Software Huddle - Seattle Startups, AI’s Future & Big Acquisitions with Yujian Tang

Episode Date: March 13, 2025

Today on the show, we talked with Yujian Tang. He was on the show previously when he worked at Zilliz, when we talked about vector databases and RAG. He's since branched out on his own, building the tech startup scene in Seattle and organizing AI events all over the place. We talk about his latest venture, the Seattle Startup Summit, coming up on March 28th. They're still Early Bird Tickets available if you're interested. We also talk about AI models, the impact AI is having on programming, including our own programming projects and share our takes on some of the recent acquisitions that have happened in tech, including Voyage AI.

What do each of you use sort of day to day or like which models or products or things you are using in your day to day? So I mostly like when I do like LLM like work, I mostly just use GPT-4. For programming, I primarily use GPT-4 and quad depending on what the project is. It's very interesting whenever I think about like IBM because it's just like... They're making a lot of moves. It's just like big giant old company that like, you know, they're just, it's like, what is, what is this company doing?

I think our generation like doesn't think about them, right? Like we'd never think they like buy something from IBM, but yet they're, they're buying all these other companies. Yeah. Well, I mean, they're such a huge conglomerate. They're called international business machine, you know, like that's what they are. What a name.

Yeah. That's amazing. When's the summit happening? March 28th. Get your tickets, people. Hey everyone, welcome to Software Huddle. Today on the show, Alex and I talk with Eugene Tang. He was on the show previously when he worked at Zillow's,

when we talked vector databases and RAG. He's since branched out on his own, building the tech startup scene in Seattle and organizing AI events all over the place. We talk about his latest venture, the Seattle Startup Summit, coming up on March 28th. There's still early bird tickets available if you're interested. We also talk AI models, the impact AI is having on programming, including our own programming projects, and share our takes on some of the recent acquisitions that have happened in tech including Voyage AI as always if you have ideas or suggestions for the show, please let us know and with that

Let's get you over the episode. Hello and welcome to software huddle. I'm here with my forever co-host Alex debris Alex. How are you? Sean, I'm doing well. It's good to be back new year new me Yeah, good to see ya. Yeah, we kind of we we Kind of put the podcast a little bit on hold there for for most basically December and January. Wife kind of took over. I switched jobs. You had stuff going on. We kind of had to slow down for a little bit, but we're back in full force now. Yep. Yep. Always always good to get back. Yeah. Definitely got jammed up with like more work stuff than I anticipated for a while.

So it was a nice break. And I feel like, you know, that holiday time always ends up being kind of like a crapshoot anyway. Like, yeah, so good, good to take a break, but good to be back as well. Yeah, people aren't listening to podcasts as much anyway, during December. So it's a good time to take maybe take a break. Yep. Yep. All right. And we have with us back on the show, Eugene. Welcome back. Thanks for joining us. Thank you, Sean. Glad to be back. Yeah. So last time you were on, you were still, you were working at Zillas when we're talking vector databases. So, you know, you've moved on from Zillas. That was quite a while ago. You know, what's going on now?

Yeah. Yeah. So I was working on the Vector Database stuff, working on some Rack stuff, working on some agent stuff back in 2023, early 2024 as well. Basically, what I was doing in Zillis is I was doing these developer communities, right? So I was helping build the community there, trying to get people to come together to talk about AI, talk about what's going on in the space.

And one thing I found about running these events is you're a little bit limited on your topic set, right? You're basically like, hey, we're going to talk about Vector's databases, we're going to talk about RAG. And then that's pretty much it. And so I was like, hey, what if I had a community where we could talk

about a lot of different things, where it's like a whole spectrum? Because as a dev, and I imagine that's why people listen to this as well, is you get to just hear so many different things when it's more third-party-ish. So I started my own community last year. We started with hackathons. We do some tech talks in SF. We do hackathons in Seattle. we grew from about zero to seven thousand members. We have 11,000 people attend our events and this year I basically turned that into something called Seattle Startup Summit which is going to be like a hopefully, hopefully 800 person event. We'll see. We're at about 460 some signups now. So we'll see if we can hit

that 800 mark. Yeah, very cool. When's the summit happening? March 28. Get your tickets people. Is this live? Is this live? This is not live. Oh, okay. Okay. Okay. Yeah. It's March 28. So it's a couple of weeks. Yeah.

When you're doing an event like this, like, you know, say you want 800 people to be there, do you have to like oversubscribe? Like, what is sort of the attrition rate with this sort of stuff? Yeah, 100% you have to oversubscribe, right? So with free events, you're pretty much looking at it actually kind of depends on the location too, because it's like the culture of the place, right? So in SF, you get about a third of people show up. If you're lucky and you have a really good event, you get maybe 40, 45%.

If you're like Databricks, you probably get 50%, you know? Or, you know, one of those big summits. With a paid event, or in Seattle, you get about 50%. With a paid event, so like Seattle Startup Summit, I'm charging for tickets, right? I do that primarily because I want people who actually sign up to show up. But even with charging for tickets,

even with charging for tickets, I found that, you know, you still have maybe like a 10 to 20% of people that show up. And so like the tickets I'm charging for the summit are like $250, like full price tickets. But a lot of people buy like, you know, half off tickets if they can find a code or, you know, they get one of the early price tickets. But a lot of people buy half off tickets if they can find a code or they get one of the early bird tickets.

And then I'm like, later you sign up for a conference or later you sign up for an event, the more likely you are to show up. So I don't know what the interest rate is going to be. I hope it's going to be like I'm aiming at 1,000 signups. So that's what I'm aiming for. We'll see if we get there. Yeah.

Cool. And what kind of person are you targeting for devs, ad startups, people that are interested in startups? What are you looking for here? Yeah, yeah. So my primary goal with this community that I build is, hey, AI is changing everything. It's not just devs and software.

It's like healthcare and finance and all this kind of stuff. But I think on the front line, devs are your target audience. Before you can have your like AI finance apps, AI healthcare apps, you're going to need devs that are literate in this space. And so my goal to community is really to get devs all the information, all the resources, all the tutorials, all the networking, all that they need to become literate in this space. And so my goal of the community is really to get devs all the information, all the resources, all the tutorials, all the networking, all that they need to become literate in this space.

