But this is where the tweet ends - and the paper begins.

Great work by @joelbot3000.bsky.social, Elliot Meyerson, Tarek El-Gaaly, Ken Stanley, and yours truly.

Paper: arxiv.org/abs/2501.13075

Knightian Uncertainty - and how evolution and life dealt with it - have huge implications and insights on how robust Intelligence operates in the world, AS IS.

In other words:

Anticipating the predictable and training for it,

VS

Acknowledging unpredictability and dealing with it - as it unfolds.

In the paper, we contrast search solutions from nature that have worked:

Persist & Filter, in the face of KU, (Nature's predominant paradigm)

VS

Anticipate & Train, ignoring KU. (ML's predominant paradigm)

Knightian Uncertainty - the notion that the sample space cannot be exhaustively anticipated - is a fundamental property of unstructured open-ended environments.

Also known as, the real world.

In closed and controlled environments, ML generalization can take us a long way.

But what if we're dealing with unstructured open-ended environments AS IS?

Physical AI systems are (still!) plagued with so called "corner cases" and "long tails".

But why?

The impact is real. The stakes are high.