02 of 12
Learning
Imagine you are playing basketball. You take a shot and miss. You can feel how far off you were, so you adjust the next one. After enough of those adjustments, the shot starts to feel like a habit. Learning is that same loop, run on examples until the machine gets less wrong.
The 2012 vision contest was a public proof: a deep stack, trained this way on ImageNet, won by a wide margin. After 2017 a lot of practice stopped needing a person to name every item. A system could train on a scrape of the public web by predicting missing words in sentences, and then adjusting when it was wrong.
How do you know it learned, not memorized?
You can check whether a system learned, rather than memorized, by testing it on questions it never practiced on. Researchers later found grade-school math problems from a supposed hidden test sitting in the same web dumps the models had trained on. A high score can just mean it had already seen the answers, which is why METR keeps private software tasks instead of trusting a public leaderboard.
Why did this only take off recently?
The method of learning from a mistake by sending it back through the layers is from the 1980s. AlexNet won in 2012 because graphics chips could run that process on ImageNet. In 2017, transformers made the same process cheap enough to run on text.
What is the second practice pass for?
The web scrape teaches what usually comes next. In 2022 OpenAI practiced on instructions and on answers people ranked. ChatGPT came out of that second pass, which is why it follows a request and refuses some of them.