The first is that LLMs are often mis-sold as doing something they're not. They are predictive algorithms that are good at producing plausible, but possibly incorrect, answers very quickly and relatively cheaply. If that's your problem space, LLMs are actually pretty powerful

1\This is the thing I think a ton of folks in the general public don't realize. LLMs and chatbots aren't really research tools; they're generators of statistically probably (to the bot) outcomes. Sometimes that can overlap with research results, but it can also lead to hallucinations.

Key thing to remember: the ‘P’ in ChatGPT stands for ‘pre trained.’ These are backward looking models, it’s literally in the name

So for example, if you are trying to do autocomplete, or generate some video frames between two other video frames as a way of performing video compression, or to upscale an image from 1080p to 8K resolution, "generate a plausible answer quickly" basically *is* your goal, and that's how we get DLSS

Cheap with the understanding that nobody is currently paying the full cost. OpenAI has a burn rate of $700k per day by some accounts. They're going to have to get that down or pass the cost onto consumers at some point.

cheaply? the power grids would disagree..

the original application of LLMs was to machine translation, and on that front it was leaps and bounds better than anything that came before it

Doesn't help that, in my experience, the people doing the selling are, well, just terrible sales people. It's the same problem that's plagued the crypto community. It's led by techbros who think they know more (or sometimes anything) about your field of expertise and refuse to consider feedback.