What Is ChatGPT Doing and Why Does It Work?

Key Highlights

• ChatGPT is “always fundamentally trying to produce a ‘reasonable continuation’ of whatever text it’s got so far — what one might expect someone to write after seeing what people have written on billions of webpages.”
• Generating an essay is just asking “given the text so far, what should the next word be?” — over and over.
• Always picking the highest-probability next word produces “flat” essays with no creativity. Introducing randomness (temperature) yields more interesting output.
• There are ~40,000 commonly used words in English. Even 2-gram probabilities require 1.6 billion estimates; 3-grams: 60 trillion. 20-word fragments: more possibilities than particles in the universe — impossible to enumerate, must be approximated.
• Neural nets — invented in the 1940s in a form close to today’s — are “simple idealizations of how brains seem to work.”
• GPT’s 175 billion parameters are “just” enough to model next-word probabilities “well enough” for essay-length text.
• Training: present input→output examples, adjust weights via gradient descent (“backprop” using the chain rule of calculus). Repeat across many “epochs.”
• “There’s never a ‘model-less model’ — any model has particular underlying structure and knobs to tune.”
• The same neural net architecture often works for seemingly very different tasks — possibly because the tasks we give nets are “human-like,” and nets capture quite general human-like processes.
• “The success of ChatGPT implicitly reveals an important scientific fact: there’s actually a lot more structure and simplicity to meaningful human language than we ever knew.”
• Wolfram’s experiment: decomposing “I bought two pounds of apples for my children” — each phrase connects to deep conceptual underpinnings (individualism/capitalism, measurement/math, agriculture, family/parenting).