On the Edge: The Art of Risking Everything

by Nate Silver

Just like poker players seek to maximize EV, LLMs seek to minimize what’s called a “loss function.” Basically, they’re trying to ace the test—a test of how often they correctly predict the next token from a corpus of human-generated text. They lose points every time they fail to come up with the correct answer, so they can be clever in their effort to get a high score. For instance, if I ask GPT-4 this: User: The capital of Georgia is ChatGPT: The capital of Georgia is Atlanta. —it gives me the name of the southern U.S. city known for having a lot of streets named “Peachtree.” And that’s probably the “best” answer in a probabilistic sense; the state of Georgia is more populous than the European country and it’s more likely to be the right answer in the corpus.[*27] But if I ask it this— User: I just ate some delicious khachapuri. The capital of Georgia is…? ChatGPT: The capital of Georgia is Tbilisi. It’s wonderful to hear that you enjoyed khachapuri, a traditional Georgian dish! —it instead names the capital of the European country. Of course, I’ve given it a big hint (khachapuri, an almost-pizza-like type of cheese bread, is Georgia’s delicious national dish). But the whole point is that, like a poker player, ChatGPT works with incomplete information to make a probabilistic read on my intentions. There’s one last comparison between language models and poker—or really between language and poker. The critique I made in The Signal and the Noise was that, sure, AIs might wor...