Machines, Creativity & Love | Dr. Lex Fridman

The episode opens by framing the exchange as something other than a technical survey. Huberman introduces Fridman as an MIT researcher specializing in machine learning, artificial intelligence and human-robot interactions, and ranks the conversation among the most fascinating of his career and his life. What he says he discovered is a personal dream about humans and machines and about how such interactions can change the way we perceive ourselves and the world. He deliberately broadens the topic to relationships of every kind, then adds relationships with machines, distinguishing machines that move from machines that do not, and describing a class of machines that could come to understand us in ways we cannot understand ourselves, and so educate us about ourselves. His culminating claim is carefully qualified: these interactions can transform the self and may very well transform humanity. He also situates the podcast outside his Stanford teaching and research, as part of an effort to bring zero-cost-to-consumer science information to the public.

The show is introduced as a discussion of science and science-based tools for everyday life, and Fridman opens with a repeated greeting that signals an established rapport between the two rather than a first meeting. That familiarity helps explain the register of what follows: the conversation repeatedly moves from technical definition to personal vision without ceremony.

Huberman then states the motivating gap from the listener's side: the terms are heard constantly, yet he judges that most people, including most scientists and himself, do not really know what artificial intelligence is or how it differs from machine learning and robotics. His request is definitional - explain what artificial intelligence is, and what machine learning is. The sequence matters, because the entire first technical movement is an attempt to answer a question about the boundaries between labels, and that same concern with definitions recurs later as a claim about the culture of the field.

Fridman resists a single definition and instead stacks several senses of artificial intelligence: (a) a philosophical longing, citing Pamela McCorduck's line that AI was born as the ancient wish to forge the gods, and our desire to create other intelligence systems, perhaps systems more powerful than us; (b) a narrow set of computational and mathematical tools to automate tasks; (c) an attempt to understand our own mind by building systems that exhibit intelligent behavior; and (d) what it means to the community in practice - a set of computational techniques that allow you to solve various problems. He insists all of these are true at once, and ties the difficulty of the question to the difficulty of defining intelligence itself.

On the historical thread, Fridman describes machine learning as the community within AI that emphasizes the task of learning: how to make a machine that knows very little in the beginning follow a process and get better and better at a particular task. He dates the most effective recent techniques to the last about 15 years, grouping them under deep learning and neural networks. He is explicit that these are inspired by the structure of the human brain only very loosely, and describes them as a network of basic computational units called artificial neurons, with an input and an output, that know nothing in the beginning and are tasked with learning something interesting.

The first breakdown he offers is by how much human supervision is required. In supervised learning, the network starts knowing nothing and is given examples - in computer vision, cats, dogs, cars, traffic signs - together with the ground truth of what is in the image, and a large database of such labeled examples lets it learn by example. Fridman immediately turns this into an open problem: how do you actually provide the truth that a given image contains a cat?

He lays out a ladder of increasingly demanding ways to supply that truth: (1) label the entire image as a picture of a cat; (2) draw a crude bounding box around the cat's face; (3) perform semantic segmentation, drawing a crisp outline around the cat. His caveat is that this is a 2D image and the computer knows nothing about our three-dimensional world, it is just looking at a set of pixels. That produces what he calls one of the fundamental open questions in computer vision: is that even a good representation of the truth? He then names the contrasting family, formerly called unsupervised learning and now commonly called self-supervised learning, whose goal is less and less and less human supervision.

Self-supervised learning, he says, has been very successful in language models and natural language processing and is increasingly successful in computer vision. The idea is to let a machine look at pictures or text on the Internet without any ground truth annotation and learn something generalizable about the core ideas of language or vision - what humans, at its best, like to call common sense, a giant base of knowledge on top of which more sophisticated knowledge is built. The dream is concrete: let a self-supervised system run around the Internet and watch YouTube videos for millions and millions of hours, then be primed to learn from very few examples, the way human children are taught a concept with only one or two examples from their parents. A cat would then be understood as more than pointy ears, or being orange, or furry. Fridman grounds the ambition in a limit of introspection: if asked what makes a cat versus a dog, most of us could not say, yet we can tell the difference, and the whole dream is that a machine would learn on its own the set of common-sense knowledge the brain uses to make that distinction. He links this to self-play mechanisms behind reinforcement learning successes, naming Go and AlphaZero that won at chess.

Huberman briefly checks his understanding that these systems play games, and Fridman confirms it, then generalizes: self-play is a system that plays against itself, knows nothing in the beginning, creates a bunch of mutations of itself, and competes against those versions. The mechanism he emphasizes is that playing against systems a little bit better than you makes you better yourself - a principle he extends from martial arts to learning in general - and that iterating this process drives everybody to get better and better until the system is several orders of magnitude better than the world champion in chess, for example.

Fridman calls this a runaway system and reports a striking claim from David Silver, a leader of the AlphaGo and AlphaZero team: they haven't found the ceiling for AlphaZero, meaning it could just arbitrarily keep improving. He argues this does not matter in chess, a domain where the system simply ran away with the game and is so much better than humans, but becomes a terrifying process if reproduced in a realm with a bigger, deeper effect on human beings and societies. His conditional resolution is supervision: it is an exciting process if you supervise it correctly, by injecting what is called value alignment, making sure the goals the AI is optimizing are aligned with human beings and human societies. He names self-supervised learning, with less and less human supervision, as the really big, exciting open area.

