00:00I understand that your company just agreed to an expanded collaboration with AMD.
00:05Could you tell me a little bit more about what's in the pipeline?
00:08What did you discuss with Lisa Su in Seoul?
00:11Right, I think the sovereignty of the AI is more important than any other times.
00:16And then one of the key ideas is that having all the data in our country.
00:20So in order to do that, we need a huge data center to run in our country.
00:25So of course we have a lot of NVIDIA chips in Korea, but we want to diversify to other chips
00:33including AMD.
00:34So that's why I asked Lisa Su to bring more than 10,000 GPUs in Korea.
00:41And then she kind of agreed on that.
00:43But we currently are negotiating how we're going to bring the chips in Korea.
00:48Are you also looking at any hardware, software optimization with AMD,
00:53or is it purely an infrastructural and financial investment here?
00:57Right, I think that is very important.
01:00Even though we have chips, we have our own LLM called Solar.
01:04The quick question is that how we can run fast and efficiently on AMD chips or any other chips.
01:10So our team itself has a vehicle, the team that makes it more efficient.
01:17But AMD is much better. They have an entire team working on that.
01:22So we definitely have technical collaborations with AMD as well.
01:25So what is your roadmap to scaling your language model?
01:28Because now it's still quite small with only I think about 10 billion parameters?
01:32Right, right, right.
01:33So we used to have 10 billion, but thanks to our Korean government support with lots of GPUs,
01:40now we have a 100 billion parameter model for the solar LLM.
01:45And then in this summer we're going to have a 200 billion model.
01:49And then hopefully end of the year we'll grow up to like 300 or even 400 billion,
01:56then which can be really a frontier level model.
02:00Tell me about your client uptake rate,
02:03because I understand you are now also looking to sell to overseas clients as well.
02:07Oh yeah, sorry, can I say that again?
02:10Tell me about your overseas expansion and how is client uptake looking like?
02:16I see, I see.
02:17So sovereignty is very important, as I said, right?
02:19So Korean, we are leading LLM companies.
02:24And when you look at other countries like Indonesia or Vietnam, UAE,
02:30they are looking for this kind of technology.
02:33They want to put their data on their country.
02:35They want to run the model in their data center.
02:39And then they ask us to help them to either, you know,
02:45they want to understand how this like frontier model is going to build.
02:48So we are currently talking with all these countries
02:51and try to transfer this technology to them
02:56so they can do all their own sovereignty on this AI and the data.
03:01How do you compete with all the AI startups like Minimax, Jupu,
03:05coming from China as well, especially when they're selling at such low cost?
03:08Right. This is a really good question.
03:10I think this low cost is coming from two angles.
03:13First of all, how can you build really smart and large-scale model with a cheap price?
03:20And the second, once we have that model, how can you serve with a very, very low price, right?
03:26For the first angle, we got a lot of GPU support from Korean government.
03:31So once we have a large, like a lot of compute, we can test a lot.
03:35We can experience a lot. We can do a lot of research.
03:38Our company has like really good technology to build high intelligence.
03:44The large model using like only small set of compute.
03:47So if we spend less money to build a very strong intelligence,
03:53of course we can make it really cheap.
03:55The second part, we, as I mentioned, we talked with,
03:59or collaborated with NVIDIA and AMD.
04:02We want to develop the technique to serve our model on their chip very, very efficiently.
04:07So the cost will be very, very low.
04:09Okay, and this is something that China doesn't have given their restricted access to US chips.
04:15I want to talk about that Squid Game version of the AI model competition in South Korea.
04:22Part of the South Korean government's plan to select a sovereign AI model.
04:26You're one of the four companies selected now in that race.
04:30How do you stand out from the pack?
04:33Right, let me just explain what is this KAI competition.
04:37So the basic idea is that we do have, you know, many companies in Korea as well.
04:42In the Korean government's perspective, which company they want to provide more GPUs?
04:47This is always a difficult problem.
04:49So the idea is that they select initially five companies and then they provide the GPUs.
04:55And then every six months they evaluate how good they are.
04:59And then if we advance in the second and third round, they're going to provide more and more GPUs.
05:05So eventually like the top two companies will get more GPUs than even maybe other, you know, large scale companies.
05:13Then we have like a compute power to do research to make the intelligence better.
05:18That's a Korean government, a very, very smart program.
05:22And then we were selected as an initial five and then we advanced in the second round.
05:28How are we going to compete and win?
05:30That's a very good question.
05:32But we are a startup, right?
05:33So we are like a very, very flexible.
05:36So currently we collaborate with many, many other startup companies.
05:41And also we developed some technology called the gaps of scaling.
05:44Yeah.
05:44Also we know how to generate, you know, data set from existing data set.
05:50We have so many technologies to compute with other companies.
05:53We've got a budget based in the first few years.
05:53I think we're going to do see that just a budget.
05:53I think you can see.
05:53I think I will see you as a budget based in the first round.
Comments