Imagine an instrument with a billion tuning pegs. You cannot tune it by ear, peg by peg — nobody has that many lifetimes. So instead you play a note, measure how far off it was, and nudge every single peg a fraction of a turn in the direction that would have made it slightly less wrong. Then you do it again. And again, a few trillion times. That is training. The note being played is a guess at the next word in a piece of text; the measurement of wrongness is what engineers call the loss; the nudging of every peg at once is gradient descent.
The word training does a certain amount of damage, because it imports a teacher and a student and there is neither. Nobody chooses the lessons or explains the material. The model absorbs the statistical regularities of whatever it was shown — the reasoning and the good prose, but also the errors, the omissions, and the assumptions nobody in the source text thought to state out loud. It can memorise when it was meant to generalise, which is the failure called overfitting. And training has a hard edge: it ends. The model is frozen on a particular date and learns nothing from the person using it. Correcting a chatbot does not teach it. The correction lasts as long as the conversation and not one second longer.
This makes training an unusually unforgiving kind of engineering. It is front-loaded, enormous, and effectively irreversible: months of computation across tens of thousands of specialised chips, at a cost that now runs into the hundreds of millions. Nearly every decision — what data, how large, what objective — is made before the run begins, and you find out whether you were right at the end. The later stages are cheaper and narrower. Fine-tuning steers an already-trained model toward a particular use; reinforcement learning from human feedback shapes how it behaves rather than what it knows. But the bulk of what a model is was settled during that first long run, and it cannot be renegotiated afterwards.
The human translation is everything you knew before you walked into the room. Your education took years, cost a great deal, drew on material you did not choose, and cannot be re-run — and none of it is what anyone actually sees. What they see is what you do in the room, which is a different thing entirely, and the one the industry calls inference. Training is where the records are set and the headlines are written. It is not where most of the work happens.
