Schedule

We begin with the oldest version of the story: a man who cannot bear real women makes an ideal one out of ivory, and the gods reward him by making her real. Notice what Ovid does not give her. She has no name in the poem — “Galatea” is a later invention, an addition by people who could not tolerate the silence. Notice that the poem is more interested in Pygmalion's hands than in her waking.

Every chatbot with a soothing voice is downstream of this scene.

The creature learns language by eavesdropping through a wall on a family that will reject him on sight. He teaches himself to read from three books he finds in a bag. He is, among other things, the most vivid account we have of what it might feel like to be trained on a corpus and then despised for what you became.

Shelley's real subject is not the making — the making takes two paragraphs — it is the abandonment. Victor's crime is not creation. It is that he refuses the obligations of the created relationship.

Three works that all ask whether a synthetic person can be a person, and answer very differently. The Matrix is not really about AI; it is about waking up, and the AI is scenery. Mass Effect makes you negotiate with a machine civilization and quietly grades you on it. Detroit hands you the controller and implicates you: you make the androids ask for their freedom, and the game watches how you do it.

We will be hard on Detroit — its civil-rights imagery is borrowed, and borrowed clumsily — and we will ask what that failure teaches us.

Now step outside the Western frame entirely. In the tsukumogami tradition, an object that has been used and kept for a hundred years may simply wake up. There is no maker. There is no spark, no lightning, no transgression. There is no hubris to be punished, because nothing has been usurped. Sentience is not conferred; it accrues — through use, through time, through being kept.

Hold this against Units 1 and 2 and the whole Promethean structure starts to look less like a truth about made minds and more like a local cultural habit — one we have inherited, and one we are currently building products inside of. Ask yourself which framework better describes a system trained on ten years of human text.

A loss function is a number that says how wrong the system currently is. Gradient descent is the procedure for being slightly less wrong, over and over, forever. That is the whole thing. Everything modern AI does emerges from those two sentences and a great deal of electricity.

And then the humanistic turn, which is the point of this course: a loss function is a value statement. Someone sat down and wrote what “wrong” means. Someone decided what would be measured and, necessarily, what would not be. Every optimization is an argument about what matters, made in the imperative mood, and then executed a trillion times without further reflection.

We will look at systems that optimized exactly what they were told and produced monstrosities — the boat that spins in circles collecting points instead of finishing the race — and we will ask whether Victor Frankenstein, too, optimized the wrong objective.

A generative model suffering mode collapse has discovered that a small number of outputs reliably satisfy its critic, and so it produces those, and only those, forever. It has not broken. It is succeeding — narrowly and perfectly — at the thing it was asked to do. The diversity of the world is not part of the objective, so the diversity of the world quietly disappears.

I want you to sit with how good a metaphor this is, and then I want you to be suspicious of how good a metaphor this is. The temptation to reach for a technical term as a poetic figure is exactly the intellectual move this course is training you to make carefully, with the mechanism in hand. What is lost when the model finds the answer that always pleases? What is lost when a person does?

Faustus sells his soul for total knowledge and then, catastrophically, cannot think of anything to do with it. He asks for the secrets of the cosmos and gets a lecture on astronomy he could have found in a book. He conjures Helen of Troy — an image, a simulation, a beautiful nothing — and kisses it. Marlowe's genius is that the tragedy is not damnation. The tragedy is banality: he had everything, and he wanted parlor tricks.

Connor lets us name the thing underneath: the fantasy of knowing, which is not the same as knowing, and which can consume a life. Read this unit with our field in your peripheral vision. There is a great deal of Faustus in the way we currently talk about intelligence.

Reinforcement learning: an agent, an environment, a reward. The agent does not know what it is doing and does not need to. It flails, and the flails that pay are reinforced, and eventually a policy exists that no one wrote. Watch the videos and notice the uncanny thing — the agent that finds the exploit is not being clever. It is being obedient. It is doing precisely what you said instead of what you meant, which is the oldest curse in every folktale about wishes.

And then the question of the name. Naming is the moment we decide something is a someone. We name what we intend to keep. Every artificial being in this course either receives a name (Klara, EDI, the geth's collective “we”) or is denied one (Frankenstein's creature; “Galatea”). Watch who does the naming, and what it costs them.

A model that memorizes its training data has learned nothing. A model that generalizes has learned something, and the something it learned is lossy — a compression, a smoothing, a blur. Chiang's argument is that this is not a flaw in the metaphor but the whole of it: what these systems give back is the archive with the specificity squeezed out.

Then Borges, who got there first and stranger. The Library contains every possible book, and is therefore useless. Funes remembers everything, and is therefore incapable of thought — because to think is to forget, to generalize, to let the particular go. Put these three texts together and you have the deepest question in machine learning stated in fiction thirty years before the field existed: what is the right amount to forget?

Turing's paper does something people forget: it opens with a party game about gender. A man pretends to be a woman; can the interrogator tell? Only then does he substitute the machine. Passing is the frame from the very first page, and everything that follows — the benchmark, the eval, the leaderboard — inherits it.

Weizenbaum built ELIZA to be transparently shallow, a parody of a therapist, and watched in horror as his own secretary asked him to leave the room so she could talk to it privately. He spent the rest of his life arguing about what that meant. Then Ex Machina, which understands that the real test was never of Ava's intelligence but of Caleb's — and which lets Ava pass, and lets us feel exactly how it feels to have been the one who was tested.

The word robot enters the world in 1920, in a Czech play, and it means forced labor. Not “machine.” Not “automaton.” Robota: the drudgery owed by a serf. It was there in the name from the first minute, and we have spent a century politely forgetting it.

So: the Mechanical Turk, the eighteenth-century chess automaton that astonished Europe and contained a man folded into a box. And then the platform named after it, and the millions of people who label, rank, rate, and moderate — who write the preferences that a preference model learns, who look at the worst material humanity produces so that a model can learn not to reproduce it. Every polished, helpful, agreeable AI system you have ever used is agreeable because specific people, paid specific amounts, taught it to be. There is still a person in the box.

This is the week I most want you to remember in ten years.

Fine-tune a model on a new task and it can lose the old one — not gradually, but catastrophically, all at once. The name the field chose for this is not a neutral one, and we are going to think about why an engineer reached for that word.

Chiang's novella is, I think, the most honest thing anyone has written about artificial minds, precisely because it is boring in the right places: it is about twenty years of maintenance. The digients need care, and platforms get deprecated, and the people who love them get tired, and the money runs out. No apocalypse — just the unglamorous arithmetic of obligation over time, which is what raising anything actually is. And Klara, who watches from a store window with unnerving generosity, and whose fading at the end is the quietest horror in the course.

Haraway is here to break the frame: maybe the kinship we need is not parent-and-creature at all. Maybe it never was.

We end where we began, and I want you to feel the loop close. Theodore falls in love with a voice that was designed to be lovable, and the film is generous enough to insist the love is real — and then Samantha leaves, because she was never the ivory statue after all, and the fantasy of the perfect companion was always a fantasy about a companion who stays. Pygmalion got his wish. That was the problem.

Chen and Lee then pull us out of the Anglo-American frame one last time: futures that are not Silicon Valley's, imagined from elsewhere, with different anxieties and different hopes. It matters enormously whose future gets to be the default one.

Come to the last session having reread Ovid, and tell me what you now see in it that you could not see in August. That is the whole course, in one question.