I have an opinion about AI, which is that the LLM scam has yet to emerge. I know a lot of people on the fediverse are just automatically against LLMs and with what seems like a good set of reasons, but really they’re lacking in detail and just going along with a tide of opinion.

For most fediverse ‘LLM haters’ (or AI haters, but not all AI is an LLM) they’ll point to outrageous use of water for cooling, outrageous use of electricity, huge amounts of processing needed, insane amounts of GPU requirements, and so on. This is not even to mention the dodgy position on original effort and copyright and ownership of matter used in training. All seems bad. And it would be if this were to be believed.

But for a moment, let’s just not believe it.

The processing requirements - yes, you need high power GPUs for training. Why are we continually training LLMs? Most of the initial training was done long ago. Much of the training now is done by ‘inheriting’ the training from the initial models and refining it a lot, adding amounts of further training on top. Are we still training models as much as we used to be? More? Not as much? As much but more efficiently? Do we have any actual reporting on this as we stand today?

The amount of GPU needed (and consequently electricity) seems to me the thin end of a wedge-shaped iceberg of rabbit mines. Yes you ‘need’ a GPU to plough through a lot of numbers. Quite why we still need a ‘graphics processing unit’ to perform vector calculation is a quaint mystery of terminology – an early movement toward general purpose computing using the parallelism of GPUs used for shader calculation shifted the GPU away from purely graphics processing but didn’t really bring forth a recognised parallel computer platform (in the way we nearly had with the transputer in the 80s).

Today the tensor calculations used in AI and LLMs are pumped through GPUs. Perhaps we’re waiting for a massively parallel computer platform using optical computing. In the meantime, scaled-up parallel vector and tensor processors will have to do. Why though? Is that where the cartel money is, in keeping the GPU trade exactly as it is, only more of it? Surely something can come in from the side unexpectedly to eat this opportunity?

More than this though, I strongly guess or suspect that the amount of investment being sucked into the AI industry is completely describable as organised crime. That much money is extorted on the basis that huge data centres are required and that they’ll need to be filled with an arrangement of running GPUs and this is where the investment money will end up. I reckon that is mostly bullshit.

The data-centre argument is because the current ‘LLM as a Service’ model needs to attend to many millions of users all at once, and their prompt processing needs to be handled with minimum delay. We get that, yes. So it needs to be better than a standard web server with a complex web site. Yes. But wait, does it really? Really?

I don’t think the future is in AI as a Service so much as we’re being convinced it is, I think the future is more in local powerful models, not one big centralised provision. This is perfectly understandable – except if you’re behind running a big centralised provision, in which case your propaganda will laugh at local models and of course promote the single big ‘as a service’ provision.

I also doubt that a lot of the training is actually what they say it is. A lot of new models are trained on the results of old models. Not copied, but using an existing model, or more than one, to train ‘from’. Most of the energy use of LLMs I suspect has already happened, and there isn’t the need for an awful lot more. Except if you’re already structured for investment money funnelling in because training.

I also doubt that a lot of the results given in a ‘as a service’ model live over the centralised server provision is artificial – I am guessing that some of it is human-faked. And that’s where a lot of the investment money goes to – into paying humans to supplant deficiencies in the LLM scheme when run live for the entire population of the world that wants to use an online LLM service.

A local LLM can certainly do all that is required, but when the whole world wants to use a big centralised one, I’m guessing that having human assistance behind the scenes becomes necessary to ensure it all flows in time. Do we have any reporting on this front, I wonder?

I think the demands of LLMs – the training time and energy, the electricity to run it, the RAM, the GPU, etc, are real but the amount is inflated and when applied to centralised ‘as a service’ business structured projects, has been allowed to be outright deceptive at the least, and at most, almost mafia-like in terms of extortion, like a protection racket, designed to keep the already outmoded ‘centralised LLM as a Service’ con going.

Anyway, this is purely my guess. I’ve not looked into this at all.
#AI #LLM