TheaterFire

South Korea’s “AI Squid Game:” a ruthless race to build sovereign AI

Posted by self-fix@reddit | LocalLLaMA | View on Reddit | 18 comments

Reply to Post

18 Comments

throwaway12junk@reddit

I still think this stupid. Korea lacks the financial might of the US and bottomless talent pool of China. If they want to stand a chance, they'd be better off forming a consortium of universities, government, and the *chaebols* and poling their collective capabilities into a unified model (preferably open). Upon completion it would automatically have buy-in by powerful institutions and key individuals, who all immediately have understanding of how it works. So the deployment of this "sovereign" would also be far faster and cheaper.
View on Reddit #76643451

TheRealMasonMac@reddit

\> ***chaebols*** and poling their collective capabilities into a unified model (preferably **open**). lol
View on Reddit #76645034

EtadanikM@reddit

There is practically zero chance South Korea will have the research talent to match the US or China in the near future.  Compute, sure, it can be built out with government support, but the research gap is huge and won’t be bridged.  Being a fast follower is already a huge achievement but it’s only made possible by standing on the shoulders of existing research, it is a pure fantasy that anything useful can be “built from scratch.”
View on Reddit #76690740

throwaway12junk@reddit

Then why structure it like a military contract with a single winner, instead of a unified national project?
View on Reddit #76645554

TheRealMasonMac@reddit

It's an incentive. Math Olympiads and the like also have winners, do they not? The "losers" actually gain as it's a learning opportunity.
View on Reddit #76645754

throwaway12junk@reddit

Maybe we're talking past each other. A Math Olympiad is similar to military contract in there being a very specific narrow goal with clear objectives. So for every challenge there's only one right answer, and the competitors are judged by how efficiently and quickly they got to it. In this case, the objective is to be "the best" in Frontier AI. Well what does that actually mean? Because you could end up with a hundred different models that are each "the best" at a hundred different fields, but 99 of them lose out on funding, then that's a huge amount of wasted potential. Imagine a situation where 20 years from now, someone builds a model that improves crop yields by 30%, only to learn a similar discovery was already made in South Korea but abandoned because it was one of the 99 losers for whatever reason. I'm coming from the belief that South Korea should strive for something like the Linux Foundation for their "sovereign AI" *then* have competitions to build things on top of it. But right now it feels they're trying to create a national equivalent of Microsoft. No doubt it will have its benefits but Linux is *far* more influential and powerful.
View on Reddit #76647401

TheRealMasonMac@reddit

Yes, perhaps I'm not understanding. Can you explain what you feel is problematic about South Korea's approach?
View on Reddit #76657260

throwaway12junk@reddit

Well what's great about it? What makes this winner-takes-all approach better than a collaborative effort? To me, all this signals is the *chaebols* leveraging their influence to get richer, and any benefits to the nation are useful bonuses. But if that's the point then good for them. Their $20 billion family fortunes will grow to $30 billion, and they can finally call themselves wealthy.
View on Reddit #76659596

Koksny@reddit

>I still think this stupid. Korea lacks the financial might of the US and bottomless talent pool of China. And yet with LG alone they are third player in global AI race, while everything that Europe has to show for is underperforming Mistral.
View on Reddit #76644340

EtadanikM@reddit

How is LG’s model better or more significant than Mistral? 
View on Reddit #76690052

self-fix@reddit (OP)

And LG achieved that without Blackwell GPUs that the Korean government is supposed to hand out starting this year
View on Reddit #76645076

RuthlessCriticismAll@reddit

> It's also built from scratch unlike China's Deepseek lmao, this is like Mistral's CEO yapping about how Deepseek owes so much to Mistral. When you are learning so much from what Deepseek shares you should act with some respect.
View on Reddit #76657591

TheRealMasonMac@reddit

What do you mean? LG is like 90% distilled from Qwen-235B.
View on Reddit #76645575

Megneous@reddit

Deepseek was trained on synthetic data from Western SOTA AIs... that's not exactly what most people mean when they say "from scratch."
View on Reddit #76651106

TheRealMasonMac@reddit

LG also used synthetic data from what seems like Chinese models, though.
View on Reddit #76655187

Serprotease@reddit

Mistral models are quite good? The 24b range especially. Mistral large is just okay, but so is the 235b one from lg. One could also note that Europe also have Bfl and the flux image models. As far as I know, Korea only has illustrious? And it’s quite a niche model based on SDXL.
View on Reddit #76656684

throwaway12junk@reddit

I have not, and am not, saying South Korea *can't* do this. I'm saying instead of trying to create an equivalent to OpenAI they should build something like the Linux Foundation. Pool the combined resources of the nation towards a single core project, the build out application and forks from there. Unless there's some nuance about Korea that I'm missing, I fail to see the disadvantages of concerted national effort.
View on Reddit #76645943

self-fix@reddit (OP)

Actually 1 company alive in the race is UpstageAI, a startup, and another that will be joining will be another startup called Motif Technologies
View on Reddit #76645175