Transcript
Use with ChatGPT / Claude Auto-scroll A
The panel opens by framing the current uncertainty around AI infrastructure, economics, and scaling decisions. Each speaker then describes how AI is being used inside their firm today, from internal labs and agents to centralized tools for research and operations.
Conviction in what AI can do is is of course rising now, but just as uncertainty grows around how to scale it, how to architect around it, and what the economics look like.We've also seen over the past 2 weeks concerns in the markets around AI infrastructure financing and the huge the enormous kind of capex has those concerns have intensified as well.This uncertainty runs straight through the technology stack and into strategy.Of course models are leapfrogging one another, costs are shifting, and open weight alternatives are also keep improving.So today for this panel what I'm going to do is move beyond the AI use cases and into questions of scale, economics, ROI.Also how can investment firms genuinely differentiate and how does the investment process itself evolve?I can't tell you how excited I am about this panel because on the same stage we have three senior leaders bringing investment, AI, and technology perspectives from some of the most sophisticated hedge funds in the world.
So let me start right there.Of course you have different perspectives and you're also approaching AI in your firms differently.Could you each, and starting with you Aaron, could you each give us the short version of what AI looks like in practice inside your firm today?What what have you productionized?What are you in experimentation mode?What what what do you find challenging?
Sure.So let us start 3 years ago.We founded AI Labs which AI stands for artificial investment associate as an internal venture to experiment and rethink how Bridgewater understands the world and trades global macro market markets using AI and ML first.We did that for a few different reasons.One of them was so that the team was focusing and but we always had a uh the the ultimate goal was was to take what we were learning there and fold that back into the rest of Bridgewater.And so, after about 3 years at the beginning of this year, our CEO, CIO, and our board all decided that then was the time.And so, now what we're doing is taking our learnings from AI Labs and imbuing that in all of our research processes.So, a few of the examples of the things that we've done is we have created a giant unstructured database to rival our our time series database.
Uh we've also created several unstructured to structured uh investment signals and spent a fair amount of effort first creating what we call the Pocket Analyst, which was able to do um between 20 minutes and 2 hours worth ofwork in May.And now that's graduated to Pocket Analyst 2.0, which is doing something like a few days to a few weeks of work.
So, um Citadel, we have, you know, Citadel hedge fund manager and Citadel Securities.Um and across both, many of our functions look the same probably as yours.The engineers, 100% adoption, don't really need to cover that.Uh we started um on, you know, with agents around uh you know, operational workloads and away from the investment process.Um that goes well and we can go into some of the details, but just as a good experiment to reconcile our trades.So, we tie out every trade, you know, you can imagine millions of trades uh across both businesses to the penny, and humans used to look at all the trade breaks.And so, as a proof of concept to say how can we get used to this kind of agentic flow?Uh now the agent resolves those or at least gives it to the human uh with the exact reason why dramatically, you know, improves the humans' time and and decreases risk.
Uh and then on the other end of the spectrum, we have some of our investment teams have built their uh agents that look like them and can do research on their own uh and be unsupervised for for someperiod of time, and then kind of everything in between.So, you know, not kind of a center or one-size-fits-all approach.Kind of every business will do it themselves.
Yeah, great.So, in our group we have a couple of different businesses we have to support, both systematic and discretionary investors, right?And and PMs.And obviously on the systematic side we've been using machine learning for decades, and so it's really sort of the first use case came up on that side.But generally what we've seen this year is really a more of a central deployment.So, primary tools we have, kind of I would say more friendly user interfaces that allow us to pick from among 10 to 15 models in which to use in that particular use case.I think what's really important is we don't have vendor lock-in, and we'll probably talk about that in a little bit, but just giving the flexibility to make sure which model do I want to use in which case, andthat supports both sales, operations, investment functions.
It allows people to then build their own workflows.So, on the discretionary side, we've built a system we called Manda, which is a Manda discretionary assistant, and that allows discretionary PMs to handle their their typical use cases.So, build me a income statement for Nvidia, for example.It'll It'll just do that.So, very basic tasks.Of course, they'll probably want to double-check it, but the idea is to sort of lower the barriers to gathering information and connecting Bloomberg, being able to throw a portfolio into the system and getting some some real-time feedback.On the systematic side, it's evolved a little bit more into one just kind of process flow.But on the research itself, thinking about delivering models into the into the market from a research perspective, and we've been doing that.We've had some good, you know, initial experiments on that side.Well, we'll go to some of the pros and cons of of that approach, but I think my my general takeaway so far, and from the experiments that we've run, that AI will find things that maybe the humans have
not, right?But it's not going to have a significantly larger sharp ratio.What we'll find is things that have been largely uncorrelated or have low correlations.So, I think that's really our view is that it's not going to be necessarily amplifier sharp but only through but maybe through diversification many more signals.So some early day good you know kind of good outcomes in early days.
