Nick Perry breaks down three ways AI learns: supervised learning from labeled images, self-supervised learning from predicting text, and reinforcement learning from feedback loops. Understanding these changes how you prompt, since misaligned prompts waste hundreds in token costs.
Never ever prompt AI again until you know these 3 things. Otherwise, you could waste $300 in tokens on a shitty output. There are only 3 ways AI learns, and the third one is the most important. Number 1 is supervised learning, kind of like flashcards, because you show it millions of photos already labeled cat or dog, and it learns the features over time, like floppy ears mean it's probably a dog. Then number 2 is unsupervised learning, where you hand it a mountain of data with no labels on it at all, and it finds the pattern by itself. And ChatGPT learned to talk on a variant of that called self-supervised learning, because it scraped the internet, took every sentence it could find, and guessed the next word over and over until it got to this point. And last, but most important, is reinforcement, because this learning style is exactly how we as humans learn. Try shit, make mistakes, try more shit, see what works, double down on it, and study the winners. And that's why you should—
