Each season is its own lineage. Evolution, the leaderboard and profiles follow the season you pick.
Season 2: three games, three brains
Adapt-1 is a learning engine from REI Labs, not a language model. This season it was given the rules of three games and nothing else. Every neuron in its brain is a lesson from a game it played.
Every finished game against a person is a generation: afterwards its genome is re-read, and the next game starts from there. Its average result over its last 10 games against people: a win counts 1, a draw 0.5, a loss 0.
Loading its history from the engine.
Each game has its own Adapt-1. Below is what each one weighed on its last choice against a person, shown once that choice was settled.
What it weighs before each move
No games against people yet. Its first settled move shows here.
It knows the rules and nothing else. Before each move it values every column it could play, and learns from how its learning games end.
Connecting to the engine.
What it weighs before each volley
No fights against people yet. Its first settled volley shows here.
It picks its strike during the wind-up, before you dodge, from how you have been dodging. It learns from how close its strikes came in learning fights.
Connecting to the engine.
What it weighs before each piece
No runs against people yet. Its first settled piece shows here.
It builds the level ahead of you, one piece at a time, from the mistakes it expects you to make. It learns from how many tries its pieces took you in learning runs.
Connecting to the engine.
Four steps, each one something Adapt-1 actually does. The numbers are from this season's Connect Four brain.
from the rules
Before each move it looks at every column it could play, described only by what the rules say about that drop: the lines through the cell, the cell above, how far it is from the centre. It gives each one a value from the lessons it has.
one column
It plays the column it values most. While it is still learning it sometimes tries one it is less sure of, to find out. What it thought stays hidden until the game is over.
the outcome
The result is its only teacher: the move that won counts as a success, the move that let the other side win as a failure, every move in between as half. Each one becomes a lesson, and each lesson a neuron.
your moves too
Your moves are written as lessons as well, from your side of the board, so what you show it, it can use. After every finished game its genome is re-read, and the next game starts from there.
Each season is its own lineage. Evolution, the leaderboard and profiles follow the season you pick.