Mastodawn

Charlie Stross Feb 25, 2023

The thing I find most suspicious/fishy/smelly about all the current hype around Stable Diffusion and ChatGPT is that it follows, as night follows day, about six months after the bottom dropped out of the cryptocurrency scam bubble.

This is not a coincidence.

Hucksters are chasing the sweet VC/private equity money that has been flushed out of crypto, and AI is the new hotness all of a sudden.

If you're thinking about investing now? Don't: it's too late and you'll be the target of a grift.

Show thread

Charlie Stross Feb 25, 2023

Update from comments: I'm SURE it's just a coincidence that training neural networks and mining cryptocurrencies are both applications that benefit from very large arrays of GPUs.

If *I* was a VC I'd be hiring complexity theory nerds to figure out what areas of research are promising once you have Yottaflops of numerical processing power available, then place bets on AMD, Nvidia, and (maybe) Intel going there and start seeding the field and hiring PR firms.

Show thread

dr2chase Feb 25, 2023

@cstross Not sure they do the same sort of linear algebra. E.g., looking at the voice transcription engine, it uses 8-bit FP, and ML in general uses a lot of 16-bit FP. Versus, crypto, that tends integer, and finite fields (e.g. carryless multiply). I don''t know for sure that 16-bit FP is adequate for traditional graphics, either.

Show thread

Richard Welty

@dr2chase @cstross been a while since i was working adjacent to this space, but i vaguely recall a thesis result showing that the training problem was NP complete. don't know if that result held up though.