:13:14
• Lisa Feldman Barrett's book 'How Emotions Are Made' argues emotions are constructed, separating bodily physical state from facial expression of emotion.
  00:24:16
• Backpropagation is described as a trivial, constrained learning algorithm compared to the largely unknown and far more complex learning process of the human brain.
  00:33:01
• Stephen Wolfram and his son Christopher worked on the alien language for the film Arrival, with Wolfram brought in as a representative human interpreter.
  00:40:50
• Wolfram's cellular automata work shows extremely complex patterns emerging from very simple local rules starting from a single cell.
  00:41:54