Notes

Learning is compression: you take what you've seen and squeeze it into a model you can act on.

If you want an operation done, the most energy-efficient hardware is the hardware that natively performs that operation.

Predictability and determinism are different axes. An architecture is preferred to be deterministic about where data moves, but the algorithm can be probabilistic about what the data means.

Short-term prices are hard to predict because any pattern in a market is made of people looking for patterns, so using it changes it.