The Innovation Civilization Podcast
00:07:21 · Waheed Nabeel, George Sivulka
[00:07:21] S0: Okay. [00:07:22] S1: But when you actually think about what happens with serious work, it's not just kicking off a prompt and then going into a direction. When you have an employee, you're collaborating with them. They're creating something. You're editing something. You're working in the same workbook or materials to come to a result. And what Hebia is all about is this idea of collaborating, of working with a variety of different... Of working with a variety of different agents so that humans and agents can coexist. And so the workbook, the matrix, basically everything that we have produced are different ways of making multistep or multi hop tasks easily understood and handleable by AI. If chat can do a single step or a single task, really complex things like investment research or drug discovery or even, like, looking for geological sites to build new mines, these are all different things that require lots of steps and process and iteration that have lots of back and forth with humans today, and we're building an interface where that back and forth can happen with humans and an agent select. [00:08:28] S0: Okay. That's pretty cool, Ashley. And is the back end actually an LLM? Is that your core technology you're using?
00:08:28 · Waheed Nabeel, George Sivulka
[00:08:28] S0: Okay. That's pretty cool, Ashley. And is the back end actually an LLM? Is that your core technology you're using? [00:08:35] S1: One of the core steps that an agent can take Okay. Is an LLM inference. Okay. But you can use existing tools too. And so as AGI approaches, we'll have LLMs or foundation models that are really good at processing text, but also ones that are really good at using tools. And we'll judge the efficacy of an AGI not by its ability to brute force things, but by its ability to use tools, including tools like Hebia. And so LLMs are fundamental tools that our agents can use, and the agents are comprised of foundation models. Yeah. Yeah. But there's a lot more richness to what they can do. [00:09:11] S0: 100%. Are these your own models, or are you using the off the shelf models from OpenAI and some of the other folks, Clot? [00:09:19] S1: These are almost always whatever model our users would like. [00:09:23] S0: So we [00:09:24] S1: built Hebia from the ground up to be model agnostic [00:09:27] S0: I see. [00:09:27] S1: Which means we triage some of the most large language model calls to OpenAI, Anthropic, Google, and others.
00:09:27 · George Sivulka, Waheed Nabeel
[00:09:27] S1: Which means we triage some of the most large language model calls to OpenAI, Anthropic, Google, and others. [00:09:33] S0: That makes sense really. On that note, so one of the kind of interesting things when I'm trying to do important tasks, let's say, with AI is, like, early on with, like, early chat GBT and others, Claude as well, it just give you the wrong answers, especially when it comes to math. Right? And if you just look, it just tremendously hallucinates. How do you guys prevent hallucination? Do you have some sort of data that how accurate your outputs are actually being, especially when it comes to these important things like you're talking about investment research and private equity or whatever, basically? [00:10:07] S1: We've ensured that every single model, when it creates an answer, cites the answer from text. So it can't create something, or it shouldn't create something that doesn't have a step even before the creation of something that does retrieval. So everything that... And is an answer that you'll get from Heavier's platform is corroborated by a highlight even before it's generated. So it's a bit of a Copernican inversion. Rather than generate text and then try to find the citations for it, We find the citations first and then generate the text, and that allows us to have a much higher accuracy [00:10:42] S0: Are you in [00:10:43] S1: terms of generating and knowing that we're not hallucinating. Yeah. [00:10:46] S0: Are you trading off, like, creativity for that... In that case? Or
00:10:46 · Waheed Nabeel, George Sivulka
[00:10:46] S0: Are you trading off, like, creativity for that... In that case? Or [00:10:51] S1: We do. We... So we leave creativity of the humans, and we're trying to build a platform that allows humans or the AI copilot to to create whatever output they'd like. What that also means, though, is when we actually go out and give people the software, yes, it's gonna be almost a 100% accurate, as close as we can get it to being a 100% accurate. But that means it'll be less creative. It'll be less suggestive. And in order to find the right investment highlights, you'll have to drive it rather than trusting it blindly. [00:11:21] S0: Okay. That's quite interesting, really. So would you say you guys have cracked an iteration of the hallucination problem in that case, or is this a different... Okay. [00:11:32] S1: So I'd say that a lot of different research and capabilities are all around are removing hallucinations. Sometimes by making a bigger model, sometimes by doing this, cite your sources even before you generate, which is what... One of the things that Hebio does, and sometimes by using advanced tool use. Or we're teaching LLMs, hey. If you're bad at math, use a calculator, and you could get a 100% accuracy every time. And as we move in those directions, what you'll actually will start to see is that hallucinations are going to be a very minimal part of the zeitgeist of of AI talk talk tracks and news. [00:12:06] S0: Yeah. That's quite awesome, actually. Now I wanna use it. But you guys don't have a consumer product in mind. Right? B to B right now, mostly?
00:12:06 · Waheed Nabeel, George Sivulka
[00:12:06] S0: Yeah. That's quite awesome, actually. Now I wanna use it. But you guys don't have a consumer product in mind. Right? B to B right now, mostly? [00:12:14] S1: Right now, I have a framework where a lot of consumer products always map to, like, flaws in human psychology. Yeah. There's a time honored one where consumer products have to map to the seven deadly sins. Are they [00:12:26] S0: just Yeah. [00:12:27] S1: B to B, we have the added benefit of being able to create things that that ostensibly should be providing value or productivity, or you can appeal to knowledge, discovery, creation rather than laziness or rather than gluttony or Yeah. A variety of other classic B to B deadly sins that people map companies into. So, yes, we currently are starting b to b. Yeah. Long long term, I'd like to create a piece of software that, like the Microsoft productivity suite, goes out to everyone. I think that b to b affords you the ability to work on an important problem that is more on productivity than maybe virality or echo chambers, all of the lower pieces of human... Humanity's relationship with technology.
00:12:27 · George Sivulka, Waheed Nabeel
[00:12:27] S1: B to B, we have the added benefit of being able to create things that that ostensibly should be providing value or productivity, or you can appeal to knowledge, discovery, creation rather than laziness or rather than gluttony or Yeah. A variety of other classic B to B deadly sins that people map companies into. So, yes, we currently are starting b to b. Yeah. Long long term, I'd like to create a piece of software that, like the Microsoft productivity suite, goes out to everyone. I think that b to b affords you the ability to work on an important problem that is more on productivity than maybe virality or echo chambers, all of the lower pieces of human... Humanity's relationship with technology. [00:13:09] S0: Yeah. That's quite cool, actually. So tell me... Paint me the picture basically that you see. So you mentioned about this multi agent future basically. Right? Where you as a human are is is basically like multiple bicycles for the mind to put in Steve Jobs' words basically. Right? Because there's so many of different workers you've got. So paint me the picture of like in X years time, whether it be five or ten of how this sort of world looks like. And yeah, what sort of tooling are we using and how and... Yeah. Just have you thought about what this kind of multi agent sort of future looks like, basically? [00:13:49] S1: Yeah. I I think one of my fun takes is that just like the Internet took work, which was all in person and turned work into something that could be hybrid or remote [00:13:59] S0: Yeah.