Large Language Models Are Like The Planet Solaris
It’s been a little while since I watched Andrei Tarkovsky’s Solaris, the Soviet director’s 1973 movie adaptation of a classic sci-fi novel of the same name written by Stanislaw Lem. But the other day, it suddenly struck me that, for being a work from the Cold War, Solaris provides a metaphor that I have been grasping for—a metaphor that helps explain how large language models are able to learn and how they can generate information based on their learnings....