Re: Sigh....
No I learnt fuck all. You are talking specifically in terms of LLMs and image diffusers, which are a small area of AI, but not the entirety of AI.
The vast majority of the energy consumption there is in training the models, not prompting them so the energy use doesn't scale as linearly as you'd like, especially when you take into account quantisation of models and ever more efficient algorithms used for prompting...if anything, over time the energy consumption of AI will plateau at worst...especially as models become more specialised...anyway that is beside the point, I don't disagree with you necessarily where LLMs and diffusers are concerned, I disagree with you referring to a very specific area of a very broad category of tech and tarring the whole industry with the same broad strokes.
I build AI solutions for a living, and most of the time those solutions run on low power equipment...like very low power equipment. A computer vision AI can run on a single core ARM processor that sucks less than 3W and can have a latency faster than you can blink...OpenCV for example can run on 4GB of conventional RAM with a single core and it is still pretty damned quick.
AI is all around you and has been for quite a long time now...at least 15 years. Image recognition, spam filtering, number plate reading, weather forecasts, online ads...all use AI to some degree.
I've built AI solutions for all kinds of things ranging from detecting nutritional deficiency, to matching body shapes to clothing, to analysing MRI data, to predicting financial markets...all kinds of stuff...and it's very rare to require massive amounts of energy sapping kit for it to work and for it to work quickly...collecting the data probably uses more energy than the resulting solution.
I don't disagree that LLM usage will increase energy use over time including diffusion models, in fact I agree with you to a certain degree (not entirely though)...but lumping in the rest of the AI industry is just stupid and shows a fundamental lack of understanding of how AI is used, developed and implemented and how it progressing over time and how different situations call for different AI implementations. Large Language Models and Image Diffusers are not used in oil exploration. Never have been, probably never will be because a large language model is the wrong kind of AI for that sort of task...same applies to image diffusers. The kind of AI they use for oil exploration, or any kind of predictive modelling is nothing like a large language model / diffuser...not even close...you typically train those on the fly, because you're using dealing with very specific data, not a nebulous amount of loosely connected data. E.g. something like this local weather dataset (https://github.com/AbdullahZahid77/Weather-Prediction-using-Machine-Learning-Python-)...that is 60 years of detailed data and it fits in 2.6MB...it doesn't take a whole heck of a lot of power or time to crunch that...even on old shitty hardware...and once it has been crunched (which on a dataset that size would take maybe a minute) running predictions against it will be near instant...even on a shitty low end CPU.
Where LLMs / diffusers are concerned, they will become more energy efficient with each new generation of hardware because even though the hardware might use a little more power with each generation, the leaps in performance and optimisation massively increase the efficiency of training. This generation might take a week on 20 GPUs using 650W to train a specific type of model. Next generation might only need 4 GPUs at 700W for a day. That's orders of magnitude better, so even though the power usage creeps up, what you get in return for that power is orders of magnitude better. So if you look at it from an "hallucinations per watt" point of view, then power consumption will in fact be decreasing, if not metric other than the sheer wattage matters then yeah power will creep up, in the short term...then on the inference side, factor in the efficiencies that are being found there...there are models that have been quantised to half their size that provide near as damn it exactly the same results and therefore require far less grunt and power consumption to operate....short term, in this area will energy use continue to increase...yes, but not for the reasons you think...is it increasing proportionally to the expected results? No, the results are getting better all the time, at a rate faster than the power consumption is increasing...there will be a point of diminishing returns eventually with these types of AI because there is only so much data you can train them on, and beyond that no amount of power will improve the results...therefore we will likely hit a plateau followed by a gradual decrease in energy consumption as things become more optimised...there are no secret herbs and spices involved with training LLMs, the difference between them all is simply the data they are trained on and the amount of money spent acquiring it, there are probably minor differences in the way the data is vectorised, tokenised, filtered, "woked" etc but those aspects aren't really that critical and don't really have an impact on energy use, they're more like fighting over a duvet with your wife, you tweak something to change and improve one aspect of the result, but inevitably you're going to make another aspect worse, that's just how it goes with these things, training models is very unforgiving when you tinker with data and the more you tinker the harder it is to objectively establish whether the results are accurate...the only reason you need more power for training LLMs is to get to market quicker and beat the competition. If you have to organisations with identical training data, one has 1000 GPUs and the other has 2000 GPUs...the one with 2000 GPUs gets to market twice as fast, but both organisations will produce pretty much identical models...the reality is the datasets vary wildly...and at the moment ChatGPT has all the money to be able to buy access to vast swathes of data and stacks and stacks of GPUs...eventually the playing field will level because data will become cheaper and cheaper to access, because the people selling it will want to continue to sell it, but the customers able to buy it will have ever shrinking budgets.
Your car example is sort of right in a way...but the context is wrong...you're essentially doing maths based on cars to argue against commercial aircraft....blaming drivers (people using models) for the fuel burn of an airline (ChatGPT, Microsoft etc)...the power consumed by people using large language models pales in comparison to the energy used to train large language models...I don't blame you for your piss poor example, because it's very easy to quantify millions of hypothetical john smiths using LLMs, because that's data that is easy to understand and extrapolate...but it is very hard to quantify the energy use of OpenAI, Microsoft etc etc because nobody has a clue, there is no data to extrapolate from and therefore that side of the equation is very abstract...if you throw an apple at a target and the lady falls in the water...she must be a witch right?