Investology: Investment Management Intelligence
00:15:30 · Andrea Carnelli Dompé, Andrea Carnelli Dompé
[00:15:30] S1: So we kind of think of agents as being software systems that are smart enough to think for themselves and to make actions with some degree of independence and [00:15:40] S0: autonomy. [00:15:41] S1: And the key words here in this definition are autonomy and taking actions, [00:15:45] S0: making decisions, taking actions. Those are the two key ingredients. [00:15:49] S1: What makes them different from AI workflows or AI co-pilots or the first wave of AI that we saw? [00:15:56] S0: we saw in the first years when everything changed in 2020. So how do we design them then in practice? So I think we follow three principles. One is we hyper-train these agents to the private markets use case. [00:16:11] S1: I think horizontal agents that kind of do everything across different use cases and sectors are not as powerful as hyper-training the system to a very specific use [00:16:23] S0: case, vertical use case, vertical language, logic. So specifically in private markets, they need to understand the jargon. They need to understand the implications. [00:16:32] S1: They need to understand that 95% cold means that you can take the commitment, [00:16:37] S0: multiply by 95% and extract the value of cold. So there's a lot of jargon that they need to understand. [00:16:43] S1: And a lot of rules that they need to understand to be able to look at private [00:16:47] S0: markets data in an effective way. [00:16:49] S1: Just like somebody needs a long training at university and in the industry to
00:16:49 · Andrea Carnelli Dompé, Andrea Carnelli Dompé
[00:16:49] S1: Just like somebody needs a long training at university and in the industry to [00:16:55] S0: actually appreciate what's inside this document. Same thing for agents. So that's one hyper-trained to private markets use case. Secondly, we like to be LLM agnostic. There's a lot of innovation happening in the market. There's a new release every week. [00:17:10] S1: Not all those models are improvements versus the previous versions, first of all. [00:17:14] S0: Not all of those models are good at handling all tasks. [00:17:18] S1: So we have built a system that basically identifies at any point in time the best available LLM for the best use case and kind of integrates them all in a crazy [00:17:28] S0: framework. So that's kind of principle number two. [00:17:32] S1: And principle number three is for mission critical tasks like finance in general, [00:17:37] S0: not just private markets, it's very important to have a very high level of accuracy. [00:17:42] S1: And the way we sort of cope with that is we make sure that we support fully the [00:17:48] S0: human in the loop. for final decisions on critical action items. [00:17:53] S1: So we have thought about the system in a way that it becomes incredibly easy to [00:17:57] S0: audit what the AI has done, why, [00:18:00] S1: and how exactly it has made implications from documents so that it becomes very [00:18:05] S0: easy for a person to click, yeah, okay, I agree with this implication. [00:18:09] S1: And that is a massive component into the productivity gain because I think a system
00:18:09 · Andrea Carnelli Dompé, Andrea Carnelli Dompé, George Aliferis
[00:18:09] S1: And that is a massive component into the productivity gain because I think a system [00:18:13] S0: that where only 99% is accurate, but not easily auditable, wouldn't save time at all, because somebody wouldn't trust it, wouldn't have an easy way to audit it, and would have to kind of do the job from scratch. So having that auditability, massively important. [00:18:29] S2: That answer brings in two reflections. First of all, how different it is [00:18:34] S1: to use data or to have a streamlined use of data in public versus private market. [00:18:40] S2: In public market, this has to some extent been solved. [00:18:44] S1: I mean, there's so much more to be done in terms of data and improving it. [00:18:48] S2: But accessing the data is already very advanced, whereas in private markets, [00:18:53] S1: there's so much to be done there and therefore so much differentiating factors for [00:18:59] S2: those that play this game. The second one is, I guess, [00:19:03] S1: more about building startups in general in this age where AI is everywhere. And I can totally get why being LLM agnostic is a good way to play this game because there's so many resources that are being built and are available pretty [00:19:23] S2: freely. [00:19:24] S0: Absolutely. [00:19:24] S1: And the way we approached it was by basically building a massive backtesting framework, in fact. So we have this beautiful proprietary data set of documents coming from different [00:19:34] S0: GPs, different asset classes, different type of data contained.
