Invest with AI
00:57:53 · Alistair Smallwood
[00:57:53] S0: Exactly. And that was something that we wrote about in our doing systematic alternative data at scale with large language models is that actually one of the crucial parts of doing any alternative data work is really understanding what the data is measuring. It's really understanding, for example, with credit card data, what the panel, how big the panel is, is there a bias in terms of cards that are on there, or a geographic bias. I remember while I was at the Fund, we were trading a European stock called CTS Aventim, which is kind of the European Ticketmaster, and there was a period where credit card data went through the roof. Basically would have implied on historic correlation that revenue is going to double year on year in this quarter for business that was ordinarily growing at 20%. Shares went absolutely bananas as you'd expect them to. People are looking at the first order read of that alternative data. One of the things that they hadn't realised was that the alternative data panel only had 50,000 people in Europe and Taylor Swift had her tour tickets that went on sale that quarter, and clearly a much higher proportion of those people on the panel had got and purchased Taylor Swift tickets. The way that we worked this out was that if you had back worked number of transactions and ticket prices to the three gigs that Taylor Swift was doing in Germany, which was the particular market that we were looking at, she would have had to have been doing something like 24 gigs in stadiums that were twice the size of what she had already booked. So again, there was that sensibility that you have to try and apply to be like, there aren't this many possible tickets out there for people to buy. And obviously, we were then short and not to take myself, pat myself on the back. However, there were lots of stuff that also went wrong.
00:59:36 · Brett Caughran
[00:59:36] S1: Yeah. Yeah. Really, this is fun. This is fun to sort of talk through this stuff. What has it been like moving from a capital deployment seat to a technology seat and what's it been like scaling Primer in this rapidly changing environment?
00:59:54 · Alistair Smallwood
[00:59:54] S0: Yeah, it's a great question. I think one of the things, if I take it back a bit, I remember moving to the buy side and having been on the sell side and spoken to our managers every day of my eight year career before that. Nobody necessarily tells you how hard it is. There's always a brave face that gets put on by managers when you go to see them. So I remember arriving and realising that you're on this kind of constant anxiety cycle of do I know enough? Am I right? And you're having this kind of tick by tick referendum on every idea that you've in the portfolio as to whether it's right or wrong. And that anxiety is never ending. It doesn't matter whether you're on annual leave or not on annual leave or it's Christmas, it's still going. And I thought that actually, super excited about the opportunity at Primera and ultimately what we can do for kind of leading the frontier of AI and equity research. And I thought that moving out of the risk taking role, that that would disappear. But ultimately the anxiety has just changed. The anxiety now is just we are fighting for our right to exist every two weeks because new models come out, our competitors move, but ultimately are we well known enough as a business and therefore the barriers to scale are really distribution. Our products can be 100% better than anything else out there, but if people are not seeing it and understanding what it does, then we kind of get in nowhere. And therefore the scaling journey has been fascinating, where we launched the product that we serve to customers in April. We've doubled every month since, which is a great position to be in. But ultimately getting to a place where we want to grow with clients as their AI usage matures. So that's not a case of us trying to just take a clip off the top of token usage. It's really around there is a huge spectrum of AI maturity across particularly the buy side at the moment, the sell side is kind of broadly all in the same place. And some people want a turnkey solution that works for them right now and can help them improve and essentially make their existing process a bit faster and more productive, and that's something that we serve to customers, which we're very happy to do.
01:01:56 · Alistair Smallwood
But ultimately, we think that an increasing number of funds, which we've seen of people building at the frontier, will want full sovereignty and ownership over their equity research system, because it really will be a system. It won't just be another tab that we're also used to adding every three or six months, oh, there's another tool out, we'll trial it, we'll have it in a new tab, and we'll forget to use it. Ultimately, these systems are going to be absolutely central to how you do that work. And so we've been super fortunate over the last few months to be working with some of the very large funds on licensing the tools that sit underneath Primer to form the kind of basis of the infrastructure or their day to day research system, which gives them the full flexibility to build and customise on top without having to worry about any of the plumbing that sits underneath it. Because ultimately the plumbing is, as we have found, difficult to do and to do it well, and you want that plumbing to improve consistently as models improve, harnesses improve. The thing that you as a PM or analyst should be focusing on is how can I use this system to discover new methods of activity research, test them, try them and fundamentally do a lot of the stuff that I was already doing faster and on autonomous a basis? And I think that the job will shift from trying to translate what I do now into a system. I think that will happen fairly quickly over the next kind of eighteen months. But ultimately, the job thereafter will be look significantly more like an AI researcher, where I'm trying to push the bounds of what every new model release can do within the system that I've built, to try and find interesting signals that I can then read, understand and potentially act on going forward.
01:03:35 · Brett Caughran
[01:03:35] S1: Yeah, that makes a lot of sense to me. I mean, you can start to get ingestion with internal RMS and model architecture, meeting notes, etc. When you sort of deploy locally, right within the within the sort of fund firewall, and that seems to have solved a lot of compliance and infosec concerns, right?
01:03:58 · Alistair Smallwood
[01:03:58] S0: Yeah, exactly. And everyone's at a different stage in terms of their infosec issues, and some people are still really keen for zero day retention when you pass anything externally to a foundational language model. Other people are very happy with the setup, and ultimately there is this big trade off of red tape versus competitiveness. And trying to toe that line is I'm not envious of it at all, but it's probably quite difficult of really trying to protect your internal IP, but at the same time of not being left behind because you're going take the next eighteen months to build your own system and frankly build it well, which is an interesting thing that we're having conversations with. We've just been through the process with a very large UK based hedge fund, where we were the first tool for them to license that had MCP connection and also had in built off-site large language model capability. And so we essentially had to work as their guinea pig to run through this, and have spent a huge amount of time with legal compliance and InfoSec. Obviously, giving that, I suppose, comfort on how we process information and why it doesn't leave us and row level security and all of those things. But definitely not something that I thought when moving into this role would be a decent chunk of what I do in terms of helping people go through that process and be comfortable with the way that we... That businesses like us process information, because it ultimately is the new wave of what will be coming. And I suppose it's always that naivety of when you move to the new job, you only think about the thing that you're super excited about, which is trying to push AI to its to its absolute limits and equity research. I don't think about the contracting and the compliance processes and frankly having to deal with the systems that those compliance systems sit on is the worst thing in the world. I mean, they were built in 2001 and impossible to interact with. They're so bad that browser use of large language models can't use them.