Fridman pivots from theory to what he calls the most exciting application: Tesla autopilot, precisely because these are systems working in the real world, this isn't an academic exercise, this is human lives at stake, this is safety critical. A short exchange with Huberman stages the field's terminological instability in real time - Huberman offers automated vehicles, then autonomous vehicles, and Fridman corrects to semiautonomous, noting we've gone through wars on these topics. The substantive point is that even though the feature is called FSD, full self-driving, it is currently not fully autonomous: human supervision is required, the human is tasked with overseeing the system, and liabilitywise the human is always responsible. Fridman treats that as a human factor psychology question within human-robot interaction, and frames the open problem as how do humans and robots dance together.

Here a concrete disagreement surfaces. Fridman reports that Elon does not see semiautonomous driving as a durable state but as a stepping stone towards fully autonomous driving, on the view that humans and robots can't dance well together - let humans and humans dance and robots and robots dance - and that the engineering task is to design a perfect robot that solves the problem. Fridman's counter-thesis is that the world will be full of problems where humans and robots always have to interact, because robots will always be flawed just as humans are, and that is what makes life beautiful: learning happens at the edge of your capabilities. The goal, in his phrasing, is to figure out how flawed robots and flawed humans can interact so that the sum is bigger than the whole, rather than focusing only on building the perfect robot.

Fridman then describes the learning loop that Andrej Karpathy, head of autopilot, calls the data engine, and notes it applies to a lot of machine learning: (1) build a system that is pretty good at doing stuff; (2) send it into the real world, where it runs into edge cases - failure cases where it screws up; (3) recognize that there are millions of edge cases, weird situations you did not expect; (4) collect those edge cases via a weird detector across hundreds of thousands of cars and send that data back to the mothership for training and retraining; (5) go back to the drawing board, learn from them, and send out a new, updated version, repeating the cycle so the system keeps getting better. His summary is to send out a pretty clever AI system and let it screw up just enough to figure out where the edge cases are.

When Huberman asks whether the updating is done by humans, Fridman answers that the annotation is done by humans: the weird examples come back and must be labeled, with some mechanisms for automatically labeling, but mostly you always have to rely on humans to understand what is happening in the weird cases. He then folds the point back into the definitional theme - what is artificial intelligence is also a question of smart people having very different opinions about what intelligence is - and caps the section with a blunt characterization of the field.

AI is basically a community of people who don't agree on anything.

Huberman names the pattern directly: as an outsider, the community seems to have very little consensus about what the terms and operational definitions even mean, with splitting not just between supervised and unsupervised learning but into intermediate conditions where machines are autonomous and then go back for more instruction - his analogy is kids coming home from college in the summer, still fed by mom, before eventually leaving the nest. He asks whether something about engineers or this realm of engineering lends itself to disagreement.

Fridman's answer is a gradient: the more specific you get the less disagreement there is. There is a lot of disagreement about what artificial intelligence is, less about machine learning, still less about active learning or machine teaching or self-supervised learning, and less and less and less about NLP language models or transformers and specific neural network architectures. He attributes the high-level friction to engineering being a little bit of an art and a science when it comes to intelligence systems: the art part creates disagreements about how easy or difficult a particular problem is. His example is that a lot of people disagreed with Elon about how difficult autonomous driving is, and nobody knows; more broadly there is disagreement about the limits of these techniques, and the terminology contains within it the disagreements.

He adds structural causes: it is a young science as a large-scale discipline - thousands, tens of thousands, hundreds of thousands of people, huge amounts of money being made - which is a very recent thing, so the terms are still being figured out; and there are egos, personalities and a lot of fame to be made. His sharpest historical claim is that neural networks have been around for many, many decades, since the sixties and arguably the forties, and that renaming them deep learning was part of the reinvigoration of the field while being really the same exact thing.

Huberman, speaking as a neuroscientist, corroborates the rebranding from the adjacent field: computational and theoretical neuroscience had their own journals and was not taken terribly seriously by experimentalists until about five to seven years ago, and he names Larry Abbott among excellent theoretical neuroscientists. He admits he did not know that neural network works in deep learning and that the two have become kind of synonymous. Fridman confirms they are always the same thing, and adds that networks probably means something a little different depending on the field, noting that neuroscience and AI people have started working together and dancing a lot more in the recent, probably, decade.

No, they were always ... well, no, they're always the same thing.

I'm a neuroscientist and I didn't know that.

Huberman is fixated on the self-play image of competing against a mutated version of yourself, and mirrors it against human learning - repetitions, errors, occasional success, his example being throwing darts at a bull's-eye, even playing darts against a version of himself wearing a visual prism. He presses on the biology: mutations sometimes give us superpowers, somebody with better than 20/20 vision who sees better than 99.9% of people, and asks whether a machine's mutation is always a negative one, or could be adaptive or maladaptive.

Fridman's answer is that you mutate first and figure out afterwards, then the variants compete; Huberman interprets this as the machine getting to evolve itself in real time. Fridman elaborates that usually you freeze a version of the system - take an Andrew of yesterday and make 10 clones of them, mutate maybe or not - and run competitions of the Andrew of today, with a fight to the death until a winner remains. He jokes about clones of himself from each day of the past year competing at podcasting, science, picking up chicks at a bar, or jujitsu, but the technical point is that reinforcement learning through self-play is a graveyard of systems that didn't do that well, where the good ones survive.