Deep Blue was also a bit of a parlor trick. It relied on a ton of special-purpose hardware - literally a rack of custom-made chess ICs. It's neat that it worked, but it didn't have any wider applicability.

It wasn't a parlor trick and could have evolved into a useful product doing a small, basic subset of what LLMs do today. The problem was IBM's leadership didn't have the slightest understanding of the technology, thought they'd invented ChatGPT and pushed it into applications far beyond its potential, e.g. diagnosing cancer.

Yeah it kind of was. Watson was essentially an NLP search engine.

They should have made an MLP search engine. Those people will buy anything.