00:01:11 · Alistair Smallwood
[00:01:11] S0: Yeah, absolutely. It was partially by accident. I worked on the sell side for eight years and was approached by a fund which was launching in London, running an originally pan European long short market neutral strategy, and the core thing for that PM was for all joining analysts to be code proficient on starting, because the modus operandi of that fund was to be state of the art from a technology and AI perspective. The PM was a big believer in potential advantages from both productivity, but also ultimately alpha generation that it could bring. His key thing was that he wanted there to be a very short space between deep domain expertise within fundamental investing and how that could translate into the building of AI tools, rather than farming it out to either an internal team or somebody that sits separately within the PM and analyst team. So I was sitting in my seat at UBS and I was approached to join this fund, but a prerequisite of joining this fund was understanding how AI works, the coding piece, this is obviously pre coding agents, this was kind of late twenty twenty two. So I kind of set about understanding how coding works, building a few tools in the background, looking back on them, they were worse than average, but ultimately there was a huge learning process that came with that. Then continued to work in the fund for two years, building and maintaining a decent amount of technology infrastructure within that, which spanned from scraping alternative data processing and machine reading of research. The And PM of that fund is the largest shareholder in Primer. It became a separate business very quickly off its founding. And the excitement that I had in terms of ultimately where AI and more broadly compute can take equity research was particularly exciting for me. And so I made the switch from the fund into Primer about twelve months ago.
00:03:08 · Brett Caughran
[00:03:08] S1: Great, great. I think you had the benefit of getting your hands dirty very early in sort of the applied AI moment around GPT 3.5. Can you sort of like, I think in the rapidity of new technology, sort of it's now like a three week cycle, it's easy to lose context of how far we've come. Can you sort of like take a step back and what are your reactions to the curve of improvement in the tools for this craft over the last four years?
00:03:39 · Alistair Smallwood
[00:03:39] S0: Yeah, absolutely. I think the best way to communicate it is really to try and give context of something that we do now versus how we were doing it back in 2023. And a great example of this would be the core of any AI and equity research system is your ability to process and machine read primary filings from companies, so 10 Ks, 10 Qs, annual reports, quarterlys, trading updates. When we first started building this at Primer, we were dealing with large language models with context windows of 4,000 tokens. They couldn't process PDFs, they couldn't process charts, they couldn't process tables. And as a result of that, we had to build our own fully custom pipeline to do that in the most equity research specific way. What I mean by that is that instead of just passing a PDF and an agent hopefully understanding the things that were important, we had to build a totally custom pipeline to make sure that some of these annual reports that were a million tokens plus were being broken down into the right sections and then being pieced together, and ultimately the output format was as agent readable as it possibly could be. Because as we know, doing this job, if you miss a number or a footnote in the accounts, you could totally break down the whole working process. So where we've come to since then is we now have million token context windows, people are no longer chunking bits of information for code to process them, and large language models can broadly natively process charts and tables, although they're still nowhere near where they need to be in terms of doing that consistently and reliably, so our custom pipeline is still running. But there is a huge amount of work that you can now do from that perspective. I think one of the key things in terms of our own internal technology development that we have seen change enormously is in previous worlds when you build software, you will have alerting systems that suggest that there's a problem here or there's an error here, and then it requires you as the individual to go in and check that and debug it and try to understand and then rebuild and fix it.
00:05:43 · Alistair Smallwood
We've now got to a stage where coding agents are so good from our perspective that every time that there's an error logged, we have internal agents that go and look at those errors, fix those errors, and then propose the fixes to us, which take significantly less time to check, which means that general reliability and uptime of platforms is so much better. But it also means that from a productivity perspective of us, the engineering team, almost all of the work that is done is go forward work rather than maintenance and tech debt that has to happen on an ongoing basis, which I think from our perspective and something that we've already experienced means that product development and velocity is just unbelievably fast. And I think that I don't think it will be news to anyone listening to this that, yes, coding agents are really good and you can build stuff really, really quickly. But now sitting within an engineering team, it's now very obvious to us that all of your time can be spent doing that rather than bug fixing and historical stuff and making sure that tech debt is fixed on an ongoing basis. And that is a huge step change for us in terms of potential velocity of the business.
00:06:54 · Brett Caughran
[00:06:54] S1: And one of the things I hear as a theme from people building AI tools for investors is just this almost sort of endless obsolescence, sort of building sandcastles that get washed away by the next model iteration. Can you sort of walk us through that dynamic and how you navigate? You build something, the next model comes out, it's obsolete, you're decommissioning, you're almost sort of recreating the spine. Maybe that's a sort of oversimplification, recreating the spine of the technology every few months.
00:07:25 · Alistair Smallwood
[00:07:25] S0: Yeah, it's great point. And I'd refer back to a great article written by Rich Sutton called the bitter lesson, which essentially says that over seventy years of AI research, ultimately, compute will always win. So if you look back at when they were building models to beat chess grandmasters, or they were building models to beat people at the game Go, which ultimately people thought was too human that you couldn't possibly have models that do these things. The key thing is that we think about and applying that lesson written by Rich Sutton is that all of the infrastructure that you build and the place that you spend your time has to be through the lens of are we just going to get subsumed by the next model, the next harness, the next tool? And ultimately, where do we think that we sit within that? And how can one be bitter lesson pilled from AI and equity research? The really core thing is that within that bitter lesson, the core thing was instead of necessarily only trying to encode human thinking into AI systems in order to get them to do a job to a standard that you recognise, It's really about trying to use compute to generate and develop novel methods of doing this type of work. A great example of this would be for one hundred years, people have modelled companies in almost exactly the same way, where every single driver cell and forecast cell is a point estimate, I. E, a single number. Ultimately, when you do that forecasting yourselves, in your head, think about a range of outcomes. And so for every single one of those numbers in your head, you are holding a probability range of outcomes. And let's say that there's a 50% probability that they grow at 2%, but there's still probably a 25% probability that they grow at three or four. Ultimately, the really exciting thing about applying AI to these types of things is that they can model things in those ways. And we're really pushing compute to the next stage. Published a paper last week where ultimately we've demonstrated that we think large language models can execute mosaic theory at scale. Every investor knows that you can have five independent pieces of information that mean nothing on their own, but actually when you pull them together, it can raise conviction levels in the target company that you're looking at.