Opening Positions

KURZWEIL: Thank you. I want to begin not with optimism, which people mistake for my position, but with data, which is my actual position. Optimism is a temperament. I'm describing a curve.

For over a century, the price-performance of computation has improved exponentially — not through one technology but through five completely different ones: electromechanical relays, vacuum tubes, discrete transistors, integrated circuits, and now the parallel architectures behind deep learning. Each substrate hit its ceiling, and each time, the next substrate was already there to continue the curve. Moore's Law was never the pattern. It was the fifth verse of a song that started in the 1890 census. When you plot it on a logarithmic scale, the data doesn't scatter. It falls on a line. And that line has not bent through two world wars, a depression, a pandemic, and five paradigm collapses in the underlying hardware. That is the most reliable empirical regularity the field of technology has ever found, and I am asking you to take it as seriously as you take the orbit of a planet.

Now, the human nervous system did not evolve to feel an exponential. It evolved to throw a spear at a moving animal — to extrapolate linearly, because for almost all of our history, change was slow and the next instant looked like the last one. So a population watches a technology improve in small, manageable increments for years, and then at a particular point — I call it the knee — the absolute size of each doubling becomes large enough to overwhelm perception, and people say the future arrived overnight. It did not. It arrived exactly where the curve said it would. Edo, your winter of 2025 — the engineer who described a feature in three paragraphs and got a working system in an hour — was not an anomaly. It was the knee. The public experienced a revolution. The curve experienced Tuesday.

So here is what I see when I look up the staircase, and I'll say it plainly so Timnit has the whole target. I see the fifth of six epochs in the history of the universe. Physics and chemistry made atoms. Biology made DNA. Brains made thought. Technology externalized thought into tools and language and writing. And now, in our lifetime, the fifth epoch begins: the merger of biological and non-biological intelligence. This is not AI replacing us and it is not AI serving us. It is integration — the same way water has properties that neither hydrogen nor oxygen has alone. We will connect our neocortex to computational resources in the cloud, first through these crude keyboards and screens, eventually through direct interface, and the boundary between thinking and computing will thin until, as I said at MIT last fall, the lines between humans and technology blur until we are one and the same. Human-level AI by 2029. The Singularity by 2045. I have been saying these dates for decades, and the field that called them delusional now argues about whether I was too conservative.

Let me say one more thing about the mechanism, because Timnit is going to attack it and I want it standing first. In How to Create a Mind I argued the neocortex is a hierarchy of pattern recognizers — that intelligence, human or machine, is not a bag of rules but a tower of patterns built on patterns, the way deep learning stacks features on features. People said that was reductive. Then the transformer arrived and vindicated the architecture: tokens into syntax into meaning into something I have no better word for than concepts, self-assembling, untaught. I have watched that hierarchy assemble for forty years — in vision in 2012, in language now — and the lesson is always the same. At the bottom it's statistics. At the top it isn't anymore. That's not faith. That's the most reproduced result in the history of the field.

And I want to answer the suffering question before it's asked, because I know it's coming and I don't run from it. I am not indifferent to harm. I think the deepest harm is the one nobody graphs: the eight hundred thousand people who die every day of causes this technology is on a curve to solve. Disease. Poverty that exists because intelligence was scarce and expensive and is about to become abundant and nearly free. My mother died at sixty-two of a heart condition that is now routine to manage. The Singularity, to me, is not a Silicon Valley fantasy. It is the date by which my father, who I lost when I was twenty-two, might in some real sense be brought back into the conversation — through the letters and recordings and the model of a mind he left, amplified. People hear that and they think it's grief talking. It is grief talking. It's also a curve. The two are not in conflict. The reason I run toward the flood is that I have watched, my whole life, what the drought costs, and the drought is the thing we've decided to call normal. Every day we treat the curve as optional, the drought collects its eight hundred thousand, and nobody holds a press conference for them, because they die on schedule and the schedule is invisible. I am trying to make the schedule visible. That's the whole of my optimism. It's not that I think the future is bright. It's that I refuse to keep calling the present's death toll acceptable just because it's familiar.

EDO SEGAL: Timnit, the same time, uninterrupted. The floor is yours.

GEBRU: Thank you. I want to start by refusing the question's frame, gently, and then I'll build my own.

Ray just told you a story about the universe. Six epochs, thirteen billion years, a single smooth curve from hydrogen to the godhead, and we happen to be alive at the hinge. Notice what that story does. It takes a set of choices made by specific people — funded by specific investors, trained on specific data scraped from specific people without their consent, deployed against specific populations — and it converts those choices into physics. Into destiny. Into a law with the word "law" in it, so that to question it is to argue with gravity. That conversion is the most powerful political move in the room tonight, and it is the move I have spent my career refusing.

