00:07:49 · Danny Fortson, George Sivulka
[00:07:49] S2: Right. So like a kind of a smart coworker. Exactly. [00:07:53] S4: If if current AI startups and apps are not actually driving real value because they're just doing a single step, that's the worst coworker of all time. [00:08:02] S2: Right. [00:08:02] S4: Even an intern, you give them an arbitrary task and the intern will go in and try to solve that task over and over and over again. They'll execute a bunch of different steps. They could be totally wrong, but you're not gonna have to stand over the intern shoulder and say, okay, now do this. Okay. Now do this. Yeah. Okay. Now do this. [00:08:18] S2: And so are you building your own large language model? Are you building this on top of OpenAI or Anthropic or one of these other ones? Like, what is the actual technology that you're building to actually create something that is a bit more agile and a bit smarter? [00:08:35] S4: So Hebia fundamentally believes that the value is not in the model layer. We think that it'll actually become relatively commoditized. [00:08:44] S2: Agreed.
00:08:44 · Danny Fortson, George Sivulka
[00:08:44] S2: Agreed. [00:08:45] S4: We actually used to train, and our bread and butter was training, and I'm I'm formerly an academic. And there was a moment in 2020 or 2021 where I I realized that because the value wasn't going to be accrued in the model layer, we should totally stop and work on some other secret sauce. And as I thought more and more about the way the industry would evolve, it seemed like, okay. There's this really important technology, but there wasn't a product that was there. And that product would need infrastructure. It would need orchestration. It would need a... An elaborate interface that would allow people to do that more complex work. And so we stopped training models, and that was like a... It tore my heart apart, you know, as a former academic. And we started actually building a product from scratch, which sits on top of any large language model. We're LLM agnostic, and our customers sometimes even use their own large language models in heavy. [00:09:37] S2: Right. [00:09:38] S4: And it orchestrates how those LLM should be called. A really good kind of similar example would be Microsoft Excel. You don't care what the back end powering calculator or computer or database is. You just care about the way that your calculations get orchestrated, the way you interface with them, that it's a container for that work. And similarly, our customers don't care about the l one behind it.
00:09:38 · George Sivulka, Danny Fortson
[00:09:38] S4: And it orchestrates how those LLM should be called. A really good kind of similar example would be Microsoft Excel. You don't care what the back end powering calculator or computer or database is. You just care about the way that your calculations get orchestrated, the way you interface with them, that it's a container for that work. And similarly, our customers don't care about the l one behind it. [00:10:01] S2: If it's, for example, Lama from Meta or Mhmm. OpenAI, you know, ChatGPT, whatever, and take... Choose any of them. They're all slightly different. They all have slightly different capabilities. They all have slight, you know, different weaknesses, different strengths, etcetera. Does that not directly impact what Hebia can do? If you're just saying, okay, we're just gonna put this on top of any LLM. Just give us an LLM. We'll make this work and make it smart and make it really capable as you're saying. But if you're doing it on top of a... One that is not good at one thing or better at another, does that not directly impact what you guys can do? [00:10:40] S4: I think the the question is, does the car drive if the engine is bad? [00:10:45] S2: Yeah.
00:10:45 · Danny Fortson, George Sivulka
[00:10:45] S2: Yeah. [00:10:46] S4: And there are certain engines that are are really good for certain tasks. Heavier, instead of just putting a... An engine in the car and then just throwing you down the... Throwing you up at hell or or or letting the the the LLM dictate everything, says that for a really difficult task, maybe we actually need to run way more large language models. And that all of these are effectively good enough at simple single step tasks. But if I go back to that investment question or this really complex work problem, we actually are orchestrating maybe a GPT four model or a thousand GPT four models to answer one question, or we're orchestrating quite a lot of LLM calls with much smaller, simpler tasks. You can think of it as a car that's driving a thousand engines. And even if the engines are terrible, I can still get up a hill because you're just looping in all of that collective power. [00:11:38] S2: I'm gonna use a different analogy, you could tell me whether it's off base or not. Heavy is kind of like an orchestra conductor where you have a whole group of musicians. Some may be good, some are amazing, but together you just make sure that the music that comes out is fantastic. [00:11:55] S4: That's right. [00:11:56] S2: Right. [00:11:56] S4: And the conductor knows exactly which, musician to point at at which point in time for which song. [00:12:02] S2: Got you. Why call it Hebiah?
00:12:02 · Danny Fortson, George Sivulka
[00:12:02] S2: Got you. Why call it Hebiah? [00:12:06] S4: When I was in grad school, I had... I'd been accepted into... It's PhD at Stanford. It was in maybe the most boring, you know, not outside department in the world. And by outside, I literally mean outside. I did my undergrad in math, my master's in applied physics and then my PhD was accepted in electrical engineering with a focus in computational neuroscience. [00:12:27] S2: What was the plan or did you have a plan? [00:12:30] S4: I all along wanted to be a professor. I did not intend at all to be an entrepreneur. I was solely focused on academic. I loved research. I loved discovering new things, and I loved teaching. Teaching was, like, the highest ideal for me. And, honestly, there's a ton of different teachers that I grew up with that that changed my life. And so I was I was mimicking them. That was the path I was going down. But I get into this PhD and I'm like, well, I'm about to go sit in a room for the next ten years, however long. I've already been in in in the math corner for for quite a long time at Stanford. And I wanna do a fun rotation. I wanna do something fun outside. I sat... I thought I thought to myself, was like, what's the most interesting or, you know, rotation or or research project that I could do now that I'm I'm kind of in this break before really deep grad school? And I wanted to go to Micronesia, And I promise this relates to this relates to the name Hepia. [00:13:25] S2: Question. Question. Why Micronesia? [00:13:29] S4: I love the ocean. [00:13:30] S2: Okay.
00:13:30 · Danny Fortson, George Sivulka
[00:13:30] S2: Okay. [00:13:30] S4: And I thought it was... There's this beautiful place, called Palau in in the coral reef triangle of the world. And I said, well, if I can make up a research project that would let me go to Palau, then I could really have a lot of fun for my first stint in grad school. And I wanted this to be relevant to my research. I wanted to be relevant to computational neuroscience. And everyone talks about AGI today, and there's a lot of different... You know, we're gonna make the models larger that become AGI. There was a whole separate subfield of computational neuroscience around mapping a connectome of animal brains. You might see where I'm going and training artificial neural networks based on like animal neural networks. And so I said, well, I'm gonna go and collect coral reef samples in Palau, and I'll get to be on a boat every day and free dive every day. And, you know, it'll it'll be in grad school. And then I'll train, an artificial neural network to go and mimic that that behavior. [00:14:29] S2: The the coral... The the different bits of the coral reef talking to other, etcetera. This sounds like a real stretch. [00:14:37] S4: I I... Hey. I I was on [00:14:39] S2: a boat. No. I appreciate it. So you actually did it. You pulled it off. I pulled it off.