Definition
Plain language
An older idea where an AI learns from its own early attempts within a single long run, getting better as it goes.
As stated in the literature
A meta-reinforcement-learning approach (RL²) that stitches multiple episodes into one trajectory so an agent carries learning across them via its recurrent hidden state; contrasted with note-carrying agents that pass a human-readable text summary instead of an opaque vector.
Also called: RL², RL squared
Why it matters: It carries learning across attempts inside one run, but stores it in an opaque internal vector rather than a summary a human can read.
For example, an agent dropped into a new maze gets faster within a single long run by remembering what its earlier attempts revealed.
Heard on the show
“… This learn-to-learn idea isn't new — there's a line of work from around twenty-sixteen, the RL-squared stuff, where you'd stitch several episodes into one long trajectory so the agent learns from its …”Episode 160 — Training an AI to Take Its Own Notes, So Its Future Self Works Better