Training vs Inference

Training

Part of: How AI Actually Works

A frontier model can cost over one hundred million dollars to build, and then answer your question for a fraction of a cent. That gap is not a contradiction. It is the difference between the two phases of an AI's life: training and inference . People mix these up constantly, and keeping them straight clears up a lot of confusion. What it is Every model lives in two phases. Training is building the model: a slow, brutally expensive, one-time process that produces the finished thing. Inference is using the model: running your prompt through the finished thing to get an answer. Training happens once, by a team you'll never meet. Inference happens billions of times, every time anyone hits "send." How it works Training plays a fill-in-the-blank game at impossible scale. Show the model mountains of text, hide the next token, let it guess, then nudge its internal numbers (its weights ) slightly toward the right answer. Repeat billions of times. One nudge looks like this: This phase takes weeks to months on thousands of GPUs and costs millions. When it ends, the weights freeze into a finished file, the model. Inference never touches those weights. When you send a prompt, the model runs you

Challenge: Gradient Descent Trainer