Technology

Other than with language models, this has already happened: Take a look at apps such as Merlin Bird ID (identifies birds fairly well by sound and somewhat okay visually), WhoBird (identifies birds by sound, ) Seek (visually identifies plants, fungi, insects, and animals). All of them work offline. IMO these are much better uses of ML than spammer-friendly text generation.

This has already started to happen. The new llama3.2 model is only 3.7GB and it WAAAAY faster than anything else. It can thow a wall of text at you in just a couple of seconds. You're still not running it on $20 hardware, but you no longer need a 3090 to have something useful.

Well, you see, that's the really hard part of LLMs. Getting good results is a direct function of the size of the model. The bigger the model, the more effective it can be at its task. However, there's something called compute efficient frontier (technical but neatly explained video about it). Basically you can't make a model more effective at their computations beyond said linear boundary for any given size. The only way to make a model better, is to make it larger (what most mega corps have been doing) or radically change the algorithms and method underlying the model. But the latter has been proving to be extraordinarily hard. Mostly because to understand what is going on inside the model you need to think in rather abstract and esoteric mathematical principles that bend your mind backwards. You can compress an already trained model to run on smaller hardware. But to train them, you still need the humongously large datasets and power hungry processing. This is compounded by the fact that larger and larger models are ever more expensive while providing rapidly diminishing returns. Oh, and we are quickly running out of quality usable data, so shoveling more data after a certain point starts to actually provide worse results unless you dedicate thousands of hours of human labor producing, collecting and cleaning the new data. That's all even before you have to address data poisoning, where previously LLM generated data is fed back to train a model but it is very hard to prevent it from devolving into incoherence after a couple of generations.

That would be innovation, which I'm convinced no company can do anymore.

It feels like I learn that one of our modern innovations was already thought up and written down into a book in the 1950s, and just wasn't possible at that time due to some limitation in memory, precision, or some other metric. All we did was do 5 decades of marginal improvement to get to it, while not innovating much at all.