00:00:00 · George Sivulka, Alex Immerman
[00:00:00] S0: A lot of my peers, if they were smart and they were lucky, they were going into financial services jobs. They would go and become an investor. And I realized that the smartest people in the world, incredibly smart kids, were going and doing the stupidest tasks. So many tedious things. A lot [00:00:13] S1: of Repetition. [00:00:15] S0: Boom. Boom. Boom. [00:00:16] S1: Yep. Night after night. [00:00:18] S0: And I saw pain. I noticed in my friends that they they literally hated their lives. And at Stanford, maybe the only other thing that I I took from the entrepreneurship community there was build a company where there's pain. And, you know, my my 22 year old brain said, well, there's a lot of pain here, and there's the most important technology in the world that can solve it. That's where I'm gonna go and build a business before anyone else catches onto it.
00:00:41 · S2, Alex Immerman, George Sivulka
[00:00:41] S2: Thanks for listening to the a sixteen z AI podcast. Today, we're exploring the intersection of professional services and AI agents via conversation between a sixteen z partner, Alex Immerman, and Hebia founder and CEO, George Svalka. Although we're very early on in the development of Magentic workflows, Hebia can safely stake its claim as an early application of them, thanks to its embrace of reasoning models and a user interface that does away with the reliance on chatting. In this discussion, George and Alex explore the rationale behind some of these design decisions. They talk about building products for an AI native user base and dive into the industry changing benefits of AI, especially in areas like financial services, investing, and law, where smart people spend way too much time on tedium and checking boxes. Oh, and George also gives his opinion on DeepSeek and dishes on his favorite AI tools. All that and more after these disclosures. As a reminder, please note that the content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security and is not directed at any investors or potential investors in any a sixteen z fund. For more details, please see a16z.com/disclosures. [00:01:54] S1: George, it's really fun to have you here. I've enjoyed getting to know you the last couple of years. I'm excited for the world at large to hear [00:02:00] S0: more from you. I'm excited to be here.
00:02:00 · George Sivulka, Alex Immerman
[00:02:00] S0: more from you. I'm excited to be here. [00:02:03] S1: Amazing. Well, we're gonna open with a lightning round, get into the heat of AI. Great. Let's rip it. Do you think scaling laws are going to hold? [00:02:14] S0: It's a good question. I think that there's two types of scaling laws. You have scaling laws for training, which is what I think you're referring to. And then more recently, people have started to talk about scaling laws at inference. And I think they're both effectively mathematical properties of the universe. I don't think they're, you know, just an experimental observed thing. You just know as you add more data and you add more compute, you know, these models get better for training, that is. And I think that they will always hold if there's enough data, and that is the question. But I do think that g p d five will be significantly better than g p d four. At the same time, you're starting to see models like kind of o one, o three doing this, like, reasoning at inference and scaling at DeepSeg as well. And this is a technique scaling at inference that was first actually pioneered at Hebia. And we quickly noticed another scaling law where if we ran more models and and basically more compute at inference time, you could get much better results for very complex tasks. And I think that scaling law has already proven to kind of extend the runway of AI, and I think it will continue to hold as a mathematical property in itself.
00:02:14 · George Sivulka, Alex Immerman
[00:02:14] S0: It's a good question. I think that there's two types of scaling laws. You have scaling laws for training, which is what I think you're referring to. And then more recently, people have started to talk about scaling laws at inference. And I think they're both effectively mathematical properties of the universe. I don't think they're, you know, just an experimental observed thing. You just know as you add more data and you add more compute, you know, these models get better for training, that is. And I think that they will always hold if there's enough data, and that is the question. But I do think that g p d five will be significantly better than g p d four. At the same time, you're starting to see models like kind of o one, o three doing this, like, reasoning at inference and scaling at DeepSeg as well. And this is a technique scaling at inference that was first actually pioneered at Hebia. And we quickly noticed another scaling law where if we ran more models and and basically more compute at inference time, you could get much better results for very complex tasks. And I think that scaling law has already proven to kind of extend the runway of AI, and I think it will continue to hold as a mathematical property in itself. [00:03:20] S1: All right. So scaling laws are going to hold. We talked about deep sea briefly there. Is deep sea a nothingburger, or is it a big deal?
00:03:20 · Alex Immerman, George Sivulka
[00:03:20] S1: All right. So scaling laws are going to hold. We talked about deep sea briefly there. Is deep sea a nothingburger, or is it a big deal? [00:03:28] S0: I am a believer that, like, it was more on the nothingburger side of things. I think China has shown time and time again that they're able to take technologies that are invented oftentimes in The United States and make them more efficient. And I think China has also shown time and time again that they're willing to obfuscate the truth or omit certain things when talking about, you know, certain technologies or the science behind a variety of different things. You know, regardless of the truth with deep sea, I generally also think it's really an American technology that's cheaper. And Americans can also make it cheaper, and it was invented here in The United States. And China has not shown that they can actually continue to play ball pushing the frontier of AI. This is not an example. [00:04:09] S1: Yeah. I think the one piece on it was also, like, pushing open source in in The US here and making us rethink potentially things on the regulatory side there.
00:04:09 · Alex Immerman, George Sivulka
[00:04:09] S1: Yeah. I think the one piece on it was also, like, pushing open source in in The US here and making us rethink potentially things on the regulatory side there. [00:04:17] S0: It's a good point. I think open source is important, and I think that as software becomes open source, it stops being a geopolitical strength of just like the closed source nation that that has invented it. But at the same time, you know, America, I believe, is so far ahead, and I I think that it will continue to be so far ahead. And these technologies, you know, every single year are so much exponentially better than the last year that we have the head start. And I think that we'll continue to have that start, and we have the world's best scientists, researchers, and minds thinking about this and working on. [00:04:50] S1: Absolutely. And then we also have the best companies building on the AppLayer who are gonna benefit from these models getting cheaper and cheaper. [00:04:56] S0: 100%, especially if they scale at inference. So [00:04:59] S1: Yeah. Exactly. So next one, what's your favorite AI tool that's not named Hebia? [00:05:04] S0: I think that my favorite tool right now is Deep Research, which was launched yesterday or the day before yesterday. [00:05:10] S1: It's pretty new. [00:05:11] S0: Very new. It has changed the amount that I use AI, and I use AI for everything. I try to automate everything. What what have you