There are a lot of specific workflows where you can't exactly verify whether something works due to the number of interactions between devices, systems, manufacturing & marketing specs, etc. So if you ask an LLM normally, it might say The ACF 2000 or whatever will do Y with the XYZ 3e, but that's not how either of those things behave together. Hidden in a lot of firm's operating data is the actual answer. So if you could, say, get all that data for a single vertical and join that into a normalized dataset, I imagine that would greatly improved accuracy for both LLMs and any model you might build for that vertical.

Title "HUMXN Research Labs Wants to Buy $50 Million Worth of Skilled Trades, Farming, and Manufacturing Data for AI & Robotics Training"

They're looking for like, plumbers' email and quickbooks which tells you what kind of plumbing jobs they do and how much they get paid, so they know what kind of things to make robots for.

I didn't say it does make it better. But it does answer the question in the title.

Here’s how I think about it: only organic data from real scenarios will provide realistic edge cases. The world is much harder to accurately simulate than people sometimes assume.

Are they repeat customers? What operations are the best to charge for? What were those prices, so on and so forth.