>>108819616
Literally every non-frontier company. Any idea someone had that relies on AI will just be stolen and added to the big players’ offerings. The best thing for these companies would be to try and get the bag and get bought out before the frontier companies spring the trap and massively increase prices to turn a profit on all these locked-in frontier wrappers.

>>108819639
Frontier companies themselves also are pets.com-tier. They all promise AGI and receive billions in investments plus extra subsidies on top. The problem is this requires literal trillion dollar market caps for some of these players to hit their investors' targets. Literally there is not enough money ON EARTH for even ONE of them to win right now. And how bad will it be when subsidies run out or when they start wanting to turn a profit? Already people point out how the latest crop of models was heavily degraded in both output quality and service quality (lower availability, slower response, worse instruction following, worse answer validity rate, higher costs). As an (((AI))) """engineer""", I always make it a point to build internal datasets to catch regressions and it's been pretty horrible lately. Literal 20% accuracy drop from gpt-4o to gpt-5.0, ~3% drop between 5.0 and 5.5, and we have collabs with openai and google to fix this but they themselves based on their own internal developments and test confirmed the same pattern, and told us their other clients told them they had the same problems we report. Let's not forget the high school-tier """deployment""" story at google following openai's lead lately (roll out 3.0 in preview, supposedly ga soon. Suddenly they deprecate 2.5 and also cancel 3.0, 3.1 preview comes out and will only be ga who-knows-when. We (clients) had to threaten them into backpedaling on 2.5 deprecation. Still no sign of ga for 3.1).
Fact is, it's all about to crumble. I give it 5 years max, but realistically likely to crash within 3.

>>108819616
every one but the hardware manufacturers

>>108819639
it's literally the opposite, once llms plateau (already did probably), every frontier corp will get fucked first. Trillions of investments with fraction of returns. Everyone else will just move to local models

>>108820015
but unlike the dotcom infrastructure investments cloud gpus won't ever become cheaper and affordable.. right?

>>108820015
Which is probably why I'm seeing a lot of FUD about how local models are unsafe/unsecure because they don't have the safety guardrails that frontier models have (hearty kek) lately.

>>108820037
A single workstation card is sufficient for an individual's use, a rack of them for a small companies. Which is expensive, sure, but not impossible to afford.

>>108820037
which would make local models cheaper too lol. There are also security concerns of leaking your trade secrets and local models can be better adjusted to your business needs.

Once llms mature i think most enterprise users will want in house solution, I'm sure some ai services will exist too. But the way it is right now is just unsustainable

>>108820052
>>108820037
Cloud gpu prices right now are so high that even if you buy brand new gpus, they'll pay for themselves with 3 months of use, plus have many advantages like not having to wait a billion years to get a popular gpu provisioned for your use. The cost of gpus also gets accounted for as depreciation which is tax-advantaged compared to paying cloud services.

>>108819616
dayum I could go for some pig ears like right now.

>>108820052
nigga that’s like saying a 1080ti is enough for an average gamer. there’s no end for needs of improvements.

>>108820015
Local models are free, uncensored, unlimited and good enough for most purposes. Plus they're improving all the time and it's inevitable that hardware keeps improving. I don't see how corporations can compete.

>>108820158
>a 1080ti is enough for an average gamer
Correct

>>108820158
It's true though. Look at what people play. A 1080ti is basically the same performance as a 5060ti by the way.

>>108820158
well in the future models and local engine might have evolve to that point, hopefully

Wikipedia, they had a meltdown over AI even though every AI uses them as base knowledge. They treat AI users as worse than vandals now.

>>108820703
And they should, since all AI can do is deface.
Then again, so do the leftards who've been infecting every wikipedia article and rewriting reality unabetted for a long time in most cases.

>>108819616
OpenAI

>>108820164
>Plus they're improving all the time and it's inevitable that hardware keeps improving

Only to a point. Have you seen the size of Pixar's render farm? That's just for them making their own movies. Goys think they're going to be making 4 hour epics free of charge and have them pumped out by the end of the day.

>>108819616
Meta AI

>>108820044
>safety guardrails
Oh no, the model might call me a nigger