OpenAI announces investment project worth 500 billion
linky (known as griftgate lmao)
At current prices of money, this project has to return 20 billion in profit, and likely operate on 50 billion revenue at least to cover the energy costs
I'm beginning to think that "flipping" the geopolitical/economic board via rampant acceleration of climate change is the US's only real plan for re-asserting hegemony.
Fossil fuel industry expansion, refusal to adhere to Paris Agreement Accords, and now data-centers being plastered across the world burning electricity to enshittify the internet coupled with China going hard on developing renewables and harnessing sustainable energy to me signals that both nations know that there is no cooperation with the West on reduction of emissions, etc.
There is no plan, there is just profits and stock evaluations (in this case, nvidia gimme), and grift. 60 year old dipshits are impressed by chat gpt talking cause that's what they do for work, they immediately think they have solved labor. kissinger invested in theranos as a reminder. These are not smart people, these are incredibly rich people who can warp society with pocket change
60 year old dipshits are impressed by chat gpt talking cause that's what they do for work, they immediately think they have solved labor
I'm sorry, I know this is a popular line of thinking around here, but this is off base.
What's actually happening is that humanity may very well have stumbled into discovering machine reasoning on a level that we haven't even realized yet.
The SOTA in this field as of a couple days ago is a Chinese open source project called Deepseek that just upended the entire industry and pantsed Sam Altman and Silicon Valley in front of the world. They're giving away the secret to their massively efficient results that are besting every benchmark, and have cracked a breakthrough unsupervised learning technique that is allowing these models to demonstrate powerful reasoning capabilities. This stuff isn't Google's botched AI telling you to smoke while pregnant, this is post-graduate mathematics.
This has been great for a couple reasons: 1. we get a lot more performance for less cost and power usage which is seriously bloomer news, and 2. it's been hilarious watching western chauvinists melt down over this massive flex out of China.
Read a good synopsis of the paper that just dropped if you're still thinking it's all a grift. I also hate a lot of the practitioners of this stuff as well as the additional strain on the grid and climate, but I'm telling you folks, this tech is legit and currently improving at an alarming rate!
Lmao literally the paper in the link you shared says (emphasis mine) “DeepSeek- R1 achieves performance comparable to OpenAI-o1-1217 on reasoning tasks” next to a graph that shows that the benchmark performance is roughly the same as OpenAI – marginally better on some benchmarks and marginally worse on others.
It is the same stuff as “Google’s botched AI” but now partially open-sourced and slightly more efficient, the latter being the reason why it “upended the entire industry”.
Yes, comparable performance to a multi billion dollar company for a fraction of the cost for training and inference. It cost $5 million to train their prior model. LLama 3 took an estimated 750 million to 1 billion to train and it's not as good.
These reductions in costs must be exciting to investors (you know, the 60 year old dipshits from the original comment you replied to) but they do not substantiate your claim that “this tech is legit and currently improving at an alarming rate” or any other beautiful epithet you used. LLMs remain hallucinating state machines with a probability attached to any word they spit out, and marginal efficiency improvements don’t make them any more exciting to the end user.
Cost means energy use in addition to hardware, so it should matter to climate concerned folks, not just investors.
Also these reasoning models are way less likely to hallucinate given time to think (test time compute). Think applications where it matters like augmenting or replacing existing labor, which will start to happen more.
The thing is, the 'efficiency' of AI isn't the problem. See Jevons Paradox. Much like our history with fossil fuels, as energy efficiency goes up, total energy consumed also goes up (because you get more for your buck now).
So the more efficient AI becomes, the more energy we will waste on pointless shit. I actually agree that anti-LLM sentiment is overly high on the left, but that argument has to be made that it has very useful applications for very specific circumstances, not that "it's getting better and that will solve all the problems". Because it getting better is actually also a problem in many ways.
I'm still convinced that these models will never be more useful than they are now. I sometimes use them to help with coding but honestly I could easily live without them, none of them can handle complex tasks with good accuracy.
Thinking they will somehow significantly improve on this is techbro shit. I'm certain there's more to cognition and intelligence than stochastic parroting.
You can keep banging this drum, and I understand what you are saying, but that isn't what we are talking about. We are talking about products and companies that companies like OpenAI and Softbank are creating and investing in. Products of which there is very little incentive for them to perfect, as then they will be out of the job, and companies that are likely little more than investment scams that will produce little results.
I mean, look at the rhetoric at it's face. Who the hell is going to fill those "hundreds of thousands of jobs"? Despite layoffs in the Midwest, we are still at basically full employment, and it's not as though secondary education is getting cheaper anytime soon. As usual, they are putting the cart before the horse, because they get paid either way when the government bails out their project.
There is no way for this project to succeed in the way they are claiming unless the U.S. actually pursues an educational and industrial policy mandate outside of what the 'market' dictates. They are building for a market that doesn't exist.
I don't disagree at all that these companies will produce little results besides grifting, but that's because they don't have a moat of any kind and they've admitted as much.
Yeah, I'm not a high tech guy, I'm a manufacturing guy, so I don't really know what that means, all I know is when I can smell horseshit coming from investement and management.
It means we're fucking lucky that the people who would probably enslave us if they could don't have any secret technological advantage thanks to these open source chads.
i think the left has been reflexively anti-LLMs because they haven't been up until recently worth the squeeze, they're emblematic of the continued ignorance of the climate crisis, the worst techbros you know are all in on them, and they're antithetical to the intentionalist view of art, and it's a convenient excuse to cut down on labour costs which are all imo perfectly good reasons to dislike LLMs but i think people are a bit blind to how accurate they've become and still have a view of their tendency to hallucinate based on GPT-3.5 and the memes about google's search summary being confidently incorrect, the latter which is representative of the absolute cheapest bottom of the barrel LLM google has (gemini-flash) and not SOTA which admittedly google has not really pushed.
I'm aware and will play with it at work. But would it not imply then 500 billion investment useless if true? if you can run full model on 4 blackwells stack (100k?) (and depending how swap works on 48 gb radeon (10 k bucks)?).
The announcement from deepseek was sucking the oxygen out of this entire dumb stargate announcement imo. It's like they're already wrong but committed. the amount I think demonstrates that they feel they MUST win this race at any cost, even if it will just mostly end up being a grift.
At least small one (8b) locally does not roleplay very smart, will need to poke the big boy later
That's mind boggling output for an 8b model compared to last year. It's funny, you can def see how that first answer is a distillation of r1's answer here:
That's at least aware of concept of free float and slippage
(i don't want to poo-poo small model over something which might have been pruned, but outside of fancy style posting, it's not that different from other small ones)The paper details more about how they got these things to train each other, so in theory you'd train these tiny 1.5b or small 8b models to specific tasks.
kind of a shit plan when it's going to decimate the already crumbling US infrastructure whilst making currently habitable areas of the US no longer habitable. As opposed to Russia, which climate change has the effect of making massive parts of it's landmass more tolerable to human habitation.
thats when you (further) annex canada to no actual resistance
xi jinping place dongfeng 21 missile launchers within alberta to deter imperialism whilst setting up the western socialist state