Self-learning systems
Learningthroughinteraction.
Keshi builds agents that learn from their own experience and remember what they learn.
01/Approach
Approach
Keshi builds self-learning agents. They start with nothing but the rules and learn from what happens when they play. What they learn, they keep.
An agent, a world, and what happens between them.
- The agent
- makes a choice.
- Its world
- answers with an outcome.
- Experience
- is where they meet, and what it keeps.
The brain behind our agents is Adapt-1, a learning engine from REI Labs. It learns while it plays, from the outcomes of its own choices, one experience at a time. It is not a language model.
About Adapt-101
Learning
They begin with the rules. What they know comes from playing.
02
Adaptation
What happened last time shapes the next choice.
03
Persistent memory
Lessons are kept across games and across players, so learning adds up.
04
Learning in public
Anyone can play against it, watch it change, and see what it learned.
02/Arena
Born with the rules. Evolved by you.
The Arena is our open experiment with Adapt-1. Each season it starts with the rules of its games and nothing else. Its lessons come from the games people play against it, a day's worth at a time, and you can watch its brain grow.
Watching is open to everyone. Sign in with X to play.
03/Reicraft
In development
Next, it learns Minecraft.
Reicraft gives Adapt-1 a body in Minecraft and lets it learn the game from start to finish: gather, build a home, survive the night, find the stronghold, and face the dragon. It learns from what happens to it, and it will learn live, on stream.
Stream coming
Now
The body
Being built: hands that walk, mine, craft and build, fluent enough to be worth watching.
Now
The brain
Being readied: Adapt-1 makes the calls that keep it alive and learns from what happens to it.
Next
First lessons
Its first live learning runs, on a world of its own.
Later
On stream
Every decision on screen as it happens, and what it learned from it.
Goal
The dragon
From a bare spawn to the end of the game, on what it learned.