Good.I'm going to talk about non-correlated returns in a few minutes but before that just on the uncertainty point um it's quite notable that in this tech cycle uh all firms actually have to make a lot of long-term decisionsand kind of capital allocation decisions when a a lot of things are changing right?Frontier models open weight models inference pricing hardware deployment patterns harnesses orchestration frameworks so many things are changing at the same time.So what's your philosophy like?How do you make these durable architecture and capital allocation decisions?Now do you find it challenging?
The discussion turns to durable AI architecture, model flexibility, and how to avoid overbuilding around fast-changing providers. The speakers also explain how they manage AI costs, optimize usage, and decide when to stop or kill projects.
Yeah I I guess it's start there.Yeah the only thing that I think we're betting on right now is that the models are going to continue to get smarter and more capable.And so what we're trying to do is ensure that whatever we are building we're excited when the new model comes out and we're not regretting spending a lot of time creating scaffolding around that model.Um and we're focusing on what model companies and other companies won't do for us.So that's things like getting our own context um uh creating positive so that's number one getting our own context together.Two would be creating an environment and we hope that there is a provider eventually that will provide this but creating environment for the agents to be able to take maximal action in without putting Bridgewater at risk.That is really difficult and it's an unsolved problem right now for us.
And then the third is creating a learning flywheel between those two things.So the agent takes action we then can improve the context and so at some level that's what we're creating and we're not trying to we're assuming that many parts of the system are going to be changing and architectingappropriately there.
We spend a tremendous amount of time also making our sandbox that keeps the agents in the walls or tries to.You know in in addition we can move pretty quickly and try and have hedges out there.So, you know, if we are mainly using one provider, one harness across call it 95% of our users, we'll make sure that 5% or always using a different harness, different model provider.I generally try to prevent people from picking model and chasing model and not spending a lot of time on that because they are going to continue to evolve.Very few of the use cases, you know, are need fable class models.Often I find it's, you know, sort of lazy and I want to make the button blue and I can just use fable to do that.So, we're, you know, trying to to focus people on what are you working on and not the the tools that you're doing.
And we move quickly.We change providers, you know, in a two-week span pretty easily.
A lot of flexibility both sides.
Yeah, same here.I If I had to project, I mean, this is always hard in the space cuz it is changing.I think what we are seeing is some convergence on the capabilities of these models both in level but also the dispersion around the models.So, I think that's important.I think we're kind of going from a game where the answer for large language models was just scale.We'll just buy more chips.We'll just crank.We'll just, you know, do as much as we can and we'll just build it all up with just pure scale.I think now it's a little more nuanced where I think maybe you want a smaller model.Maybe I can be a little bit more efficient in my search so that the next level, I guess, or the next, you know, the next few years will Yes, we'll I think we'll see some development but maybe open
weight becomes more competitive.You always have to be flexible to identify that but I think it's going to be how you figure out the amount of data that we have, right?So, if you're sitting in a mid-frequency space, forget high frequency for a second, there's only so much data, right?You can scale up the number of hypotheses that you want to look at but it's hard to scale up the data.There's only so many global financial crisis in the data set.There There's some fancy things you can do and sort of create additional synthetic data but that kind of fails after a few times kind of model failure.So, it's really you just have to be careful.I think it goes back to the organizational kind of how it how it brings that into this into the how the humans interact with the data given that you want to have good good research practice.
So, a lot of the investment is harnesses.How can we develop a man group researcher in in sort of synthetic format, right?The rules and the culture and all of that.Can we put Can we codify that in the system uh and try to be flexible.
And uh Andrew, you you mentioned one thing about model routing and I imagine that's the same across.Uh but you know, you mentioned it for the reason for cost.I'm just wondering whether the costs of um of AI have now become a topic.We've moved very fast from token maxing to okay, the cost of these agentic systems.Uh and agents, of course, uh it's it's like maybe 1,000 x multiple versus a simple prompt, right?If you're trying to build a full agentic workflow.Has it really come into your organization yet in terms of we have to optimize on cost or the spends can go out of control.
Yeah, well, so we you know, sort of Opus 4 5 early this year kind of I think awakened many people and and that definitely pushed us harder.Um that uh our goal was like 100% adoption.How do we get as many people as possible using this not in a token maxing way.Thankfully, we weren't ever bit by that.Um but it you know, the costs in absolute terms are real.It's a substantial uh dollar amount.But for us, it's more about waste.So, cost is an input to the investment process and should generate returns.If you are working on either a bad project that is unlikely to be successful and it was unlikely to be successful with linear regression, it's unlikely to be successful with traditional machine learning, it's unlikely to be successful withthese products.Um that is likely wasteful.And then if you're doing it in a wasteful way, so you know, do you have a huge context window?