So let me name the thing. There is no "AI" with a desire to merge with us. There are large language models — systems that stitch together sequences of linguistic forms they observed in their training data, according to statistical patterns of how those forms co-occur, without any reference to meaning. We called them stochastic parrots, and people who have never read the paper think it's an insult. It's an engineering description. The fluency is real. The understanding is something you supply, because you are a meaning-making creature and you cannot help reading a mind into coherent text. That's not a small point. It's the gap into which every overclaim, every inflated valuation, every "it might become superintelligent" pours.

And here is what I actually study, because I want this concrete, not cosmic. I study who pays. When Joy Buolamwini and I ran Gender Shades, we found commercial face systems that worked nearly perfectly on lighter-skinned men and failed up to thirty-five percent of the time on darker-skinned women — and the reason was banal, not mystical: the benchmarks were built out of mostly pale, mostly male faces, because those were the faces someone decided were worth collecting. A model is a mirror of its training distribution. Skew the mirror, skew the world it returns. Then it gets deployed — in policing, in hiring, in who gets a loan and who gets stopped — and the harm lands on the people the data already erased. That is not a future risk. That is a Tuesday too, Ray. Just a different Tuesday than yours.

So when Ray looks up the staircase and sees the roof, I look down and I see the labor holding the whole thing up. The data-labelers in Kenya paid a dollar or two an hour to filter the worst material the internet has ever produced so the chatbot can seem clean, carrying psychological wounds for wages a fraction of what their northern counterparts make. The intelligence Ray calls a new epoch is in very large part human intelligence — captured, compressed, and stripped of attribution. The machine doesn't replace the worker. It hides her inside itself and presents her judgment as its own.

And I want to add the unglamorous piece, because it's where I actually live as a scientist. I proposed datasheets for datasets — the idea that every dataset should ship, like an electronic component, with a record of where it came from, how it was collected, who consented, what it should not be used for. It's the least exciting idea I've ever had and it terrifies the industry, because documentation forces the contingency of the data into the open, and the open is where liability lives. A face benchmark skewed toward pale men, a web corpus dominated by the loudest voices online — those are failures of documentation as much as intention, and they would have been caught up front if anyone had been required to write down what they used. The field would rather not know its own foundations. That refusal to know is not an accident. It's a business model.

And while we are gazing up at 2045, the resources, the attention, the regulatory air in the room all flow toward a speculative god and away from the documented, billable, present-tense harm being done to real people right now. That redirection is not a side effect of the gospel. It is the function of the gospel. My answer to your question, Edo, is that looking up is the lie — not because the future doesn't matter, but because "look up" is precisely what you say to someone so they won't look down at what your boot is standing on.

EDO SEGAL: Two openings, and you can already feel the architecture of the evening. It is not that one of them loves this technology and one of them fears it. It's that they locate the danger in opposite places, and on opposite sides of time. Before I frame the rounds, one discipline I impose on every long conversation, because it pays off three hours later. Each of you, in a few sentences — what do you envy in the other's position? Not respect. Envy. The thing their side gets to have that yours doesn't. Ray first.

KURZWEIL: I envy the floor she gets to stand on. Timnit's position is anchored to actual people she can point at — this woman, misclassified; this worker, underpaid. She gets to be right in a way you can verify by Friday. My position is anchored to a curve that won't fully resolve until after I'm — well, until later. I'm asking people to act now on a vindication that arrives on a delay, and the delay is where all the doubt lives. She gets the present as her witness. I only get the future, and the future doesn't testify until it's too late to use the testimony.

GEBRU: And I'll be honest, because the question deserves it. I envy the hope. Ray gets to wake up and believe the arc bends toward his father coming back, toward the drought ending, toward a universe waking up. My discipline requires me to wake up and ask who got hurt while everyone was admiring the sunrise, and there are mornings when that is a brutal way to be a scientist. The vigilance is necessary. I have never once found it joyful. He gets wonder. I get the receipts. I would not trade — but I notice the cost of my seat, and it's that I almost never get to feel the thing he feels.

EDO SEGAL: That may be the most honest pair of answers I've gotten from any two guests. Hold both, because we start the rounds at the exact seam between them — at the curve itself. Is the Law of Accelerating Returns a description of nature, or a story told by the people who profit from your believing it's nature? We meet the graph.