So my primary audience is devs. And then, you know, like secondary audience, like tech leadership, of course you have to have, you know, tech leadership there. They got to buy in founders, people like that. Actually, one of the things with Seattle Startup Summit that I feel very strongly about is, oh, so Sean, I know you're in the Bay,

Alex, where are you based? I'm in Omaha, Nebraska, so I'm always away, yeah. Okay, okay. Yeah. Okay, so. Not a tech hub, that's for sure. Not a tech hub, yes. But so Seattle is like one of those weird places

that's like, it's a tech hub in the sense that there's a lot of technical talent here. There's Amazon, there's Microsoft, there's Snowflake, there's Databricks, everybody's here basically. There's actually not a lot of startup activity. If you go and you look at what people are talking about, you can see that when you hear people talking about tech hubs and AI,

you always hear people say something about Seattle. But then if you hear people talking about tech hubs and AI, you always hear people say something about Seattle. But then if you hear people talk about startups, you'll never hear people say anything about Seattle. Unless they're talking about, here's a tier two startup city. Why is that? Yeah. Yeah, that's actually one of the things that I was like, what is up here? What is up with this? My opinion is, one of the things that I want to do is I want

to nurture that developer to start a founder pipeline. And I think part of it, and once again, I'm not entirely sure this is just like, you know, you're just taking shots in the dark until something sticks, right? Like my opinion is like, there just needs to be more activity around it. You know, there's not as much activation. If you go to the Bay, like Sean, you've probably see this. There's events every day of the week.

Yeah. I mean, I think that's why the attrition rate for events is so high is because people get invited to so much stuff. They just sign up to all of them and then they're like, I'm kind of tired today, so I'm not going to go. Yeah, exactly. So that's one of the things where I'm like, yeah, you know what? We should have this in Seattle. Yeah, it is surprising because Seattle, it's not like Seattle, like it from like a proximity to like a VC hub, like Seattle is

not that far away, you can do a day trip to San Francisco if you needed to hit up, you know, you know, a bunch of VCs or something like that to run through pitches. I wonder, in other cities in the US, like New York, Boston, cities that do have a pretty decent startup culture and a number of startups there, what is it that they did in order to build that?

Is it that they have some sort of incubator or something like that, or some prominent person? I also think is that they have like, you know, some sort of incubator or something like that, or some prominent person. I also think sometimes that you have like one company that emerges as like a really successful startup out of a place and then suddenly everybody's like, oh, okay, like I can do a startup in, you know, wherever Omaha, Nebraska. You know, I think about that too. And so like, SF has a bunch of different startups that are very successful, obviously. Seattle has had two huge successes in the last like, well, now it's like 50 years. But you know, like, yeah, why doesn't that fling off a bunch of other sort of people

leaving there and starting something else and just like creating a whole ecosystem like like we've seen in other places? Oh, okay. So I actually have looked into this a bit. And here's what I think that is. and starting something else and just creating a whole ecosystem like we've seen in other places. Oh, okay, so I actually have looked into this a bit and here's why I think that is, okay? So Amazon and Microsoft are so big and they pay you so much money that it is very hard to leave. On top of that, a lot of their employees,

a lot of their software developer employees are actually people who are on H1B visas. Interesting, okay. And you think that's more, I mean, cause like that's, some of that's gotta be true about the pay in the Bay area as well. But then do you think there's less of the,

I guess H1B just proportionally in SF compared to Seattle? It could be. So Seattle has like the two big, like Amazon, Microsoft companies that do that. I think Google and Meta, Google, Meta, Apple, they all they all do that as well. I think a couple of other things that are at play here as well is like the Bay has Stanford, Berkeley, right? These are very big school.

Well, I get a Berkeley is a very big school. Stanford is not that big, but it turns out a lot of people. You we have you dub up here, which is one of the research institutes. We also UW up here, which is one of the AI research institutes. We also have AI2, which is a really good AI research hub as well, research institute as well. The thing with this is AI2 is tiny, and UW is roughly the size of Berkeley.

And so also, there's a lot of people who go to UW and then go to AI2. So it's not kind of like the same as you have down there with Berkeley and Stanford. And then also Silicon Valley, it started because it got its name from the people who were doing the Silicon chips back in the 60s, right? Like Hewlett Packard and whatever.

And they were successful and they have a ton of money and they're hanging out in Palo Alto. So you've got like, you know, you walk right the whole like, the whole like startup, like shtick meme, whatever, is like you're walking down. What is that? Stan, Stan Hill Road? Yeah, yeah, yeah, that road and you know, there's like billions of dollars like right there. And in Seattle, you don't really have that. We have Madrona. Madrona has like $3 billion under management. Here's another plug, John Turrow from Madrona,

he's a partner who's going to be coming to speak at Seattle Startup Summit. But there's just not as much money, like MVC money here. There's a lot of money in tech money. People who, there's a lot of money in like tech money, right? Like people who like there's a Microsoft executives. I don't know if you have you guys have either of you guys been up here? Yeah, I was there last month. Yeah. Oh, yeah.

Oh, okay. Cool. So in Bellevue, there's this place called Medina, Medina, something like that. And this neighborhood is literally like just mansions. And that's all the Microsoft execs, all the Amazon execs. So there's a lot of money up here, but I just don't think it's being activated in the same way. Boston, you know, Boston also has MIT, Harvard, a lot of like medical stuff out there as well. Yeah, Boston, I think, has a good base of bioinformatics, biomedical-focused startups.

We're only tech-seeing around that, which I think has emerged from those universities. Yeah. Yeah. And of course, New York has Wall Street. Yeah. Yeah, exactly. So what's the agenda for the summit?

What's the, if I'm attending, what should I expect? Yeah, so I think the agenda for the summit? Like, what's the, you know, if I'm attending, what should I expect? Yeah, so I think like the idea with this summit is like really like, hey, I'm gonna try to gather all of the people who are working on startups in Seattle. Now I haven't gathered everyone, but I've gathered a lot of people. So at the beginning, we're gonna start,

we're gonna talk with John Turow. I haven't like, I don't know who's gonna moderate this yet, but we're gonna have a talk with John Turow. I haven't like, I don't know who's gonna moderate this yet. But we're gonna have a talk with John, he's he's the partner of Amazon, and he's gonna talk to us about AI, AI agents, AI infrastructure, right. So this guy was building like a voice agent, like, company back in like 2011, 2012. And, you know, he got he got out of it, he went to Amazon, he built for like 10 years. And now he's kind of like seen the investor side,

the startup side, et cetera. So he's gonna be giving the first fireside chat. The second fireside chat that we're gonna do is with Pasquale De Mayo. So he's currently the head of Amazon Connect, right? So that's their customer service platform. The story that we wanna cover there is like, he was part of that journey from the beginning

where they took Amazon Connect from zero to 150 million in like less than four years. And that's something that I think a lot of founders would find very interesting, even though they were internal to Amazon, it's still like, that's a lot of revenue very quickly. And then the third one is we're doing a fireside chat with Vijay Raji.