Are you uh you know, using the the the higher effort than you need, more uh larger model than you need.And so, rather than give people the these draconian, you know, uh walls, we just give them in their own insights page that says like, here's how you're spending.If you made these three changes, it would reduce your spend last month by X thousand dollars.And people generally try and try and do that.So, we focus more on waste and our we have two charts that we look at.One is, can we get the number of requests to keep going up?So, can can people use it more?And then, can we get the absolute spend to go down?And so, those things thankfully are moving in a K a K shape right now.Um, and if that continues, we we, you know, feel pretty good about it.
Yeah, I think we're definitely in this optimization phase.How do we use this better rather than just using it?Again, it's widely available and in house, but how do we become smarter at it?So, I think part of it is understanding the tasks.You know, personal example, I use a eight to nine agent research deck review research proposal comes to me other ideas.It's got eight or nine agents associated with it or I'm writing a paper what it might be.And so, one of the agents might be my, you know, the checking of the bibliography, my sources.Like, I don't think I need a fable 5.1 for that or you can pick.And so, I'm optimizing with our tech team.I have some test cases and then, you know, with these nine agents, you can sort of score them and say, okay, you can dumb that one down.
This one makes a lot of lot of mistakes, so upgrade that one.Oh, fable is out here.Can we should you upgrade this other part of the process?I think that that organization I think that it's key is that how do you take that knowledge and spend it across the entire organization?So, even if you're in operations or investment function, how do you know that you're getting better with these tools?Um, every I don't know if your organizations, but everybody's built their own dashboard, right?Everybody's got their own proprietary dashboard.Now, we have several that probably do exactly the same thing.So, we're we're sort of trying to get past that phase and have some just direct cases.And I think ultimately, some of it has to be top down.I like to have a lot of it bottom up as people are creative.But, you're right, you sometimes you have to have people stop working on projects.
Knowing when to kill something is as valuable as knowing when knowing when to invest.
The panel explores how far up the infrastructure stack firms should go, and how much AI capability should be centralized versus embedded in teams. They then debate hidden correlation risk, human oversight, validation, and the danger of cognitive surrender as AI speeds up research and coding.
And a quick question, curious for all of you actually, how far up the infrastructure stack are you willing to go?There's been a lot of news about I won't name firms, but about investment firms actually investing in data centers directly.So, how how how important is that to you like right from you know, dedicated GPUs, reserve compute, direct investment, or are you happy to have pay as you go inference?
So, I there's a lot of great attributes to pay as you go.You can scale really quickly, which is wonderful.And so, but also you're kind of at the mercy of those providers.And so, we think about our portfolio of those providers to ensure that we're hopefully immune to any shock in any one of them.And there's a lot of also security concerns with several over the past few weeks.Um And so, we there's definitely a place for those inference providers.However, we also found good value and uh and actually increase capability in some form of RL fine-tuning.And that's only available generally on your own hardware, or there's a few providers right now that are helping you do that as well.Um and then lastly, uh for the most possibly for some of your most secure data, if you really don't want that leaving your walls or your VPC, then then there's also a place for GPUs.
So, that's mostly how we're looking at it today.
We also we generally like using a hyper scaler in front of one of the you know, frontier labs.We don't have to share We already trust those hyper scalers with our data.We trust them with you know, the the storage of our documents that describe investment strategies in detail.So, you know, we don't have to add another one and you know, vertex or bedrock, they you know, run those things for us.So, that's generally our choice.We also have an on-prem footprint with our own hardware for similar reasons.
Yeah, I think we've generally been a pay as you go in capability.Again, it's flexible, scalable.You can change models very, very, very quickly.I think the only kind of cost pressures obviously token demands going cost is going down and obviously we're using more of them.So, you can think about the math on that, but there are some ancillary spillovers in terms of hardware costs that we need to send broader GPUs, other things where the demand just in the whole ecosystem is so high,you start to think, well, should I should I delay this purchase of this machine cuz maybe in 3 years it'll be really cheap.
I mean, you know, there's there's some sort of thought about how to the optimal time to buy your own hardware.I think we're entering into that phase now I mean, hardware cost has always gone up and data costs, but in this environment how to optimize that is important.
Switching quickly to your your operating model within the firm.So, AI of course has a lot of fixed costs, right?You have data infrastructure, you have model access, compute, security, eval, all of these things.These are hard to reproduce team by team.So, in some sense of course centralization makes sense, but in investment firms traditionally many firms actually kept the data and tooling quite close to the individual investment teams, the PMs and the researchers because they can shape the stack,right?So, is that old model changing a bit?Do you feel AI is pushing more centralization or are you trying to keep the edge very close to investment team directly?
