AI, DREAMS, AND THE FUTURE OF LEARNING

Mammals dream because reality is slow.

A rat that ran a maze yesterday may, that night, run the same maze again in its sleep – faster, in fragments, sometimes backwards. Neuroscientists first recorded hippocampal replay in rats in the 1990s. Later work showed just how strange it can get.

The body is still. The learning isn't.

Experience is something we live through, then replay, compress and recombine.

Which is what some of the most capable AI systems do.

AlphaGo Zero learned Go by playing itself. DeepMind's Dreamer series learns by imagining futures inside its own model of the world. MuZero plans by rolling possibilities forward before anything happens. Reasoning models disappear into private scratchwork before returning the neat answer we get to see.

And Anthropic recently introduced a feature it calls ‘dreaming’: between sessions, an agent reviews past work, updates memory and carries lessons forward.

These aren't all the same technique, but they do rhyme.

If experiential learning is the route to more capable AI, imagined experience is how you outrun real time.

But, if the machine learns from dreams, who checks the dreams?

One to sleep on.

Happy Monday.