00:19:34 · Andrea Carnelli Dompé, Andrea Carnelli Dompé, George Aliferis
[00:19:34] S0: GPs, different asset classes, different type of data contained. [00:19:38] S1: And we have the ground truth that we have been able to establish by working with our clients. [00:19:43] S0: Every time a new model comes out, [00:19:45] S1: we can very quickly check if using this model alongside the entire infrastructure [00:19:50] S0: that we have set up, there's an element of interaction with our systems, but tell me, [00:19:58] S1: very quickly if this new model can improve the performance of our systems and tell [00:20:03] S0: me exactly where it improves it, where it does not improve it. And based on that, [00:20:07] S1: we can integrate that new model or totally discard it where it makes more sense in [00:20:12] S0: our architecture. And that's super powerful. Otherwise, every time a new announcement came out, [00:20:17] S1: it could be absolutely in the dark whether that is going to be helpful or not. And that makes me reflect on something that I've seen from the CEO of Bright Data, [00:20:27] S2: and I will not quote him exactly, [00:20:29] S1: saying that the calculation power is always going to increase exponentially, [00:20:35] S2: but you can build data modes. [00:20:39] S1: And I guess by that he means accessing and processing the data is where you can [00:20:45] S2: really make a difference and a sustainable advantage for yourself. [00:20:51] S0: Yeah, well, yes. I think, I mean, at least from a pure technology perspective, [00:20:55] S1: precisely because these systems have to be highly trained domain to be really
00:20:55 · Andrea Carnelli Dompé, Andrea Carnelli Dompé, George Aliferis
[00:20:55] S1: precisely because these systems have to be highly trained domain to be really [00:21:01] S0: driving big efficiency gains. They will do okay if they're not trained. [00:21:06] S1: There's so much to be gained if you train them specifically on the domain that having access to the information in that domain is a moat for you because you'll be able to get a better mousetrap out of that versus somebody who doesn't have access [00:21:19] S0: to that data. [00:21:20] S1: And as you accumulate more data, you're going to have a quicker edge, for sure. [00:21:25] S0: I think eventually that kind of advantage might level out. [00:21:29] S1: But as of today, it's all about gaining that data and optimizing the LLMs on that data. [00:21:36] S2: And I guess it's also a technical perspective. [00:21:38] S1: It reminds me of the time when Uber was pretty much offering free rides and WeWork [00:21:45] S2: were incentivizing people to come in their offices with a lot of perks. I can see as well the evolution. Gmail, not so long ago, was trying to make me pay for the use of Gemini, and now they're trying to force it into my email account. It's a good opportunity that may not last forever. [00:22:03] S1: Every single penny lost in P&L by one of the LLN providers is one percent gained by [00:22:09] S0: the consumers of these LLNs.
00:22:09 · Andrea Carnelli Dompé, George Aliferis, Andrea Carnelli Dompé
[00:22:09] S0: the consumers of these LLNs. [00:22:11] S2: what you're trying to reach and how much accuracy you're trying to reach, are you aiming for 100%? Then it's probably going to be exhausting. Perhaps in what you do, if you're aiming for just 90%, then this 10% means that there's absolutely no efficiency. [00:22:26] S1: Whereas if you reach, let's say 99%, then that's something that's really effective. [00:22:33] S0: Absolutely. [00:22:33] S1: So I think the best way to look at it is in terms of time saved, the person overseeing process. It's a good proxy to think in terms of extraction accuracy, if you like, but it's just a proxy. And there comes a point where 1% of additional improvement doesn't really help the [00:22:52] S0: person because he still needs to check it. So we try to think in terms of time saved. Ultimately, that's what we want to optimize for. [00:23:00] S2: Now I realize, Andrea, we also have to reflect on this democratization of alternatives, which is such a big topic, and investing in private assets. So you're talking to sophisticated people and make it easier for them. But it's also part of this democratization or enlarging the access. Now that you don't need that many resources, that's when you may be able to start in this very complicated world. [00:23:29] S0: I think democratization could be great if done properly. So in principle, if you give more options, more investment options to a wider audience, more people will be able to choose what's best for them. That's in theory. What's happening often, though,