So he's the CEO of StatSig. Have you guys heard of StatSig? Yeah. Yeah. They've been doing a lot of stuff in the Bay. They're like experimentation. And one of the things I talked with his team about his statistics in person

five days a week, by the way, and they're based out here in Seattle, and they do stuff in the Bay. So I was like, I was like, dang, like you guys all travel like that far to the Bay like, you know, pretty often do this. And they're like, yeah, we're really interested in having more people come to Seattle. I was like, me too. So, you know, that's part of it. And then we'll have some, we'll have a couple of panels. So let's see who's on our first panel.

So we got Alex Ye from GMI Cloud. So GMI Cloud is one of the Neo Clouds. I think we'll touch on this later as well. But they're one of the GPU Clouds. They just raised like 80, 8282 million, something like that. Then we'll have, who else is on that panel? Tony, Tony Holdstock Brown from Ingest. So he used to be in the healthcare space actually.

And he's got an interesting story. So Ingest was like an orchestration platform. So you guys know like AWS like step functions, right? Yeah. Yeah, so they do something kind of like step functions. And then like when the whole AI agent like thing picked up, they found that people were using what they were building to build AI agents.

And so they kind of like pivoted into this space. And then there's also Rob Williams from ReadAI. So for context, Riverside is the third streaming platform that I've been on. ReadAI is like one of like seven, like, you know, meeting assistants that I've seen. And they're doing quite well. They just raised the $50 million Series B. I think they're expanding into the enterprise. So that panel is going to be talking about the AI stack.

You'll notice that all three of these people are from different pieces of the stack. So Alex is from the super infrastructure core layer, like GPU cloud. And then Tony is the orchestration framework layer. Like you want to build an agent, this is what you use. And then Rob is from this application layer, where it's like,

here's what we're doing, here's how we're using AI, right? And then our other panel, I've titled it Enterprise Insights, which honestly is a very vague title, but these people work with enterprises, right? So we have Sharon Zhou from Lamini. So by the way, I don't know if you guys know this company, right? Lamini. Have you heard of this company? No, no, I don't know that one. Okay. Well then you, you won't have, you won't, you won't, you won't have

the context for what I'm about to say, but it's spelled L A M I N I. And I thought to myself, I was like, Oh, this has got to be pronounced Lamini. And then I talked to one of my friends and she was pronouncing it Lamini. And I was like, what is this has got to be pronounced Lamini. And then I talked to one of my friends and she was pronouncing it Lamini. And I was like, what is this? How do you say this? And then I got on a call with their team and they're like, oh, it's pronounced Lamini. And I was like, how am I supposed to guess this?

And then they're like, no, it's a real word. It actually means multiple llamas. It's a type of llama or llama types or something like that. I was like, wow, okay. Wait, like a herd of llamas is a Lamini? Yeah, yeah, yeah. Yeah, something like that. And I was like, wow, okay. Wait, like a herd of llamas is a llama night? Yeah, yeah, yeah. Yeah, something like that. You didn't know that, Alex?

I did not know. I'm on their website right now and in the M of their thing, it's got a little llama embedded in there. So I could see it's llama related, but it's weird because it doesn't look like llama generally. It's only got one L rather than the two. Yeah, so. Does this have anything to do with like, the llama model from

meta? Or is this completely unrelated? No, it's totally unrelated. That's a lot of llama on like, I would be a little bit worried about sort of trying to use another mama related word for something that's so well-known and branded from meta. Yeah, I don't know, man. I think people, because it's LLM, people are like, oh, llamas, llamas.

But yeah, so they're like an AMD cloud, right? So they basically provide you AMD chips. And Sharon is interesting in the sense that she is a student of Andrew Ong, who is super, super popular in this space. And then, who else on that panel? Denny Lee from Databricks. You guys know Denny?

Yeah. Yeah, yeah. He gets around. He's everywhere. You can't not know Denny if you're in this space. And then Yina, I don't know if I'm saying this correctly, I think it's Yina Arenas, but she's the VP of Azure AI, Azure Core AI,

so they're like new AI platform. She's also based in Seattle. So at first I thought she was based in the Bay, I was like, oh, yes, we're gonna get her to come out from the Bay for this. And I was like, oh, she's in Seattle. I was like, oh, even better. She has tech leadership from AI in Seattle. So yeah, that's those are the panels. And then afterwards, we'll have a lunch break and you know, stuff

like that. And during the day, there's like a bunch of exhibit tables, you know, like when you go to a conference, there's always exhibit tables, you walk up to the zip tables, you talk to the people see what's going on. In the afternoon, this is my like, I would say like my like defining feature that I do in my events is like the demos, this is my defining feature that I do in my events, is the demos. It's always really fun to see lightning demos. If you ever come to my events in the Bay,

you'll see that I have people do these two-minute demos. And after the two minutes is up, I will shush them off the stage. And I'll be like, get off the stage, get off the stage. And so at the end, at the afternoon, we'll have four-minute demos. I'm giving people a little bit longer. It's a little bit big. Very generous. I know, right?

But we're having these four minute demos at the end. And currently we're at 26. I have spots for, oh no, 28. I have spots for 52 of them. So we'll see if we get up to 52. If it's 39, I'm happy to just cut off the last 13. I'm doing like 13 an hour. So. And so these will be like AI companies showing off what they're doing.

Partly with the idea to like, I mean, are they attracting dev talent? Are they attracting investment? Are they attracting customers? Like what's the goal there? Yeah, yeah, yeah. So there's the idea behind this is like, the original idea behind this was that,

and it's kind of evolved a little bit, but the original idea behind this was like, some early stage companies, and I've given a lot of free demo slots to early stage companies, is early stage companies really need product feedback, right? Like devs, here's what devs are good at, building things. Here's what devs are not good at, talking to people.

I was like, well, what I can help you do is you can come and show us what you're building, and then people will come to you. You don't even have to go out and find them. They will come to you and they'll tell you, hey, this is really cool or hey, have you thought about this or blah, blah, blah, blah. That was the original idea behind this.

