Games and reinforcement learning
Learning from reward rather than from labels: the Go move no human would have played, how many games it takes to learn Atari, and the search behind AlphaZero.
Articles

How a language model is turned into an assistant
A model that has read the internet continues text; it does not answer questions. Three stages turn it into an assistant: demonstrations, comparisons, and a reward with a leash. The InstructGPT numbers of 2022, and the Bradley-Terry probability worked for two answers.

How many games does it take to learn Atari?
In 2015 one network learned 49 Atari games from the pixels on the screen, with the same settings for every game. It needed 50 million frames per game, about 38 days of continuous play. A person needs an evening. The gap is the central problem of reinforcement learning.

Move 37, the Go move no human would have played
On 10 March 2016, in the second game against Lee Sedol, AlphaGo placed a stone on the fifth line that its own policy network rated as a 1 in 10,000 choice for a human player. Commentators thought it was a mistake. It won the game.