I use AI fairly extensively, and I self host some models, but I'm not really an expert.

My take:

• The advances of SOTA models have achieved nearly all of the functionality gains you are going to see....from here on out they are going to be diminishing set of returns for their future capabilities.
• Simultaneously, due to newer training / quant methods, you are going to see smaller models that can more or less achieve some level of parity with SOTA models.
• The big big growth is going to be in small, hyper focused domain specific models. Your doorbell will run faceNet, your coffee maker will have a tiny model in it to optimize temp, flow, etc....Almost all of these will not be "LLMs" but will be vision style models that are doing classification....but they will all link into LLM APIs to make use of generative features so your doorbell will text you a daily update like: "Suzy came in at 3pm with groceries, a UPS delivery guy came at 3:30pm, Suzy left to walk the dog at 4 and came back 15 mins later, while gone an unknown person rang door bell, here is his pic...."