Is it possible to build a bridge in a single day with a million workers? Can a lemon produce a thousand liters of juice if squeezed for long enough? These “disanalogies” show how surprising AI’s scalability laws are. Jensen Huang illustrates how the power of artificial intelligence systems grows exponentially from training to fine tuning, to execution, surpassing our intuitions about the physical world. The video also explores the generalization of Moore’s Law and Ray Kurzweil’s approach to accelerating returns, revealing how innovation doesn’t run out but continues to expand AI’s possibilities.
I want to illustrate why the power of artificial intelligence in modern architectures is so unexpected, even unreasonable, using a series of examples that I call “disanalogies,” paradoxical analogies that show how what we expect in the physical world doesn’t correspond at all to what happens in the world of artificial intelligence.
I start from the slides that Jensen Huang, founder and CEO of NVIDIA, presented at the Consumer Electronics Show in January 2025. In this slide, in particular, he shows how scalability laws apply in successive waves across different phases of development and application of modern artificial intelligence engines. Each of these laws develops exponential power, and it’s the envelope of these various scalability laws that we observe in the increasingly powerful applications we use every day.
What happens when we train the engine? Naturally, we would expect to maximize artificial intelligence power as we would in building a bridge. If it takes a thousand people to build a bridge in a thousand days, is it reasonable to expect that a million people could build that same bridge in just one day? Paradoxical, right? Yet that’s what happens in artificial intelligence systems. Or, if we want to “squeeze” the system to produce intelligence and we keep squeezing more and more, it’s a bit like squeezing a lemon that produces a certain volume of juice and deciding, no, let’s squeeze it longer. And after ten hours, we discover that single lemon has produced a thousand liters of juice. This is what happens in artificial intelligence, where we manage to extract more “intelligence juice” by training longer.
When the system is trained, it is balanced so that its characteristics allow certain yields in different application areas. It’s then possible to intervene with what is called fine-tuning mode, that is, specializing the model in a particular area to highlight capabilities that were previously latent but not sufficiently accessible or developed. It’s a bit like what we see every day: you give your car a new coat of paint and, gleaming, it starts to fly. Right? The car’s latent capacity, which no one presumed, manifests itself: it’s not only shiny, but it flies too. Or, knowing that ten new guests are coming to your house, you don’t worry because the house develops rooms that weren’t there before, where you can accommodate them. And if a thousand arrive, that’s fine too.
More recently, we have begun to introduce models that are able to adapt their behavior not during training, not during fine-tuning, but during execution, through the dialogue itself with users, where the system can be invited to reflect on what it does, focusing more. Perhaps that’s what happens when you decide that, faced with a particularly complicated problem, the right thing to do is drink a thousand cups of coffee, because this will allow you to focus better and support the necessary reasoning. And no, that’s not how our brain works, our physiology, but it is exactly what happens in artificial intelligence systems.
The ensemble, the envelope of these unexpected and unreasonable scalability laws, which we continue to observe, represents a generalization of Moore’s Law, a generalization of the approach expressed and analyzed by Ray Kurzweil in his law of accelerating returns. And it’s fantastic to be able to take advantage of it. Just as in the past those who said Moore’s Law was dead were wrong, because the creativity and passion of engineers worldwide have allowed it to continue through increasingly different and innovative approaches, so today those who claim that AI is already finished, that this wave is running out, don’t realize they’re looking at a single exponential curve instead of recognizing the paradigm, the envelope of curves that outline the innovation we will continue to explore.