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.
00:09:40 · Alistair Smallwood
In order to do that for a single company as a human, you have to have unbelievably good understanding of the full ecosystem of customers, providers, ultimately the macro that affects those things. And you spend a huge amount of time trying to understand economic transition or transmission between those three businesses or four businesses or whatever it might be. Ultimately, doing that at scale with large language models is genuinely really possible and something that we've written about last week. So to come back to your question specifically, like how do we protect ourselves from every next model upgrade or every next harness upgrade? Ultimately, what are we providing to our underlying customers? We are providing state of the art infrastructure and tools in order for them to build on top of. And the key thing that we think about is that all of our tooling and infrastructure should get better with every harness upgrade and every model upgrade. There's a huge amount of discussion out there around harness engineering and how to get that to be as good as it possibly can be. State of the art harnesses at the moment are Claude Poe and Podex, and what one should do and what we do as a business is we spend our time doing small improvements around the edges to those harnesses. If we try to replace them and build them ourselves, we will get carried out in three, four or five months time, where those harnesses have improved so much that all of that customization and months that we've spent building disappear. So essentially, we're trying to build an infrastructure that allows all model upgrades, all tool upgrades and all harness upgrades to flow through into our products rather than replace what we've already built.
00:11:24 · Brett Caughran
[00:11:24] S1: Yeah. Yeah, no, I think and we've become simpatico online friends because I think the sort of the two concepts of Mosaic theory of scale and systematizing alternative data and data gathering are sort of two concepts I've been navigating through in my own head as well too. And I agree are so much more powerful than the sort of consensus use case now of using Claude as a chatbot and asking questions in sort of an active way. Maybe on the first one, can we go deeper on that? What does scaling Mosaic theory in a systematic way, what does that mean? What does the tooling, the data look like to be able to do that? Then how do investors probably more challengingly start to rewire the way they think about learning, analyzing, tracking companies?
00:12:14 · Alistair Smallwood
[00:12:14] S0: Yeah, it's a great question. I'll probably take the first one first and then the second one afterwards. So in terms of the way that this is actually set up, it's kind of it's built in a few very, very key layers and to try and generalize this as much as possible. Ultimately, the first thing that you need in any of these types of processes, whether you're trying to process alternative data, whether you're trying to do mosaic theory at scale, or ultimately whether you're just trying to find new repeatable signals. The first thing that you need is really, really high quality access to primary data, so company filings, announcements, presentations, etc. But not only do you need access, but you need a state of the art way of retrieving that information 100% accurately, and being able to do that reliably at scale and frankly, efficiently, both from a latency perspective and from a cost perspective. That is something ultimately one of the key pieces that's allowed us to do Mosaic theory at scale is that it's the primer tools that sit underneath that to allow you to pull from, I mean, every primary document for 15,000 companies globally over the last at least fifteen years, and being able to do that in a scalable way. One of the really key things here is that lots of the data providers that provide those tools essentially just serve you a PDF. And we have seen and tracked in our own evaluations internally that the large language modelers love to get lazy when you've given them multiple PDFs over a huge amount of time. But within the MOSAIC theory at scale, was looking at the bioprocessing industry, where I was looking at somewhere between 10,012 documents at a given time and constantly cycling between them to try and make the connections. And so that was the first key piece, you need access to a very good way of retrieving data accurately. The second piece is ultimately the kind of engine that runs the whole thing. We use GPT-six Astra just because it was the most intelligent model at the time of writing the paper. But actually what sits within that is the skills that we taught the agent, which kind of split into a number of key things. The first thing is really explaining mosaic theory and ultimately what they're trying to achieve. You're not necessarily trying to achieve what you might be trying to achieve with looking at alternative data at scale.
00:14:34 · Alistair Smallwood
You're not trying to look for a 0.95 R squared correlation between two data sets, because mosaic theory is often qualitative, not quantitative, I. E. You want management commentary saying all has improved in this business unit and whether the second indicator had a similar direction. And so really explaining how to map those things and what matters and what doesn't matter. We also then have another skill which talks about the specific industry that we were looking at and how it constantly writes to that skill and updates it. And then also thinking about how you would think about actually doing this as an individual, economic transition and how you would qualify whether these two companies are related, and how you would check that through either management commentary or trying to understand whether there's any relation between their inventory build and ultimate revenues of a different business. Those types of things that you would do ordinarily as an investor, really trying to codify those into a series of skills so that the agent has access to those consistently. So you now have got the kind of data access and you've got the engine plus the skills. And then the third thing is arguably probably one of the most important is the evaluation engine. So every time the agent that's running thinks that it's found a relationship, you have to have a really high quality evaluation engine that sits underneath, that tries to qualify these leads, rechecks them, runs adverse aerial agents in the background to really understand these things. Ultimately, once all of those things are qualified, we actually then used JEV that sat underneath. It's obviously been huge hype across multiple platforms, but a really key thing when you're doing mosaic theory is ultimately you want to be asking a very simple question. If I had all of the information just from the company, what is the probability of the next quarter that I haven't shown you yet being positive or negative in terms of growth? And then I show you all of the information that I've collected from a mosaic basis, and then does that change the answer from Jeb suggesting that forecasts or the next quarter's growth would be more positive or more negative than the previous measure?
00:16:46 · Alistair Smallwood, Brett Caughran
And the fact that that was essentially deterministic massively unlocked our ability to do that work at scale and run that evaluation engine in a loop until we found some really exciting examples off the back of it. In terms of the second question, sorry, go, Brian. [00:17:01] S1: I want to just add one. I've had debates with a number of quants over the years that have made the claim that you can systematize fundamental investing with the technology. And I think the technology is there. The pushback I always have that no one has convinced me of yet is the fundamental investors or traders decision point is based off a concept of completeness of research, right? So there are times where maybe there's a signal in a transcript that this number is conservative, and that might be where I'm hanging my hat. But more often it's, I met with a management team at a conference and they're backing off of guidance, or I met with a competitor who's telling me this end market's slowing down and I can connect the dots. It's context that is gathered in meetings on the phone in messy sort of non agent legible ways. How do you think about that? Like if we really want to scale Mosaic theory, how do you think about scaling that completeness of context that is so critical to do that consistently?