Proof of Work: AI Value Creation
00:30:25 · Joe Zein
[00:30:25] S2: In an ideal world, too, right? You have different tools that a specific fund manager is using, right? They're using the portfolio monitoring system, they're using their CRM, they're using all sorts of different tools, right? And you want to go with the, I'm going with an ideal world that is still not completely ideal, where they don't have files, everything is living in some sort of system, right? So you would think that, okay, we have the model, but in reality you don't. You have parts of the model, you know, some subsets of models living each in their own SaaS. And the exercise is so, well, how can I, and this is the whole, you know, like 2020 data warehouse, bringing everything into one place and joining them. How do I make sure that this entity that is in 73 strings is the same as that entity that I have in uh Salesforce? And this is the whole practice of MDM or data management that has then been rebranded as analytics engineer, that is now being rebranded as AI infrastructure, of basically matching those records across different systems and putting them into one place. Because at the end of the day, 73 strings is looking at your data from the lens of the portfolio. Something like Salesforce is looking at your data from the lens of like BD and underwriting, right? And so you want to bring those lenses into one lens, which is the representation of how you think about your own company.
00:31:44 · Stuart Willson, Yann Magnan
[00:31:44] S0: Yeah, I'm glad you went there because I was going to ask Yann as a next question. How does 73 strings handle the entity resolution problem when the same company has three different names or four different names across three or four different source systems? How do you go about unifying that? And one could what can people learn from that approach? [00:32:06] S1: So uh I think we used to do that when we were uh back uh back a couple of years ago. Uh most of that uh issue was basically handled uh at uh the uh implementation point of the platform when we are implementing the platform. Well, basically, again, you you have to develop anyway that uh consistent data model across the board. Uh so one of the work that you do is working with the with the client, identifying all these types of issues, and so identifying the uh correct uh legal entities were uh one of this uh topic. What we're seeing happening is that actually that's typically the area where LLMs can help because they can make this type of connections. So you can accelerate that work now with with LLMs. Again, as long as you have put the data in the relevant place. So that's what we're seeing happening.
00:32:06 · Yann Magnan, Stuart Willson
[00:32:06] S1: So uh I think we used to do that when we were uh back uh back a couple of years ago. Uh most of that uh issue was basically handled uh at uh the uh implementation point of the platform when we are implementing the platform. Well, basically, again, you you have to develop anyway that uh consistent data model across the board. Uh so one of the work that you do is working with the with the client, identifying all these types of issues, and so identifying the uh correct uh legal entities were uh one of this uh topic. What we're seeing happening is that actually that's typically the area where LLMs can help because they can make this type of connections. So you can accelerate that work now with with LLMs. Again, as long as you have put the data in the relevant place. So that's what we're seeing happening. [00:32:50] S0: Great. I'm gonna take two questions from the audience because they are relevant to this conversation. One is from Johnny. Uh he asks, given solid data infrastructure and a rag-based retrieval lever, how how sophisticated does the underlying LLM actually need to be? Is there a case for running lighter, more efficient models like smaller open source options, Gemma 4, or do the complexity of financial signals and edge cases demand frontier models?
00:32:50 · Stuart Willson, Yann Magnan
[00:32:50] S0: Great. I'm gonna take two questions from the audience because they are relevant to this conversation. One is from Johnny. Uh he asks, given solid data infrastructure and a rag-based retrieval lever, how how sophisticated does the underlying LLM actually need to be? Is there a case for running lighter, more efficient models like smaller open source options, Gemma 4, or do the complexity of financial signals and edge cases demand frontier models? [00:33:18] S1: I can give a give it a start, and then Joe, I'm sure you have uh you're you're you'll be more in the wheels on this one. But my view is that um as long again, getting back to I've said that maybe three, four times now, but uh as long as you get the data consistently structured, consistently labeled, standardized across the board, then that's a very, very strong uh starting point. And then there are lighter models, smaller models out there uh that can help you extract the first layer of intelligence at a lower cost, 100% certain. Um that being said, when you want to do something more sophisticated, you know, that's where you know you have uh you have to you have potentially have to uh to to get into the um into the LLMs.
00:34:04 · Joe Zein, Yann Magnan
[00:34:04] S2: Yeah, I completely agree with Yann here. I think there is easier you make it on the models, the more you can go, I don't want to say the on-market, but you can go on uh towards lighter models. That being said, if you're trying to run some crazy task that is that needs to scour the internet, even if you have amazing data, right, you won't have all of the world's information into that into your data model. And so this is where I think you know you would still need to, it really depends on the task at the end of the day, but there's definitely we've definitely been seeing people and interest about like, I know exactly what task is going to be running on this model. I don't need to have a sophisticated model that allows me to catch all sorts of edge cases. But the the the premise of going that way and using and optimizing how you you choose your models needs and requires uh uh a data foundation. [00:34:54] S1: Just uh an example real quick on that font on our side. So our ingestion technology actually is a combination of uh machine learning, NLP, and LLM type of algos. And if we would be using LLMs across the board, one, I'm not sure that we would get uh an accuracy as good as what we have, actually. I'm quite sure of the opposite. And second, the cost would be just crazy. So there are areas where PML and NLP algos that we have developed do a much better job at a much lower cost than bringing the LLMC.
00:34:54 · Yann Magnan, Joe Zein, Stuart Willson
[00:34:54] S1: Just uh an example real quick on that font on our side. So our ingestion technology actually is a combination of uh machine learning, NLP, and LLM type of algos. And if we would be using LLMs across the board, one, I'm not sure that we would get uh an accuracy as good as what we have, actually. I'm quite sure of the opposite. And second, the cost would be just crazy. So there are areas where PML and NLP algos that we have developed do a much better job at a much lower cost than bringing the LLMC. [00:35:25] S2: Yep, there's techniques that have been there to parse receipts and to do stuff like that, like OCR. That this is one of the nice things of like someone who's like 73 strings who started before that whole AI craze is that they had to go the hard way about doing the extraction. And then, you know, they they layered in LLMs, so but they have a foundation that is pretty deterministic compared to an LLM that allows them to do the extraction at a lower cost on and but and and have it more accurate too. [00:35:55] S0: Great. Another LLM question from the audience, a very common question I get from organizations of all sorts from Sean. What about the sensitivity around confidentiality? If we put the information in LLM, won't the LLM be able to use that information for queries from other companies?