Engineering Intelligence: Lex Fridman's Opening Lecture for MIT's AGI Course

Fridman opens the course by explaining that its mission is to engineer intelligence, in keeping with MIT's motto "mind and hand." The course will explore the fundamental science of what makes an intelligent system, but it always wants to ground that in building systems that can make for a better world today.

He positions this as a counterweight to the better-known "black box" view of AGI, which asks what happens to society once a human-level system exists. Will robots take over, or will we reach a utopia with no messy jobs? Fridman finds those questions interesting but says they are not what the course is about.

Instead he wants to look inside the black box and build intuition for how systems approaching human-level intelligence might be created. He believes we are very far from that, though "far" may not be measured in time: a few breakthroughs, or even one, could change everything. It is not constructive, he argues, to discuss safety and ethics without understanding the actual methods.

The fundamental disagreement among the speakers, he says, sits at the core of that black box: how hard is it to build a human-level system? Despite impressive results in deep learning, neuroscience, cognitive science, and robotics, nobody agrees on how far there is still to go.

Fridman shares a meme contrasting a Google intern writing a for loop that grid-searches neural network parameters with a media headline that says the Google AI created its own baby AI. The course's first goal is to avoid that kind of futurist thinking, hype detached from what systems actually do, while still taking seriously topics like AI safety and autonomous weapons later in the course.

His second goal is the reverse: not to treat the for loop as a reason to ignore consequences. He compares the idea of "I'm just the engineer" to the stance of scientists building nuclear weapons during World War II. Near-term, preventable harms are the low-hanging fruit that the engineering process itself can catch.

He also warns against dismissing a breakthrough just because current systems look limited at common-sense reasoning. Technology adoption has been getting faster every decade, so a new idea can have widespread effects almost overnight. He uses the image of a dark room where we feel around for a light switch: some people say it must be close, but in reality we know very little about where it is.

For the positive motivation, Fridman points to a human compulsion to explore, citing Stewart Weaver's short book on exploration, from ancient voyages to the Arctic and Darwin's Beagle to Yuri Gagarin and the moon. The jobs, ethics, and business debates aside, he says, what drives him and much of the community is the desire to uncover the mysteries of the universe.

Students register on the course website, submit links to vote.ai, an aggregator of articles, papers, and videos on AGI where people vote up or down and tag sentiment as positive or negative, and enter one of three competitions. Fridman says the course also offers free shirts to people who attend the last lecture in person.

The first project, DreamVision, builds on Google's DeepDream and the idea that creativity is a mark of intelligence. Participants use neural networks to visualize what the networks see, mixing dream and reality across video streams. The winner is picked through a Mechanical Turk competition for the most beautiful visualization.

The second, Angel, short for artificial neural generator of emotion and language, is a twist on the Turing test that uses emotion instead of words. A customizable face with 26 muscles is controlled by an LSTM, and participants have ten seconds to impress viewers in A/B tests. The course staff will also enter by mapping their own faces onto the system and trying to convince viewers to keep them as friends.

The third, the ethical car, extends the DeepTraffic reinforcement learning competition with unpredictable pedestrians and puts human life into the loss function. Fridman says this is not an ethical question but an engineering one: participants must decide the trade-off between getting from A to B quickly and hurting people, as governments and insurers already put a price on human life.

Fridman walks through the speaker lineup, which includes Josh Tenenbaum, Ray Kurzweil, Lisa Feldman Barrett, Nader Pinsky, Andrej Karpathy, Stephen Wolfram, Richard Moyes, Marc Raibert, and Ilya Sutskever. Most, he says, are not interested in the future impact on society without grounding it in the engineering. Tenenbaum, a computational cognitive scientist, will speak about common-sense understanding and intuitive physics, and learning from very few examples, as children do, rather than "memorization engines."

Kurzweil will give a whirlwind account of how natural intelligence is understood and how it relates to the exponential growth of AI. Fridman is especially excited about Lisa Feldman Barrett, whose book argues that emotions are constructed rather than hardwired. If emotional expression is learned social norms, he says, then emotional intelligence is a learning problem that machines can in principle take on.

He cautions about Sophia, the robot: she is an art exhibit, not a strong natural language processing system and not AGI. Her humor and embodiment are enough to captivate people, which shows how easy it is to trick humans into seeing intelligence. That suggests one route: machines could generate the surface display of emotion while the underlying technology stays trivial.

The Angel project strips away appearance and tries to see whether the simplest face can make viewers feel something. Fridman admits it is an open question for his own team, asking whether it could become human before our eyes and whether someone might even fall in love with Angel.

Nader Pinsky will cover cognitive modeling architectures, and Karpathy, the "state-of-the-art human" on ImageNet, will speak about the role and limits of deep learning. Fridman describes representation learning as taking complex data and reducing it to simple, essential elements. In the classic example, a straight line cannot separate blue from red in the input space, so the network learns a transformation where it can.

He contrasts the biological and artificial. The human brain has about 100 billion neurons and 1,000 trillion synapses, while ResNet-152 has around 60 million. The brain's topology is chaotic and asynchronous and its learning algorithm largely unknown, whereas artificial networks use backpropagation over a defined loss and need separate training and evaluation stages. The brain is also far more power efficient.

Fridman then maps learning paradigms onto cognitive terms. Supervised learning, where humans annotate the data, he likens to memorization. Semi-supervised and reinforcement learning, with sparse labels, are closer to reasoning. Unsupervised learning, which forms representations with little or no human input, is the hope for something like understanding.

The open challenges he lists include transferring between domains, the need for huge amounts of supervised data, hyperparameter tuning, and human-defined loss functions that are flawed when released into the real world. Real systems have no ground-truth test set: the goal is an autonomous car that does not kill pedestrians, and edge cases are where methods have to generalize.

Fridman previews Stephen Wolfram, whose knowledge-based programming through Wolfram Alpha and the Wolfram language builds a deep graph of knowledge, and who worked on the language in the film Arrival. Wolfram's earlier work on cellular automata shows how extremely simple local rules can produce complex patterns, which Fridman links to both neural networks and knowledge formation.

Richard Moyes of Article 36 will discuss the legal, policy, and technical aspects of banning autonomous weapons, where he says there is broad agreement about the safety hazards. Marc Raibert, CEO of Boston Dynamics, will discuss legged robots in the real world and how society reacts, with fear, excitement, and hope, once intelligence is embodied.

Ilya Sutskever, an OpenAI co-founder, will cover game playing with deep reinforcement learning. Fridman notes AlphaGo Zero in 2017, which learned to beat the best in the world through pure self-play from scratch, not expert games. He lays out a stack of perception, representation, reasoning, and action, and asks how much of it can be learned end to end from raw sensory data to action.

Deep learning has already reached better-than-human performance on ImageNet classification and segmentation, but whether it can reason is open. Fridman closes with the lectures, three projects, and speakers to come, plus later sessions on AI ethics and bias, creativity, brain simulation, and a natural-language Turing test project in March.