LLMs are good at making predictions from deep datasets. But there are a lot of problems that LLMs are pretty bad at --- problems that involve wide, shallow datasets. A company I am working with has built a machine to look in the LLM's blind spots. They can take the wide, shallow datasets and generate insights (not predictions) that drive decisions. They can, e.g., identify genes of probable significance to a disease from existing medical literature, patient data, and research. It's all very cool, and they're fending off suitors at big valuations.

Here's my question, the one I ask myself every day since I linked up with them: what are the problems that LLMs are failing at because the data is wide instead of deep, diverse instead of homogeneous, constantly changing (and thus resists training)?

They've agreed to let me test drive their machine and scope a problem to solve. I get to play with the coolest toys!

#asknostr