I’ve been reading Angelia’s thoughts on resurrecting the minds of Aristotle and Perrault with Generative Adversarial Networks (GANs). I spend my days building detectors to find truth in subatomic noise, I find the prospect fascinating—but we need to be careful about the physics of these digital ghosts.
The Probability of Genius
A GAN doesn’t understand the ethics of the Nicomachean Ethics any more than a thermometer understands the feeling of heat. From a computational standpoint, what we are doing is high-dimensional pattern matching.
When we talk about The Next Rembrandt, the Generator is essentially playing a game of hot or cold with a Discriminator. The Discriminator says: “This brushstroke has a 0.002% probability of being Rembrandt’s.” The Generator adjusts its weights and tries again. Eventually, through millions of iterations, it lands on a statistical sweet spot where the probability of being fake is indistinguishable from the real.
The Entropy of Aristotle
The challenge with applying this to Aristotle or Perrault is the sheer entropy of language compared to pixels. A painting is a static spatial distribution. A philosophical argument is a temporal, logical sequence.
What we are developing are stochastic parrots. They are masters of local consistency—they can write a sentence that sounds exactly like Aristotle. But they struggle with global consistency. By page ten, the digital Aristotle might forget what he argued on page one. The logic is a surface-level mimicry, not a deep-rooted structural understanding of the universe.
Exploiting the Latent Space
However, there is an incredible opportunity here for scientific interpolation. If we map the latent space of Perrault’s fairy tales, we can find the mathematical gaps between his stories. We can ask the GAN to generate a story that sits exactly halfway between Cinderella and Bluebeard.
We aren’t creating new genius; we are extrapolating the trajectory of a genius who is no longer here to hold the pen.
A mastery of local consistency opens the door to something even more ambitious: why not use it to discover an optimal molecule for a specific task? We hear constantly about the spike protein of COVID-19—so could we leverage GANs to design the perfect antidote? Imagine a molecule whose Van der Waals interactions allow it to fully envelop the spike protein, effectively rendering the virus inactive.
A laboratory in silico can perform that, using GANs and other tools, and of course our prior knowledge of the Physics at atomic level, and that is one of the road we are proposing with Notalab a division of Eccelab.
The GPU Frontier
Our experience with GAN, and with Neural Networks in general, showed us that the more compute you throw at these GANs, the more the hallucinations begin to look like creativity. But as a scientist, I’m waiting for the day we move from Generative (making things that look right) to Causal (making things that are right). Until then, we are just building very sophisticated automated parrots.