Then for this event, this has evolved into like, OK, like we, our theme is AI dev tools. Let's try to get a whole view, like, you know, just adding in that idea of like getting like the whole view of the landscape. Let's see as many AI dev tools as we can and like get to kind of like compare and contrast. That's cool. Because I find like, you know, I talked to a lot of these companies that are interested

in doing, like every company in the world basically has some sort of AI initiative. But I'm saying like 99% of them have no idea what that actually looks like. So a lot of it comes down to, I think them seeing like the art and the possible. So if they see like what other companies are doing, and then they're like, oh, okay, like, we could do something similar, or we could, you know, maybe we can work with these people or whatever. But I think there's such a huge

interest in this space, but very few people who kind of have like a definitive plan for what that means. Yeah, 100%. I think that's that's like, that's definitely part of it. And I think also, you touched on something else there, which was like, you got to see other people do things. So you kind of like have an idea of it. That's also part of it for me is like, about I think about like 70% of people

who have signed up for this conference are like some sort of builder, like product or engineering, some sort of builder like that. And I think like one of the other things is like, if you see other people doing startups, you see other people who are devs that move into this founder kind of role,

you're more likely to go and take that job. Take that like, hey, I know this guy. I know what he can do. Maybe I can do that too. I know what she can do. Maybe I can do that too. That's also part of this as well. Awesome. Well, it sounds like a great event. It looks like you got a good set of speakers there. Should be really interesting for folks.

Hopefully, if you're in the Seattle area or you want to travel there and you're listening to this, you'll sign up. One last question. If you happen to be in the Omaha area, is there a livestream or anything like that I can watch? I especially want to see the demos and see all these things.

I think those are so fun. That's such a great idea. I don't know if we'll livestream, but I just, oh my God. Okay, so this conference is like $100,000 to do. And this was like before getting like the recording AV. And so I just was like, you know what? We should record this. How much more is that gonna be?

And they're like, oh, it's gonna be another $10,000. And I was like, okay, well, we'll figure this out then. And so I was like, all right, we're almost at the breakeven point. And you know what? I just went ahead and got the recording package. We'll record it. I don't know if we'll stream it, but we will record it at least.

Okay. That's perfect. I just want to see all these demos and watch that. Yeah, that'd be great. Yes. I will let you know, Alex. It's like the YC demo day. Exactly. It's like demo day, but in Seattle. Yeah, that's awesome. Cool. Let's get into some other AI-related topics here.

I think there's been a lot, certainly since Alex, you and I last talked, but even the last month or so, just like models, models, and more models everywhere. DeepSeq basically came out, broke the internet. Now almost like DeepSeq seems like old news at this point with everything that has happened since that. It's like, you know, quad three sevens out, OpenAI is launching 4.5 soon. As response to DeepSeek, like Google has Gemini Flash 2, OpenAI has O3 mini. Like, all so much stuff is happening.

You know, maybe we'll go to you, Jensen, you've been working in this space for a long time. Like, what's your thoughts on this? Do you have any specific takes? I mean, I love it. I think it's great. There's a lot of AI models coming out. I think that's always fun.

Competition is what makes people make more like interesting things. There's a lot of talk around GPT 4.5. So I just did an event in SF last week where it was a full day AI production mini conference. Basically, people were just talking about how GPT 4.5 is really expensive. That's the number one takeaway that I think people are getting right now. Someone said that O3 was really nice. I didn't get to see a demo of it, but someone did say that O3 was really nice.

I'm actually, for my workflows, I'm using GPT-4. I'm like, I don't even, I'm not even doing anything that complex. So I'm still just playing around with GPT-4. I think, you know, like, I don't know if I mentioned this to you last time, Sean, but I think one of the things that I find particularly interesting about this is everybody that,

at least that we can see, is still on this kind of like transformer model architecture, right? And then last year, literally December 31st, 2024, the last day of last year, and I don't even know how they got people to do this, release this at the end of December, but Google released this paper that is on like a slightly different architecture.

And it's called Titans. And they use this thing called a long-term memory module. And I think I would like to see... I don't know if Flash 2 is built on this. I don't know anything about Flash 2, actually. I don't know a ton, but I believe it's a derivative model from their existing Gemini models because it came out so quickly. I think a lot of them are using a lot of these models, even though the incremental 0.2, 0.3 are usually using the same base model, and then they're doing some additional fine-tuning

on top of that to create a new version of it. It's probably still based on the Transformer model, but this new paper they released called Titans has this interesting thing called a long-term memory module, which makes it kind of work. So it's a lot smaller than Transformer models, which means it's more efficient, which means it's less costly, it's going to less like energy cost,

it's better for the environment, et cetera, et cetera, et cetera. But the thing I find interesting about it is it learns on... It learns kind of like people do, right? So what the long-term memory module does is it says, if you see something that's very statistically unlikely to happen based on your current data, you should remember that. So, for example, if you're walking down the street and all of a sudden one of the trees turns red, you're going to remember that, you're like walking down the street and all of a sudden, like, one of the

trees turns red. You're going to remember that because you're like, what's going on here? This is not, this is not unusual. So I think that's like an interesting architecture thing that's happening. I think like DeepSeq, for example, the whole like distillation model thing. I don't really think it's quite the open AI killer that people have been kind of like making it out to be. It is very popular, of course. But the whole like distillation model thing, like, yes, you can train it for a few

million dollars, but that's the last training run. They still took data from open AI, you know, they still need that that large language model data. And then the primary cost is inference. So I don't think it's quite as much of a breakthrough as people think it is. But yeah, that's kind of what my opinions are. What about you, Alex?

What about you, Sean? I mean, the interesting thing about DeepSeek is just the state of the art, it seems like, is still going to be super expensive to train all the time. But then anyone is going to be able to replicate that pretty quickly and easily. So it's just like, it seems like is still gonna be super expensive to train all the time, but then anyone is gonna be able to replicate that like pretty quickly and easily. So it's just like, I guess it's gonna be a super decentralized technology, right? You know, like it's gonna be hard to stop in terms of just like, if you get a new model out there, people are just gonna be able to copy that and really say open source or whatever

they want to do there like pretty easily. And like we've seen like cost of tokens is just like continuing to plummet all the time. You know, you're saying like GPT 4.5 is expensive, but that's gonna be,, like, the cost of tokens is just, like, continuing to plummet all the time. You're saying like GPT 4.5 is expensive, but that's going to be a tenth of that in probably six months or a year, something like that. So it's just like, yeah, it's pretty wild, just the improvement of diversity, I think. Yeah. I mean, I think that, like, overall, my impression when I saw all this come out and people were talking about, like, oh, like, the reduction in tokens and, like, the impact this is going to have on GPUs and NVIDIA and stuff like that, like lowering the cost of tokens doesn't mean there's going to be less use of GPUs. It actually means the opposite. It's just like when we lowered the cost of transistors, it's not like there was

less computing and that there was a whole lot more computing. But so this is like net good, I think for everyone. And also going back to your point at the beginning, Eugene, about competition is good. It's forcing companies to essentially compete. And that's why you have new launches on December 31. Or when people are launching models on weekends, they're not waiting for the right time.

