World Models & AGI
You can picture a glass falling before it hits the floor. Nobody taught you that. You watched the world long enough to guess what happens next. That guess in your head is a world model.
Chatbots don't have one. They've read every sentence ever written about glasses and never seen one fall. Yann LeCun likes to say a house cat understands the world better than the smartest chatbot, because the cat learned by watching, not reading.
One way to teach a machine is trial and error. It tries something, gets rewarded or punished, and slowly figures out what to do. The simplest version just remembers what worked and what hurt. To learn that lava is bad, it has to fall in. Again and again. It learns to stay away from lava without ever understanding what lava does.
The smarter version builds a little copy of the world in its head first, then practices in there. DeepMind's Dreamer works like this. It imagines thousands of possible futures, falls into the lava in its dreams, learns from every imagined mistake, and wakes up knowing what to do. It barely has to get hurt for real.
LeCun thinks this is how machines will finally get smart: watch the world, understand how it works, think before acting. Not just read more.
Snoop, down below, has read everything and seen nothing. Watch him make one real mistake, then show him how his world works. After that, he dreams his way to the diamond.
Fun fact: Dreamer was the first AI to find diamonds in Minecraft entirely on its own, with no human gameplay to copy. It figured it out by dreaming.
