No “Zero-Shot” Without Exponential Data: Pretraining Concept Frequency Determines Multimodal Model Performance

I’m not going to pretend that I can understand the minutiae as someone who basically just has an undergrad degree in IT, but it seems to be implying that exponential data is required for further advancements in generative AI.

A cool paper and all, but I can’t stop thinking about how this shit won’t matter at all to C suites and MBAs who just want to AUTOMATE AUTOMATE AUTOMATE everything under the sun. Reminds me of how the experts in whatever specialized field research a problem and business people just throw away the results and make up their own (i.e. marketing).

The conversation should have always been “Yeah your job will eventually become automated, but it’s not for the reason you think.”

Will be very interesting to see how the next few years play out