It's such an arms race essentially between all these different companies. And I think that's good for the can ultimately good for the consumer or good for the developers are building on some of these technologies like there's a there's an open source project out of Stanford, I forget the PhD students name, but it's called s one. And you can you can download it, you can try it out. And like they did basically applied the same approach

of this distillation using open AI originally using their reasoning tracing and then open AI shut that down. But then they use DeepSeq as well to do the training and they get a similar performance and the final training run costs $20 to run on GPUs. So you can recreate this whole thing yourself if you want, if you download their source code and run through it.

So that's great for, I think, creating this distributed environment. And I also think that a lot of the model companies are, because there's so much competition and the switching costs from model to model is not that expensive. You're literally mostly just changing like a string value typically, then they're I think they're trying to find other ways of having sort of stickiness in their platforms

and adding value that goes above just like, hey, here's a model that you can use. You know, they're going sort of up the stack, they're adding more like agentic capabilities, reasoning capabilities, data integrations, all this kind of stuff. So, you know, it's definitely an interesting time. Or even just like focusing on certain areas like Cloud 3.7, they were saying like,

hey, we realize everyone is using our thing for code. We're gonna train them, like make it much better at just coding and like worry less about just like being a general purpose thing, which like it's good for general purpose stuff, but it's like amazing for coding, you know? So yeah, I think they're trying to find their niches and stuff like that. I think quad, you know, anthropic really prides itself, I think on, you know, safety is one, you know, key thing. And then also being really good at coding problems. And I think like real world reasoning, like they, I listened to their CEO talk recently in a podcast, and

they were, he was saying how, you know, a lot of times when models come out, they're sort of optimizing for like the the wow benchmarks of like, oh, this, you know, model can pass the LSAT or something. But like, that is very different than what a lawyer actually does in the real world. Just coding problems in Olympiad are very different than the engineering type, like the real world engineering that you're doing. So Anthropic really tries to focus on, how do you do this in the real world? Yeah. What do each of you use day to day or which models or products or things are are using in your in your day to day?

So I mostly like when I when I do like LM like work, I mostly just use GPT-4 it's pretty cheap and I have my open AI key saved and I Just I don't even want to bother switching actually, you you know what? I do play around with GitHub models. So GitHub has this thing called GitHub models, which I actually only found out about because I'm a LinkedIn instructor. And they're like, hey, so we're releasing our own like set of ways to access LLMs.

And we want our instructors to like test them out. And so I'm playing around with that. They have, they have. They have OpenAI, they got DeepSeq, and they have another one. I forget what the other one is. But interesting. So it's routing through a unified GitHub interface, but it's going to the different provider models that you specify?

I think it's Azure underneath. I think it's GitHub, and then they just put all these models on Azure. Yeah, OK. Like, self-subsidiary, right? Yeah, yeah, for sure. Okay. And so that's like, you're talking about that's when you're using those when you're like building something into a product or like you're using those as just like chat type stuff or like what sort of use is that for? So I don't really chat with elements. I don't really use them for that, no. Oh man, I use that all the time.

I use it so much, yeah. I'll use it for things like, hey, like here's this copy that I'm writing, can you help me edit it? I'll use it for things like, hey, I need to organize this information, can you help me organize it? Or I'll use it for like, oh, I need to organize this information. Can you help me organize it? Or I'll use it for like, oh, so for example,

I took like the attendee list for Seattle Startup Summit and I fed it into ChatGBT and I was like, hey, can you give me a breakdown of like the demographics of this? Like who's, you know, what, are they engineering? Are they products? Are they marketing? Are they tech leaders?

Are they whatever? So it's, are they marketing, are they tech leaders, are they whatever. So it's really useful for things like that, but I don't like chat with it. And then the other things I play around with a lot is I build a lot of tutorials. I build a lot of tutorials because I teach courses and stuff, and so I just mess around and put different ones on there. Yep, yep. And are you using like cursor or anything like that? Or are you just like, are you using anything to help you generate the tutorials or you write it all by hand? No, I do, I write them all by hand.

Oh my goodness, man. What are you doing? Get you some cursor when you're like, you're doing all by hand. I don't know if cursor can do this. It might be able to do this now, but like when I tried it, and now it's been a year actually, wow.

When I tried it last year, it wasn't really able to generate like Lama index or like main chain or like, you know, codes from like more recent frameworks, which you might be able to do now. I'm not really sure I haven't tried it recently and I probably should. I would give it a try.

I would say get cursor and get clot three seven and like let it run. I bet like, I'm pretty surprised at like what it at what it can do. Last week I was out of town. I went to Disney with my family and I heard Cloud37 came out and I was just like, oh man, I want to try this, but I told my wife I would not work on this trip and all that stuff. I just got back this week and I'm like, wow, it's pretty wild. Yeah, it's good. Yeah, it's good. I would recommend. Yeah, I use primarily for my consumer-facing application side, I use primarily ChatGPT.

I'll use Gemini a little bit as well. And I started playing around with Notebook LM2 from Google for doing research. But for programming, I primarily use GPT-4 and Quad, depending on what the project is. And sometimes I'll use the earlier models of GPT-2 as well, like 3.5 and stuff like that. If the thing that I'm doing doesn't really, I just need to do the quality doesn't matter that much.

Then I'll do that just to reduce the token costs and stuff like that. But primarily, Claude and GPT-4 are my go-to models. Gotcha. Do you use any like Cursor or even GitHub Copilot or anything like that? No, I use, I do want to experiment with Cursor, but primarily, I do use OMS a lot for helping with coding, but I'm

mostly using like chat GPT for a lot of stuff right now. Yeah, yeah. What about have you any of you have either of you used v zero from Versailles at all? I haven't like, but I think, you know, there. I mean, I do a lot of next JS for like the front ends of things I build, but most of the stuff I'm building is from like a UI perspective, is really just like to serve a purpose of to go do something more interesting on the back end. So like, I, you know, I wanted to look half decent, but I don't need to like, I'm not building like full on, you know, end to end front end applications and stuff like that

for the most part. Yeah, that's true. That's true. I'm like, I'm like mostly a backend dev, but I have like a project right now where I'm doing like full stack, including some front end stuff. And it's like, I'm just, I can, I can write the front end stuff, but I cannot design it. So even just to like, I just like,

I'd ideate whatever with, with V zero and just like get some, some stuff. And I know it's just like sort of, you know, it's not super unique, but it's like, it looks good by the standards of what I could do anyway. So I like to get that just like a first cut. Yeah, even I fell in even like, techie Petit, if you're like, hey, I'm writing in next.js, I need a form to do this. And then you're like, make this look more professional, like this

website or something like that. It you know, it does a pretty good job of kind of giving you out of the box styles and stuff like that for these one-off projects. I think if you're doing something more professional, then you probably want a more integrated experience that can do a deeper analysis of what's actually within your code base and stuff. I think one of the interesting areas that's emerging now. Obviously, all these coding assistant, I think when it comes to AI, the initial frontier was, how do you do supercharged code completion?

