LLMs are incredible about mapping to a space of already seen things. When this space is unimaginably large, you can be fooled for a long time. But, they clearl

LLMs are incredible about mapping to a space of already seen things. When this space is unimaginably large, you can be fooled for a long time. But, they clearly struggle with generalization and rule following. This failure to generalize (extrapolate, deduce, compute) is why we still can't fire all of our DBAs. Has anyone encountered an LLM-based text-to-SQL engine that actually gets the job done? I think that's your best canary. I stopped caring somewhere around "transpose these 2 letters of the alphabet" not working consistently.