also, there are plenty of text to speech tools that just record and don't analyze and they've existed for nearly my entire life. we've got to stop allowing people to conflate direct, non-generative tools with genAI when they have a financial incentive to do so

I went back and looked at the SoCar presentations at this conference and feel better now.

Sorry for the freak out.

The point is, I am all for AI, and I can see LLMs/Generative AI having some value, but I think higher ed, and my field are getting high on vaporware.

I'm really concerned about how we've just outsourced this stuff to the cloud. Most of us have devices that can run small-medium models on device, but phone assistants are still trash, and still require connectivity in many cases.

Last year I got a preview of someone doing cutting edge work with LLMS proposing transformer modification to better fit knowledge forms, that seems like good work. I don't know that building large data sets based on bullshit is all that good.

These models are in the 80-95% accuracy range, because of advances on pre-release training practices, but you're misusing the technology when you apply it to knowledge work. Because knowledge is human and LLM is machine bullshit.

And hardly any of it is cutting edge, just a lot of assessment, which is both a fools errand (because of the speed of model evolution) and not useful (because the models are literal bullshit, and the researchers aren't dealing with model revisions, which would actually be helping things).