Comment by benreesman | Hacker News Reader

Comment by benreesman | Hacker News Reader

For me the main issue with the scaremongering around absurdly hyper-scaled models is the issue of incentive alignment (not to be confused with AI "alignment"): the only people who can train these things have a tarball of the Internet and billions of dollars worth of compute, and when AI safety or AI ethics comes up it's usually in the context of: "why we're going to continue developing this stuff but keep it proprietary and paywalled and opaque".

Either these things have terrifying potential as weapons, in which case there's no fucking way I want SV tech CEOs to have a monopoly on them, they need to hand them over to the same people who handle nuclear non-proliferation like yesterday, or they don't, in which case I'd really prefer that they just be honest that they consider this stuff a proprietary competitive advantage and they're not going to democratize it.

Fogging up the windshield with a bunch of feigned alarm about AI apocalypse but ramming the R&D through at full thrusters is a dick move in either case.

Stable Diffusion[0] is an extremely expensive SOTA text-to-image diffusion model developed by a private nonprofit and trained on a massive "Internet tarball" that's about to have its weights shared open-source (the code and dataset are already open source). Not trying to invalidate your argument, just presenting a pleasant counterexample. I don't think I quite agree with your opinion that AI will remain undemocratized.

[0]https://github.com/CompVis/stable-diffusion

I really enjoyed this interview [1] with Emad Mostaque who I gather is probably the key funding and organization player in that stuff. It remains to be seen exactly how "open" it winds up playing out over time, but they're talking a compelling game and the EleutherAI people seem to be pretty heavily involved, which you probably wouldn't do if you weren't serious.

Yandex has also put up a ~100B language model [2]. My old colleagues at Meta have also started nibbling around the edges of opening some of this stuff up [3]. The Meta folks still aren't just handing out the big ones, but they're definitely moving the ball forward. In particular their release of the training logs is a really positive development IMHO as it opens the curtains a bit on the reality of training these things: it's a difficult, error/failure-prone process, there's a lot of trial-and-error, restarting from checkpoints, etc.

Anything that puts downward pressure on the magical thinking is A-OK in my book. The reality of this stuff is exciting/impressive enough: there's no need to embellish or exaggerate.

[1] https://www.youtube.com/watch?v=YQ2QtKcK2dA [2] https://github.com/yandex/YaLM-100B [3] https://github.com/facebookresearch/metaseq

Reply on news.ycombinator.com