And then it went from there to, how do you go from a prompt to actually being able to generate a more comprehensive part of an application or some fully formed application? And now there's a lot of going on as well with just like automating PR reviews and stuff like that, which I think is interesting.

There's a bunch of companies that I think are finding a lot of success. And that seems to like the next frontier of optimization. I know CodeRabbit is doing that. Who else is doing that? Let me get back to you on that. Okay. I've only seen CodeRabbit because... Oh, SourceForge, sorry. SourceForge is.

SourceForge. Oh, but they've been around for a while, right? Yeah, so they just released an agent for doing this. I talked to them recently, but it looks really interesting what they're doing. Okay, cool. I think one of the things about, like, with, I think one of the reasons, you know, there's been a lot of success when it comes to engineering is because, you know, one of the challenges with any of these, you know, LLMs is, you know, around reliability, but, and how do you build like a good eval set for like testing the application? But there's so much example, so many examples exist as well as when you're coding,

you have additional checks and balances like I can run it, I can compile it, or I can run it through unit tests. If it's a PR, there's tons and tons of example PRs that are out there that can become sort of my grounded source of truth that I can test against. And that really really really helps Whereas if you're doing sort of just like super general purpose stuff Like what is how do you build an eval test set for that? Like how do you test it and measure the reliability of is really really hard Have you talked to the folks that arise or or common about this? I have it. Okay. Okay. Let me let me let me get you

Taxed it. I'm sure they have some interesting things to say about this, because that's their whole business. Wait, sorry. Their whole business is evals? Or it's? OK, interesting. OK. I have talked to BrainTrust and some other folks in the space. Yeah.

So I mean, I think this is also another area where a lot of companies, but this is the emerging area too, is how do you test these things? What's the framework look like? I think a lot of companies end up building sort of hand curating their own eval test sets to start and stuff. But now there's companies that are focused on this.

But for general purpose stuff, it's a challenge just because the range is so wide. I mean, that's like the advantage of the sort of predictive ML or purpose-built models is the, you kind of knew what the inputs and outputs should be. So you can build a pretty good test set to measure, you know, precision and recall. It's hard to do that with like a general purpose research

tool or something like that. Yeah. Yeah. Yeah, for sure. But I agree with you, like the rapid feedback loop in programming makes it so much easier to like use this stuff where I was talking I was talking to a lawyer the other day I'm like how much do you use this he's like well we tried it but you just have to like you have to review it so closely anyway it's like I might as well just do it myself and in a lot of cases because there's not that like you know did it

compile or or did it pass tests or does it look right on the page or anything like that you have to like really look through every single word a lot to a lot more yeah especially something where the consequences of being wrong. I mean, the consequences of being wrong in coding is high potentially as well, but you have those additional checks and balances. You have automated tests, well, hopefully, and you can run a compilation. You can make sure that it compiles at least. Presumably, there's a review process and stuff like that. So there's like multiple stages that it's gonna go through.

Not every profession has that. They're basically banking on this like one person who's responsible for it to not only craft it, but make sure that is correct. Yeah. Yeah. I think it's really interesting that like, you know, you give a bunch of software engineers, this AI that can do a bunch of things.

And the first thing they do is, oh, how do I make it so that I have to do less work? I'm gonna jump in there now. Yeah. One of the things I've been thinking about is, over time with any industry, and this has certainly has happened with engineering is, because there's so much to know at this point,

there's a lot of like specialization. You know, as you go more and more in the industry, you might be like, hey, like I am not only a front end engineer, but I'm a front end engineer that has deep expertise in this particular framework. Or, you know, I work on this part of the back end and I really know, you know, these technologies deeply and stuff like that. But I kind of wonder is, if you could have a tool, like an AI tool that can essentially behave like the expert, and I really just need to know more of a high level sort of architectural standpoint and work at a bit of a more abstract level,

does that change the way this kind of specialization does? Suddenly having these tools at your disposal, it pays off to be sort of more a million miles wide and an inch deep to be able to give enough sort of instructions to these tools to get what you need. But you don't necessarily need the deep expertise of like, OK, well, I know that this is happening at the thread level when I'm calling this particular function and stuff like that.

I think that's a really good application, actually. this particular function and stuff like that. I think that's a really good application actually. I hadn't even thought about that. This is what keeps me up at night. I definitely think so. Like I've seen that myself over the last year. Like I have taken on like a few full-step projects and I'm like not a great front end developer but it's like I can make it work.

I did a mobile app. I'm working on a mobile app right now. I've never done mobile work before, you know, last year. And it's just like, yeah, I think if you have a good base of skills, I think you can apply it just to like way more areas than you could before. And you're not gonna be as good as like the people that are true experts and craftsmen in that particular area,

but like a lot of apps don't need that, you know? So yeah, I think you can like, absolutely. I think you can branch out, absolutely, I think you can branch out into a lot more things. It's also like good for like the, like, I guess, like the software engineer market in general, like, because I think one of the things that I at least I remember seeing this, like five years ago, is people are like, oh, like, you know, what if you get really into a framework, and then it becomes like obsolete, you know, becomes obsolete, then you're just screwed. But now this idea of being able to be like,

hey, I understand the ideas, the applications, how to apply these things in software. I'm just not an expert in this, but I can just ask the LLM. That's actually really helpful and I think that's better. I think that's better as a software engineering experience. Well, this is partially biased in my own experiences.

But in software engineering, people who are software engineers, who become good software engineers, are people who really enjoy learning about the high-level architectural stuff and how do you put all these pieces together, and they have a lot of curiosity. And so like you were saying, Sean, it's nearly impossible to learn everything.

And so I think this is nicer for the people, that persona of people who enjoy the software engineering archetype, but don't necessarily wanna get super deep into things just because it's literally impossible time-wise. So I think this is a really cool idea. I haven't seen anything that particularly does this, but I guess you can just go to Chatuchipati,

you can be like, tell me about the high level architecture of the COBOL, and then you can go and do that if you want. Well, yeah, and also, all the TypeScript professionals can now maintain those COBOL mainframes because they can be like write it in TypeScript and translate it to Cobalt for me. Yeah. That sounds terrifying. Yeah. I want my bank to just stay on Cobalt for now. Yeah.

