Welcome to Waifu Labs v2: How Do AIs Create?

Server processors gets you more bang for the buck... iff you're planning to run the hardware flat out for literally years. You save on power, but the up-front cost takes up most of that, so for a system that's mostly idle you wouldn't use them. On the other hand, any CPU with fewer PCIe lanes than a TR won't be able to run multiple GPUs optimally, and TRs are relatively cheap enough to make the reduction in PSUs/chassises worth it.

Not to mention that there are some approaches to training you can only use if you have multiple GPUs on the same motherboard, aka. sharding a single model across GPUs without communications overhead killing any benefit of that.

There's a similar situation ongoing with fiction writing, by way of NovelAI. (And some competitors, but NovelAI is head and shoulders ahead of the pack. Thankfully; they seem to be the nicest of the lot.)

I'm a fairly prolific (fan-)fiction writer, and also AI enthusiast, so of course I jumped on that bandwagon as soon as I could. What I've found is...

- AI cannot write stories on its own. It just can't, full stop. Some people try, including me, but the results are nonsensical without significant tweaking. I expect that to change eventually, but not without a conceptual breakthrough or two.

- AI is immensely useful as a prosthetic imagination.

What I use it for isn't to write the story for me. It's to, in case I ever get stuck at some point, offer me suggestions for how the story can continue -- suggestions that I can accept or deny. Even if I deny it, it's useful as a way of illuminating my own ideas for the story. There's got to be a reason I don't like that continuation, and that is often enough to think of something I do like.

In other words, it's mostly eliminated writer's block.

It's also handy for expanding my vocabulary. English is my third language, and while I like to think I'm good enough for daily life -- I've lived in Ireland for over a decade, after all -- there's a big difference between 'good enough for daily life' and 'good enough to write good fiction'. Prior to using NovelAI, my writing was... dry. Conceptually heavy SF doesn't necessarily require high-end wordcrafting, but it helps.

The AI, especially when told to emulate Sheridan Le Fanu or any of the other great authors, is better than me at this. And since I can ask it to jump in at any point, it's become the most attentive, capable cowriter I've ever had. Perhaps noticing this, NovelAI now calls their default AI tuning 'Co-Writer'.

It's still likely to write something I can't immediately use, but that just means I need to absorb its ideas and make them my own. Repeat a hundred times per day, and I end up learning much, much faster than I ever did when I was writing on my own.

To summarize, I don't use AI to write my stories for me. I use it to get better at writing.

I think it should be possible to do the same for other forms of art.

People keep forgetting that you really can only fit to data you have. Extrapolation exists as a concept but the requisite intuitive knowledge needed in order to create something new that can be successful is hard to understand even by humans at this point (how much of the business press is full of gimmicky blogposts about how to be successful, full of contradictory anecdata, opinions, advices), I don't even know how AI researchers would go about tackling that. I am not an AI person but as a plain old scientist I know extrapolation without intuition is almost always a fraught effort.

Sure it can, but then it's not considered anime anymore. I think this sentiment is confusing genre with training set constraints. A new model is not required for some artist to do something different from anime or even an anime artist to do something different. Humans can self-adjust all with approximately the same base model (whatever that is).