[thenaturalist]

This is such a short sighted take glaringly ommitting a crucial ingredient in learning or improvement - both for humans or machines alike: feedback loops.

And you can't really hack / outsmart feedback loops.

Just because something is conceptually possible, interaction with the real rest of the world separates a possible from an optimal solution.

The low hanging fruits/ obvious incremental improvements might be quickly implemented by LLMs based on established patterns in their training data.

That doesn't get you from 0 to 1 dollar, though and that's what it's all about.

[bwfan123]

this. Was highlighted by Sutton in a recent podcast rather starkly.

LLMs are a great tool. But, the real world is far too nuanced to be captured in text and tokens. So, LLMs will be a great productivity boosting tool like a calculator or a spreadsheet. Expecting it to do more is science fiction.