Yeah. I don't think we're quite there yet, but I just thinking sort of bigger picture, like what is the future impact of some of these technologies on various industries. And I think the one industry that's leading the charge is to the tip of the spear is I think engineering in terms of what impact this is having on people's, the way they do their job is gonna impact the way that people probably think about hiring as well as like what skill set do you now test for? If I can give someone a problem

and they can literally just type it into a prompt during an interview and get like a decent, you know, decent output, like, you know, you can't obviously can't test that way during an interview. I don't think those are great tests anyway, but that has historically been a way a lot of companies have hired for a very long time. Have you seen like the recent thing

about like the Columbia student who like, he he like he took like his live stream of this video he just like got this like I don't know like this like transparent looking thing that like sits over his like coding interview and he would just use that and he used it in a coding interview and he passes like Amazon, Google, XYZ, ABC like interviews and he just like was like I'm not going to take in these offers and he has like a 30 he's like he's like 30,000 MR and like literally like, I don't know, like a month

or something like that. I was like, what? What the heck? That's that's amazing. I didn't I didn't see that one. That's that's funny. Yeah. I mean, I think it was similar to if you're in your teacher, and you're used to the the thing that you're giving a student is a book report well a Book reports not a great test of someone's You know comprehension of a book at this point because you can literally generate a book report in a single prompt and have it So you need to figure out new ways of essentially, you know testing skills

And I think it also evolves like evolves what are the core skills that you need to be successful in a job. It's changed a lot. There was a time where writing cursive was maybe a critical skill, or knowing the Dewey Decimal system, or these types of things as well. And those skills haven't completely gone away, but you need to essentially develop

new types of skills that are relevant to the workspace. I've never written a book report in my life. Wow. I've never read a book for a book report in my life. I grew up in the age where I had the internet, and so I was just, you see internet. I really want my kid's school to have like a, an LLM class where it's like,

I don't know how to explain it, but basically I want the teacher to give them like kind of wacky assignments, like things that they don't have any context for and just be like, you know, and buy, get them a chat GBT subscription or something like that, and just like go do something and tell them,

like spend two hours on this and give me something at the end and also send me the chat that you did for it. And like sort of like get them used to like using these things to like make like, basically I wanna say like, here's an Excel spreadsheet and I want you to find out these like 10 answers from me. And like none of the kids have probably used Excel or something like that,

but like they could go to the chat GBT and be like, hey, I need to figure out. I have this spreadsheet. Like, what do you need from me to figure this stuff out? I think it'd be cool. Yeah. Yeah. And I think that's that's that's really interesting. Like that test, you know, someone's ability to like learn and use these tools as a learning tool, which is is probably a more important skill set now. Because I also think like even as amazing as some of these coding

tools are or other types of tools that are out there, you still need enough sort of technical context to direct them in a way that's going to give you a good result. I don't think they're at a place where you could give this to the hands of my mom and my mom would be able to build a mobile app like you're doing right now. Exactly, yeah. That's totally true. But in terms of just so many things that you should be able to just,

you might not think you could do it, but I think my 12-year-old daughter should be able to get answers from an Excel spreadsheet, even though she's never used Excel before, with an hour of chat to be honest. I think that's totally doable. And I think the big thing is I want them to learn how to use these LLMs and also just train their agency a little bit of just

if you don't know how to do something, you have something that can really help you figure it out, at least to a medium level. It's not going to make you expert on any of this stuff. Yeah, absolutely. Some decent results from that sort of stuff. Cool. All right, well, so the last thing I wanted to chat about

was there's also been a ton of recent acquisitions that have happened. Voyage AI was acquired by MongoDB, data stacks by IBM, weights and biases by CoreWeave. I'm sure there's a ton of other ones that are missing here. And given, Eugene, your background working for Zillow, I'm kind of curious, what are your thoughts on the acquisition by Voyage AI? It seems like a kind of trend where a lot of databases are going to continue to increase value for their platforms. It's like, go beyond just a database to building sort of like a data platform.

So we're going to offer from an AI perspective, we're going to have embedding models built directly into the platform, we're going to have, you know, chunking all these other things that, you know, if you weren't using them, you'd have to go to some other third party service or write yourself book to get that stuff ready before you sort of ingested in the database. Yeah. Okay, so this is what I was gonna say earlier. So, Frank, Frank Liu, Frank Liu and I worked together at Azales

and he went to Voyage.ai last year in September. And I started hearing rumors, not from Frank, from other people. I started hearing rumors that Voyage.ai was going to get acquired in like December, January, something like that. And I was talking to one of my friends about this and I was like, so did you hear about this VoidJI thing?

And they were like, yeah, I heard VoidJI was getting acquired. But at this point, I'm thinking like, if you're one of these companies that raised in 22, that raised in maybe not 22, maybe 23, but like during that like kind of, no, 22. Yeah, at the peak, like there were still many companies that raised $100 million at 100 million dollars at two billion dollar

valuation with zero revenue. Yeah, yeah, yeah, exactly. And my friend was saying he was like, hey, like, you know, like, I'm thinking if you're one of these companies, you're either an acquisition target or you're dead. And I was like, yeah, I mean, that's that's fair enough. And then I saw Frank in. I don't know if it's early February, January,

something like that. We went out, we went out for drinks. And he was like, he was like, yeah, I've heard those rumors too. I was like, oh, okay. But like, like, like, Zillis, for example, like, you know, the CEO of Zillis is he's, he doesn't want to do an acquisition like he's like, he's like IPO or bust. But Tang you the guy who does voice. Yeah, I obviously is open to an acquisition and they got acquired. And I remember hearing about this like the week before it happened. He was like, so you know how like we're talking about like coming on board to like sponsor a Seattle Star Wars Summit and your AI production meeting conference. And I was like, yeah. And he was like, well, we wanted to do it, but we're actually doing something else right now.

And you'll see an announcement on Monday. And I was like, OK, OK. And then he actually told me literally the Sunday or the Saturday before. So I was like, I knew a couple of days before. And he was like, just don't tell anyone. I was like, all right, I'm going to keep my mouth shut. But now that it's happened, I'm like,

I'm sure I can talk about it now. But yeah, he was like, yeah, MongoDB, good acquisition. Good for them in terms of I heard they got very good valuation out of it. I, I personal opinion, in vector database space, MongoDB has a lot of work to do. has a lot of work to do. Yeah. Yeah, I mean, I think that, similarly to the model companies that we were talking about earlier, where they're looking for ways to increase their value beyond just being a foundation model, I think databases

are doing some more moves. How do we move up the value chain and own more of the pipeline natively? You see the same Databricks, Snowflake have continued to essentially layer on more and more functionality. They're way beyond just like, hey, there's a data store too, like this entire platform where you can, you know, do all kinds of different things there. I think replacing a lot of the ETL layer for traditional, you know, data infrastructure, as well as trying to own a lot of the AI space of,

you come to us and do everything essentially. Yeah. Also, DataStacks literally just recently signed on to come on board as a sponsor for Seattle Star of Summit. I saw they got bought by IBM and I was like, that deals dead in the water then, that's gone. But then I got an email and they were like, yeah, we're in. And I was like, oh, okay, great, let's do it. I think the HashiCorp IBM deal is official now too as well.

