Financial Modeler's Corner
00:16:14 · Peter Fuller, Giles Male
[00:16:14] S1: Our customers are primarily in management consulting and high finance. And the... I think it's... Yeah. That's fair to say that, you know, the the big early share of those customers work in areas related to private market transactions. Right? So that's valuations and modeling practices in in management consultancies. It's financial due diligence, commercial due diligence, as well as the investors and investment bankers who are kind of working on the other side of the deal. Right? These are very, very high stakes, very schleppy as you guys know. Yeah. They're just just a very important work. Right? And those users benefit a lot from a from a product that is better than the kind of basic horizontal products you can you can buy anywhere else. So, yeah, that's where we focus. [00:17:07] S4: Got a question for you. I... And and I thought... Has the terminology that is now, like, you know, in the kind of for most people like harness and orchestration, we just caught up to where teams like you were before? I'm assuming they aren't new words for this. [00:17:25] S1: Yeah. Maybe. Yeah. I I think we've been talking about harnesses we've been talking about harnesses and orchestration for for quite a long time. [00:17:32] S4: But but it is like harness has probably been the word of July and now it's all... I don't know what it would be now, but loops was one recently and it got knows what else. But orchestration and harnesses are definitely buzzwords.
00:17:43 · Peter Fuller
[00:17:43] S1: Yeah. And like good reason. Right? They they are really important. I mean, you know, I think people like... Look, harness is a is a fancy word. It's like the environment that you build around the LLN, the agent, like the tools that you give it to do the work that you care about. But, you know, I think people have made made mistakes there. Right? And I I probably suffered as a as a consequence. I think a lot of people built harnesses that just didn't price in, like, the improvement in the models. Right? So you can build a super opinionated harness where you, like, really put the model on rails and you do... You know, back in 2024 or early twenty five when everyone was doing that, they they their products were performing better. Right? Because, you know, because they had all these specifics baked in. But then the next model generation comes out and you you gotta throw all of that stuff out or or usually teams actually just aren't intellectually honest enough to to throw it out. So mistakes have been made at the at the harness layer. I think we've done a very, very good job on the on the harness. Orchestration, I think, like, it's very important to be model independent. Right? That's another thing I was gonna say when you asked about the... Particularly about the big labs is, like, there's a lot of nervousness in our customer base about, like, betting the house on Anthropic, like betting your entire organization just on Anthropic models when OpenAI could overtake them, like, next month. And, you know, maybe even Gemini one day. Who knows? Right? But, like, I just... A lot of our customers just don't think that's a good idea. And I agree with them. Right? I don't think you need to take that risk. I think you need to, you know, they they buy us because they bought... They wanna buy the the vertically focused products that is built specifically for their needs. That's just gonna be able to send them to the best model for that... Any given task. So, yeah, those things are very, very important.
00:19:47 · Paul Barnhurst, Peter Fuller
[00:19:47] S2: I have a couple questions and I'll let Anand Giles go again. The first, when you mentioned, you know, the models, I'd love to get your thought on... Are they gonna all kind of become a commodity at some point? The models will just all be pretty similar. Where do you think we're heading? Because it feels like outside of customized special, the big huge ones, how much can you differentiate yourself in a model when you're all training off similar data? So I just love a little bit of your thought there. You're already smiling, so I know you have an opinion. [00:20:17] S1: I'm I'm not gonna claim to know the answer. I don't think anyone knows the answer to that question. [00:20:20] S2: Sure. I just love your thoughts. [00:20:22] S1: I mean Yeah. [00:20:23] S2: Yeah. If you're right, we'll come back and congratulate you. And if you're wrong, you'll be like most of
00:20:26 · Peter Fuller
[00:20:26] S1: us. Yeah. Yeah. Well, look. I'll take a couple of things. The commodity thing, I mean, it's kind of baked into their mission statements. Right? Like, OpenAI always talk about intelligence too cheap to meter. And I I read that and I think, is that good for your business model? Like... But they're always talking about that. Right? Well, that means it's a commodity. Right? So I think I think there is there is gonna be a lot of that and people will make a lot of money selling the compliments to that commodity. Right? Our customers included. It's already happening, by the way. Right? As as the power of the models grows, this intelligence grows, our customers, like our consulting customers, I think BCG grew AI transformation related revenue at, like, 25% last year, something crazy like that. And they're just making so much money implementing AI in, you know, helping customers implement AI. So you're already seeing that effect whether... With the compliments are doing really, really well. The other thing is, I mean, I I I really I really don't know. There's been, like, architecturally, there's just been so little diversity in the frontier models. I mean, everyone's kind of pursuing the same thing, right, and scaling it. That's that's been the winning play for the... That's how many years. I don't know. It's perfectly possible that startup or someone like Ilya releases a model that is architecturally different and has very different capabilities, maybe different spikes and troughs versus the existing models. And that would be amazing for companies like us. Right? Diversity at the model layer increases the value of orchestration between those models. Right?
00:22:23 · Giles Male, Peter Fuller, Ian Schnoor, Paul Barnhurst
[00:22:23] S4: And Because you... Because your users can choose between them. [00:22:27] S1: So easy. I mean, it would be so... Let's say, Ilya releases this, like, fantastic model, which is just very different from from Anthropic models. And then I... You know, customers would be able to use Tracelight, and we would use both the models for different tasks. Right? And so the value of orchestration would grow. But people have been thinking that for a long time and it hasn't happened and it might not happen. Right? It might just be scaling all the way. I don't know. [00:22:55] S4: And just to delve on that, because the way you just used the word orchestration, I'd never thought of. Because to me, I was thinking sort of within, I don't know, like a cloud environment, you have a sub agent team of agents. But you're actually talking about the orchestration of potentially the use of different LLMs [00:23:15] S3: We already know. [00:23:15] S4: Within this... Yeah. Okay. Oh, wow. Okay. [00:23:17] S2: Well, think about... If I can say one thing, I think, Giles, when you think about it, you're selecting which model to use within an LLM. They're just selecting which model to use of all LLMs. [00:23:30] S4: But but but but also by the sounds of it, but this is what the... I wanna just make sure I understand. Orchestration for me would be like within a a a a workflow, you're actually delving within different LLMs from different providers within a workflow.
00:23:30 · Giles Male, Peter Fuller
[00:23:30] S4: But but but but also by the sounds of it, but this is what the... I wanna just make sure I understand. Orchestration for me would be like within a a a a workflow, you're actually delving within different LLMs from different providers within a workflow. [00:23:44] S1: Yeah. So for example, if you upload 500 pages of PDFs into Tracelight and say, here's contracts, read contracts, transcribe, and build a model. Right? It would be Gemini reading the... Reading those PDFs, and then it would be today, Claude, doing the modeling for you. Right? So it already happens in our in our product. And, you know, on our kind of pro tier as well and our max tier, you can choose between using the kind of frontier intelligence. Right? Or we have efficient mode as well where we'll use, yeah, like a smaller cheaper model. Right? But people get way more usage out of their out of their subscription. And today, those are from different providers. [00:24:36] S4: Do you have a strong view on... And again, this has come up in a couple of places recently. Do you have a strong view on, you know, choosing the most powerful model out there at the moment versus having a really strong harness and set up, however that is kind of implemented. Do you get a sense of like what is better if that's fair?