Prediction is underrated: if you can predict, you can optimise

Headshot of Adam Jones

Adam Jones

A businessman in a suit and red tie, hands cupped around a brightly glowing crystal ball on a velvet-draped table. Photo: InfoWire.dk, Flickr, public domain

People underrate the power of prediction. I think this is because they don't realise prediction is all you need...

Once you have accurate and cheap prediction, you can turn it into an algorithm for action:

  1. Enumerate your possible actions.
  2. Use your prediction function to evaluate each one.
  3. Execute the best one.
In practice, the action space might be huge, and prediction usually has some cost,1 so enumerating everything might not be feasible. But with heuristics to prune the space, the algorithm works. You can also distill it by running it explicitly the first few times you do the thing.
  • Empathy is underrated: Getting good at empathy lets you predict the behaviour of others, including future versions of yourself. Combined with this algorithm, you can more effectively optimise for your goals.
  • Test-time compute and distillation helps you run this algorithm implicitly without as much cognitive effort.
  • AIXI is effectively the above algorithm formalised - and presented as an algorithm for AGI.

Footnotes

  1. I mean cost in any resource, not necessarily money.

    E.g. even on a small decision, like where to go out for dinner, it takes a bit of time and brainpower to explicitly predict how much you might enjoy different options. ↩