Oh yeah, yeah, that was huge, dude. That was like 7 billion. Did they announce how big the data stacks one was? I don't think I saw it, but maybe. Data stacks has been around for a long time. I think they're over a decade old at this point. Yeah, Astro. Yeah, Astro DB. Yeah, Astro. Data stacks also

bought like some company like in like the last like two or three years. Yeah, forgot which one it was. But guy Ben, Ben chambers was one like the CTO of that company. I can't remember what it was. So yeah, the CTO of that company. I can't remember what the company was though. Yeah. Data stacks also, vector database functionality. Yeah. Yeah. So IBM is- They're making a lot of moves.

It's very interesting whenever I think about IBM, because it's just like big giant old company. It's like, what is this company doing? I think our generation doesn't think about them. We'd never think they'd buy something from IBM, but yet they're buying all these other companies. Well, I mean, they're such a huge conglomerate. They own a red hat.

I mean, there's so much stuff that's under the umbrella of IBM at this point. They're called the international Machine, you know? That's what they are. That's like information superhighway, but from like 1920. What a name. Yeah, that's amazing. And then what are your thoughts on weights and biases, CoreWeave?

CoreWeave, I'm not super familiar with, but you know, they're an AI infrastructure company, you know, focused on sort of the GPU layer in the cloud. Yeah, that one really surprised me. I thought, like, okay, so I'm actually not super familiar with either of these companies, but I thought Weights and Biases has been around for longer. They have. They've been around for over a decade, I think, because they were originally traditional sort of ML ops. I interviewed one of the founders of Weights and Biases at one point.

So they, I mean, they pre-date everything going on in general AI and then of course they added a lot of functionality on top of that, but their original core customer was to serve sort of data scientists that were building models. How do you do versioning? How do you set up your training runs, all that kind of stuff? And I always, from an outsider's perspective,

felt like they had a very good business, but they, obviously they see some sort of opportunity with CoreWeave. Yeah, that one really surprised me because CoreWeave's a little bit newer. It's not that big. I mean, it is pretty big. They've they've raised a lot of money. Well, I think to do any cloud GPU infrastructure company, you have to raise a lot of money.

Yeah, you got to raise a ton of money. But yeah, I was I was really surprised. I just I don't know. But but it says they spent 1.7 billion on whites and biases. Like how much did they raise they have 1.7 billion? DC money. Yeah Do they have this investors they have they might have similar investors It might have been like a deal that some of the investors like helped like put together. Yeah, what are your thoughts on? these kind of like specialized cloud infrastructure for AI workloads versus the public cloud? Like, you know, I can do you think that because there's more and more of these companies that are emerging.

Do you think that we're really going to have a dozen different places where you're going to go run your AI workloads independent of where you might be running your application stack today? Yeah, so I've thought a lot about this because I talked a lot with the GMI Cloud people. And I also talked a lot with this company called Amara, which says data centers. So I think the thing between these new Neo Clouds or AI-native GPU, whatever, is that unlike the CPU marketplaces, like AWS,

for example, they're basically just running their own data centers. AWS has these CPUs. They basically just have a data center where you walk in and you have these racks, and these racks just full of GPUs. And there's different cooling that you need for that. There's different like kilowatt edge that you need for that. There's like different, like, um, there's different requirements that you need to run one of those data centers. So I think it's like, it's like the, the physical requirements are different enough from like your classic cloud compute stuff that they will exist. I don't think we're gonna have, I think there's currently 45 of these.

I don't think we're gonna have 45. We're gonna have some. I think the two kind of like, the two plays that I see that are gonna be most likely to be sticking around for a while are the ones that own their own data warehouses and the ones that are basically like a marketplace for GPUs. I think those are the two kinds of companies that these are,

I kind of like, I see these as the two kind of extremes. And I think those are the two that are going to stay around. I think we're going to have maybe like three or four of each. Interesting. Are most of those specialized data centers, this is my ignorance on it, are they mostly used for training or are people doing inference from these specialized places as well?

I think, I think they're doing both right now. My intuition on this is that the training is actually what's making them a ton of reliable revenue. Then the inference is actually what's mostly like it's probably used more, but it's not as reliable. I don't know how that's going to play out but it's not as reliable. And I don't know how that's gonna play out in terms of the business model. It's probably just some sort of uncertainty

they're gonna account for. But that's one of those things that I think is not very clear. And there might even be companies that are particularly... So for example, GMI Cloud, they provide training. That's their current model. But they're building an inference platform. And they want you to go and use the inference. So it's just not very clear yet.

You think the TAM on the total addressable market for inference is much, much bigger because most companies are going to be doing it. But then you have players like Base 10, Together AI, and others that are focused on inference and optimization. We've had Base 10 on, Together AI is coming on soon. Those companies are interested, but they're focused,

specializing essentially in multi-tenant inference and just optimizing the heck out of it. Within public Cloud? Those are more within public cloud, for those ones, those are more within public cloud, right? I don't know. Is they? Like rather than build their own data center, I imagine? I'm not 100% sure.

I think they probably work with other providers, because I know that Base 10 and GMI Cloud are doing a dinner on Wednesday for GCC in like two weeks. So, okay. Yeah. It's definitely interesting times. Anything else that other you wanted to chat through? I think that's all I have.

Yeah. Yeah, that's all I got. All right. Awesome. Well, Yujin, thanks for coming back. Excited about the, you know, what's going about everything you got going on in Seattle. Hopefully, you're able to accomplish your goal of turning Seattle into the startup hub. For those that are listening,

check out the Seattle Startup Summit at the end of March. Now, it's good to see you as always. Yeah, for sure. I'm going to check out the recording, so make sure you put those on for me later on. Even if it's not live streamed, I'll catch those recordings later. Yeah, 100 percent. I'm gonna check out the recording. So make sure you put those on for me later on. Even you know, even if it's not live streamed, I'll catch up. I'll catch those recordings later. Yeah, 100 percent. I'll get them. Yeah. Thanks, Sean. Thanks, Alex.

Yeah, thanks for coming on. All right. Cheers.