AI with Alec. Get smarter on AI. The easy way.
00:05:14 · Alec Coughlin
[00:05:14] S0: I love that. And let me play something back to you. You you tell me if I'm I'm picking up what you just put down there properly. So if I'm one of 15 managing directors and each of us have our own slightly different style and some slightly difference might be more than slightly different. Am I to understand that that's where Robo's gonna learn rather quickly all those unique characteristics and the way I like to tell stories, the type of data I prioritize, you know, that that that whole kit and caboodle is that what we're what we're getting into here is that it learns rather quickly and enables that type of almost personalization or customization? I mean, think about it this way.
00:05:56 · Tumas Rackaitis
[00:05:56] S1: And, you know, you... Whenever you try making AI systems, you kinda wanna model how people would do it normally. Right? And the normal way people make PowerPoints on Wall Street is they find precedent decks. And, you know, anything... If you build a good AI system, it's gonna learn from examples pretty quickly. Right? Yep. In fact, it's it's a pretty bad AI system, it does. And the other thing is that, you know, traditionally, people have tried making PowerPoint automation tools with UiPath or all... A bunch of rule based systems, and they never really worked. And the reason why is that this new modality of AI allows you to be a lot more flexible. When little things don't, you know, quite work, it still somewhat figures it out. You know, different types of intentions, finding the right logo happened to be, like, something really interesting we started working on. Anthropic is a lot like anthropology. And finding the right logo is honestly like a ton of work in finding a PowerPoint, especially if you have, like, a bunch of logos on a PowerPoint. That can save you hours of time, and we'll just be able to find them. We have a lot of good computer vision verification. Going back to the company knowledge graph. Right? What do we store? We store really good PowerPoint friendly logos with every single company we identify. Right? So it's this really good powerful tool that we built on top of it.
00:07:05 · Alec Coughlin
[00:07:05] S0: I love it. I love it. So another line that that really caught my attention inside your recent fundraising announcement and congrats on that. It's awesome. Really smart and amazing investors you guys have. Was this idea of of... I'm gonna I'm gonna look at my notes real quick as I butcher this. Bankers need more than just a chatbot. Right? So if you're talking to... And there's a lot of them in my, you know, network and followers and what have you, as we talked about previously, the the finance folks in my network that are really smart and and they've done great things but they're not technical. Right? So they might think chatbot is a chatbot. Right? And and so on and so forth. So how do you help them better understand what you mean by the idea that that bankers don't need another chatbot? So maybe what is a chatbot in the general sense? And then how are you guys different? And I know that but I'm gonna lay it up for you. Right? And then kinda go down that path a little bit. Because I think there's some people in my network that would love to hear the answer to that question.
00:08:05 · Tumas Rackaitis
[00:08:05] S1: Well, I think the way to think about it is you don't want to have a chatbot for every single point application. I don't wanna have a chatbot from, you know, an Excel plug in I have. I don't wanna then go to my expert interview platform and then have a chatbot there. I wanna have one generalized place that can hook in with my private data and connect it all of my private data to anything public, any type of acquisition, any type of, you know, knowledge that you have publicly about a company, whether it's private or actually traded publicly, and to have that be a centralized platform there. So what we do is we take in any type of data your firm has. Oftentimes, it's unstructured or it's structured. And what we'll do is we'll link it to our knowledge Right? And that's like... It sort of builds on top of each other. And what that does is this allows us to, you know, be able to be an AI interface for all sorts of different parts of your application. One of the ways that we have users interact with it is is our chat application. But there is another thing. I think this actually requires us to explain our stack a little bit. Right? Yep. Our stack is basically four parts. Right? Our tech stack. Bottom layer is the data stack. Right? That's part of our... Us linking, crawling, you know, doing a lot... Building, like, a really, really powerful financial search engine. On top of that is actually our search stack, like actually executing and all of the indexing that we do, building the knowledge graph, the AI native security master, and so on. Top of that is our chat stack. Right? Every time you search chat, you know, every time you, you know, you ask a question, we might search for information. This is basically the very common request response type of chat you get from ChatGPT. And on top of that is the actual agential level work. That's where we're actually generating PowerPoints. We're thinking, how do we use the basic knowledge that we treat a rogo chat response as a unit of work? And how can we build other things on top of that to be able to fully automate your your work end to end?
00:09:53 · Alec Coughlin
[00:09:53] S0: That's a killer answer, and I'm so glad that you framed it that way because I think that will help that side of of of the collective audience and followers better understand that bit. So here comes the Peter Thiel question as we talked about. Right? So what's kind of one thing within the world of AI that you believe that perhaps a lot of people wouldn't necessarily agree with you and and why?
00:10:19 · Tumas Rackaitis
[00:10:19] S1: One thing I think a lot of people have, you know, rather interesting notions, and this is what RAG became. The entire idea was I have some data... Some files. What I'm gonna do is I'm gonna run an embedding model on those files. Maybe it's a super fine tuned one I've been working on. Maybe it's not. And what I'll do is I'll take the embeddings and... Which are, you know, vector representations of text, and I'm gonna try to, like, take my question. I might generate a hypothetical document, whatever, and it'll find relevant documents that are like it this way, and then I'll return them. And to to some extent, this works. Right? Like, we have an embedding model. We do embedding search as well. But we found that it's not actually that good. And, fundamentally, the reason why when we looked at it... And fun fact, embeddings are technically reversible if you can solve a convex optimization problem. But, you know, if you take an embedding, it's... It compresses far too much information, ultimately. And the one thing that you can kind of think about doing instead is realizing, well, if, you know, an embedding really is just like the semantic representation of text in a vector, the thing that really understands text better than anything else, the best machine, is honestly the attention layers inside an LLM. So how can you start working about... Around using the internal levels of attention to actually... And using the actual way that these LLMs work and harnessing the actual attention layers to be able to use that as a search space instead for your files. And now what this means is that it's gonna cost you a lot of VRAM on, you know, NVIDIA GPUs, but you can actually coordinate and make a really, really successful searches across, you know, attention space. And that's been incredibly successful for us. So that's been the one way thing that, like, when when people ask, like, what do you think about AI that other people think is totally wrong? I think it's like embedding is just... I don't think are the best way to go.