First question: I'm not 100% certain what you're asking here. They are limited for useful purposes because there are only so many uses for generating images or text. But there are multiple LLMs trained on different data sets with different potential uses.
One non-llm example that circulates a lot on tumblr is that AI image recognition tool that was trained to tell bear claws apart from croissants, but that researchers were able to adapt to recognize different kinds of cancer cells. That's image recognition, not an LLM, but it demonstrates that yeah different training sets can be used for different purposes with specific training, and general training sets can be refined for specific purposes.
Second question: A lot of the information about the energy required for LLMs is tremendously misleading - I've seen a post circulating on tumblr claiming that every ChatGPT prompt was like pouring a bottle of fresh water into the ground just to cool the servers and that is a serious misstatement of how these things work.
Energy requirements for LLMs are frontloaded in training the LLM; they require a lot of energy to train but comparatively little energy to use.
This chart breaks down some of the energy requirements for training an LLM:
50-800-ish Megawatt Hours for training. Let's take the high middle and consider 800MWh for our question.
800MWh is 8 Million KWh. The average US household uses 10500 KWhr per day. You divide those two and you get 762, so if we're rounding up you could train the custom model for the same amount of energy that it would take to run 800 US homes for a year. That is certainly not a *small* amount of energy, but it is also not an *enormous and catastrophic* amount of energy.
GPT-4 took about 50 Gigawatt Hours to train, which is a big jump up from the models on this chart. Doing the same math as above, it works out to powering about five thousand homes for a year. That is, again, not a *small* amount of power (apparently about twice as much as the Dallas Cowboys' Stadium uses in a year).
But training isn't the only cost of these models; sending queries uses power too, and using a GPT-backed search uses up to 68 times more energy than a google search does. At 0.0029 KWh per query.
Now. This is *not trivial.* I don't want to dismiss this as something that does not matter. The climate cost of the AI revolution is not something to write off. If we were to multiply the carbon footprint of Google by 70, that would be a bad thing. My attitude mirrors that of Wim Vanderbauwhede, the source of the last two links, who I am quoting here:
For many tasks an LLM or other large-scale model is at best total overkill, and at worst unsuitable, and a conventional Machine Learning or Information Retrieval technique will be orders of magnitude more energy efficient and cost effective to run. Especially in the context of chat-based search, the energy consumption could be reduced significantly through generalised forms of caching, replacement of the LLM with a rule-based engines for much-posed queries, or of course simply defaulting to non-AI search.
*however,* that said, the energy cost of LLMs as a whole is not as ridiculously, overwhelmingly high as I've seen some people claim that it is. It *does* consume a lot of power, we should reduce the amount of power it uses, and we can do that by using it less and limiting its use to appropriate contexts (which is not what is currently happening, largely because the AI hype machine is out promoting the use of AI in tons of inappropriate contexts).
Ditto for water use; datacenters do use water. It is a significant amount of water and they are not terribly transparent about their usage. But the amount required for "AI" isn't going to be magically higher because it's AI, it's going to be related to power use and the efficiency of the systems; 50GWh for training GPT-4 will require just as much cooling as 50GWh for streaming Netflix (which uses over 400k GWh annually). The way to improve that is to make these systems (all of them) more efficient in their power use, which is already a goal for datacenters.
And while LLMs do have an impact on the GPU market my experience is that it has been minimal compared to bullshit like bitcoin mining, which actually does use absurd amounts of power.
(Quick comparison here, it's frequently said that ChatGPT uses over half a million kilowatt hours of power a day; that seems like a big number but it's half a gigawatt hour. Multiply that by 365 and you've got 182.5 GWh. It's estimated that Bitcoin uses in the neighborhood of 110 *Terawatt* hours annually, which is 11000 GWh, which is about 600 times more energy than ChatGPT with *significantly* fewer people using Bitcoin and drastically less utility worldwide.)
A lot of computer manufacturers are actually currently developing specific ML processors (these are being offered in things like the Microsoft copilot PCs and in the Intel sapphire processors) so reliance on GPUs for AI is already receding (these processors should theoretically also be more efficient for AI than GPUs are, reducing energy use).
So i guess my roundabout point is that we're in the middle of a hype bubble that I think is going to pop soon but also hasn't been quite as drastic as a lot of people are claiming; it doesn't use unspeakable amounts of power or water compared to streaming video or running a football stadium and it is in the process of becoming more energy efficient (something the developers want! high energy costs are not an inextricable function of LLMs the way that they are with proof-of-work cryptocurrencies).
I don't know about GPUs, but I know GPU manufacturers are attempting to increase production to meet LLM demand, so I suspect their impact on the availability of GPUs is also going to drop off in time.
Author's note here: I am dyslexic and dyscalculic and fucking about with kilowatt hours, megawatt hours, gigawatt hours, and terawatt hours has likely meant I've missed zero somewhere along the line. I tried to check my work as best I could but there's likely a mistake some where in there that is pretty significant; if you find that mistake with my power conversion math PLEASE CORRECT ME because it is not my intent here to be misleading I just literally need to count the zeros with the tip of a pencil and hope that I put the correct number of zeroes into the converter tools.