Invest Like the Best with Patrick O'Shaughnessy
00:18:40 · Patrick O'Shaughnessy, Gabe Stengel
[00:18:40] S3: Is there a favorite example of that of some like pain in the ass thing that you've had to wire up last mile things? Yeah. [00:18:46] S1: I mean think about if you are ingesting MMPI because you are working on transactions or you're working on deals that have an effect on the market, the compliance requirements you have just in the auditability or what you can flag and how it feeds into the internal systems of an investment firm or a bank so that if you ever get audited or a regulator ever wants to see what you did, do you have all that plumbing in place that is pixel perfect? That's one example. Another example is if you actually wanna transact, say you are a big public company buying another big public company and you need to send data back and forth, you actually need some sort of data room. Something that is compliant, safe, and secure that coordinates. And ideally it's not just some static kinda Dropbox folder but something that's plugged into the way that you do work, your agents, your workflows. And I don't think Open Air and Thropic will ever wanna build a data room business. If you actually wanna be the exchange for all of high finance and all of capital markets, you not just need to own the intelligence, you need to own the transaction venue, the communication venue, the workflows, and all the data inputs that go into it. [00:19:51] S3: It's interesting. I asked about data and models, but actually it sounds like harness or infrastructure is probably the most important of the three.
00:19:51 · Patrick O'Shaughnessy, Gabe Stengel
[00:19:51] S3: It's interesting. I asked about data and models, but actually it sounds like harness or infrastructure is probably the most important of the three. [00:19:58] S1: Think about the fundamental difference between Cloud Code when it came out and Cloud Cowork versus OpenAI and ChatGPT. The models were actually fairly similar but the harness and the way that it was presented from Claude was far better and it just allowed the models to exercise more of their long running capabilities and that's why they had a run up in usage and a huge amount of expansion. And it just shows the way that you harness these models is so so important. And I think people under appreciate that intelligence, the reasons that humans are high agency and can do a lot is not just because we have high raw recall and IQ and knowledge, but there's all these different micro services in your brain. How do you put knowledge away? How do you retrieve it? How do you trigger things? Emotions are a way to trigger all these different micro services. That's why a great investor might have great judgment is because they have good spidey sense for when they see this sort of thing in the market. It actually triggers the recall from this event that informs a creative decision. All those kind of small microservices are things that need to be built out.
00:19:58 · Gabe Stengel, Patrick O'Shaughnessy
[00:19:58] S1: Think about the fundamental difference between Cloud Code when it came out and Cloud Cowork versus OpenAI and ChatGPT. The models were actually fairly similar but the harness and the way that it was presented from Claude was far better and it just allowed the models to exercise more of their long running capabilities and that's why they had a run up in usage and a huge amount of expansion. And it just shows the way that you harness these models is so so important. And I think people under appreciate that intelligence, the reasons that humans are high agency and can do a lot is not just because we have high raw recall and IQ and knowledge, but there's all these different micro services in your brain. How do you put knowledge away? How do you retrieve it? How do you trigger things? Emotions are a way to trigger all these different micro services. That's why a great investor might have great judgment is because they have good spidey sense for when they see this sort of thing in the market. It actually triggers the recall from this event that informs a creative decision. All those kind of small microservices are things that need to be built out. [00:20:59] S3: But if I force you to become an investor and your only goal is to invest in roguelike businesses, like one of these businesses that's, let's say, a vertical application that's wiring up the capabilities of AI to an industry, what features would you look for that would get you the most excited either in the industry or in the founder builder and their approach based on what you've learned?
00:21:20 · Gabe Stengel
[00:21:20] S1: A few things. One is the industry actually does have to have enough complexity and depth in the types of data, the types of systems of record that people use, the types of deployment models that you can spend a lot of effort solving those problems in order to have a wedge to solve everything else. Because if it's an industry that anyone can just walk into and sell the basic version of Chad GPT or co work or co pilot immediately and you don't have to solve all these weird integrations, you're not gonna have enough time to build all those things that are perpendicular to what you're doing. And so the industry itself needs to be adjacent enough to the core market. So that's one thing. The second thing I would look for is domain expertise from the team and the founders. And I don't have unique domain expertise but I had enough to get started and then I was so curious about finance and money and capital markets. And I grew up in New York and I was surrounded by people that all they could think about, all they could talk about was high finance and I was fascinated by it and I wanted to learn about it. And then we assembled a team that was uniquely passionate about it too. And so I have over a 100 people that have spent time within investment banks or within investment firms across the world's best institutions and so we can constantly take the models as they're released and harness them for finance. And our job is really to catch the change in the models and figure out how to apply it in these institutions. The final big thing I would look for is a business that's willing to constantly reinvent the core product and constantly willing to slash it to nothing. And I think anyone who their delivery method of their product is not something that they can fully cannibalize quickly like a terminal or like a very specific UX or interface is not going to be agile enough to constantly reinvent every six months when there's a step change.
00:23:03 · Patrick O'Shaughnessy, Gabe Stengel
[00:23:03] S3: What's an example of that that you've done like a tear down and rebuild? [00:23:07] S1: The story, the anecdote I'm most inspired by is Max Levchin who talks about how at a firm they rebuild the fundamental ledger technology every year. And they rebuild it for a few reasons. One is it's the most interesting engineering problem and so all the engineers wanna work on it. So it's a good way to retain talent to teach engineers about the core fundamental business of a firm. But But then number two, it's a good way to make sure that system doesn't ossify and it's constantly improving. And so we do that same exact thing for our harness and the core agentic system. We are constantly looking at it realizing we're not even at a local minima. It would be impossible if we were at a local minima because the models are changing so quickly and we need to redo the whole thing. [00:23:49] S3: What's something that the models currently cannot do that if they could would really change the nature of the product?
00:23:49 · Patrick O'Shaughnessy, Gabe Stengel
[00:23:49] S3: What's something that the models currently cannot do that if they could would really change the nature of the product? [00:23:55] S1: Compaction. So it's if you have a 100 conversations with a single agent, how does it make sure that it's actually remembering the right things and compacting its memory into an amount of tokens that it can use every time and it has enough coherence and context on who you are and what you care about to make it feel like it's a true person that you're speaking to that you know learns more and more about you. That's a hard problem and it compounds exponentially when you think about agents not just as a one to one. Right now almost all agents are one to one. You use ChatGPT individually. You use Copa individually. You use Gemini individually. As soon as these are actually things that consort with a lot of colleagues or in a Slack channel with a 100 people have to work across an entire company, now the compaction problem just scaled exponentially because it's having conversations with a 100 different people and needs to be able to coordinate across those things. And so being able to take all of that memory, those interactions and actually lodge it into the mental model or brain of that agent so that it can be persistent so that it can actually maintain context over the course of a bunch of interactions, that's something that the models are not great at today.