[delichon]

On the contrary. The heart of an LLM is a next word predictor, based on statistics. They do much the same with concepts, making them essentially consensus distillation devices. They are zeitgeisters. They get weird mainly when their training data is too sparse to find actual consensus, so instead tell you to stick cheese to your pizza with glue.

[astrange]

> They get weird mainly when their training data is too sparse to find actual consensus, so instead tell you to stick cheese to your pizza with glue.

That's exactly not how that happened. That happened because Google's summaries are based on their search results and one of the search results contained that.

[ericjmorey]

This is only useful if you know what data was used to train the model.