I think we we still keep a lot with the investment team.So, you know, Citadel we try and provide a platform that is useful.And so, you have to decide when you're going to platform something cuz if you platform too much, it's sort of dead weight and slows everybody down and every business owner can say, I couldn't move fast because of theplatform.So, we try and make sure people can run their, you know, individual teams.And then when we spot, you know, duplication, we've decided which is the right solution, we'll sort of suck it into the to the center.We do that, you know, if you just measure in head count, our our AI team that's central is very small, probably under 10% of the the team that's working on AI across the firm.Um most of that is spread out in in the investment teams.
Yeah, I I'd say for us, I think part of the bit we we try to keep the data in as central as possible in terms of the allowing both discretionary and systematic portfolio managers access to it.Again, security masters, all of that.I think that's the goal, but there is a bit of a trade-off in terms of speed.But, I think one of the things that I I I've liked about AI, it's in the new new technology, it's allowed us to take on projects that I don't think we would otherwise take on because of the costin terms of particularly data management side.You know, taking different security masters, other things which you could probably hire 10 people in the past to do, and you do it for a year, and then you come out with the same thing that you already have.It's just in the central place.
Probably not worth it.But, in this new environment, you can start to lower the the barriers to to entry on centralization.Uh so, that's another piece and that that I think is new or that I'm kind of interested in in terms of our stack.
Yeah, Bridgewater's generally been fairly centralized with our platform.But, given how fast the capabilities are changing, we've found that and I pushed teams now out into the various um asset class groups, and are getting out and not only are those technology teams delivering or creating capabilities thatwe're bringing back into the platform, but also the investors actually are now coding and bringing that back into the platform.
Yeah.Uh and you mentioned the uncorrelated returns.Very important to to firms like yours.Is there a concern at all that if um teams increasingly use the same models, same data, same retrieval systems, same kind of research tools, same prompts, they they start thinking the same way, and um maybe there's some hiddencorrelation risk also because of the commonality of the platform and the
Yeah, I I I did grow up on the systematic side, so I've heard that argument systematic investing for 20 years, I think, that you know, because everybody's using systematic, we're all going to have the same the same trade.So, that's that's always been there and and you know, obviously a concern.I'll say for you know, in our experience both with the Agietic systems, um you know, you you do see some uncorrelated ideas, not necessarily powerful.We've also done a couple of natural experiments and more from you know, people.So, think about competitive firms, right?So, in in one case, uh we had a machine learning algorithm that we were using on a certain piece of data.We got introduced to a startup with eight people sort of very focused for like 5 years building their own systems.Yeah, this is more machine learning than traditional machine learning.
And so, I wanted to have the experiment of like we got some smart people over here, we got some smart people over here.Let's have them look at the same data and what comes out of that.So, again, from that experiment we saw uh uncorrelated but similarly powered models.So, that was one.And then we even within the broader Man Group, we've had a couple of different equity teams, one in London and one in Boston over the years.And again, similar outcomes.So, I think part of the machine learning in particular, the hyper parameters, how I trade it, does create a lot of potential differences in the market.I guess we need you know, we need to be concerned about overlap, but again, humans are designing these systems.And at the end of the day, you know, the markets are are human driven.And and so, the model today might be different than the model tomorrow.
So, there's a lot of really cool things that you can look at, but I think you have to have some experimentation to see if this is actually an effect.I mean, we we see teams that cover the same stock, take differing views of that stock sort of regardless of of how the view comes up.Um one thing we do worry about is people outsourcing the most critical parts of their job.So, if you're fundamental investor, you know, there's lots of things uh that you could use this for, right?Write my thesis on this stock given all of my data.Okay, that's one that's one uh extreme or uh prepare me for this meeting with management that I have or that's probably more appropriate, less outsourcing.There's more um red team this for me.So, be my counter thesis in debate.So, I see great PMs that will go back and forth.
Uh you know, having this thing truly disrupt try to dispute them and sometimes it will convince them that they were wrong and they might they might actually change directions on the stock.Um and you know, those are I think great outcomes.And then we see on the other end, I'm curious to know what what your investors creating the So I'll call them amateurs, but like the amateur programmers programming, like how that ends because the ones that I see, it'slike the ideas aren't great, lots of dashboards, like always trying to kill those things and say, you know, if you want a programmer to do something, probably still programmer plus, you know, AI might be better.POC maybe yourself, but I'm still a buyer of the the
of the That's good for me.I'm a technologist.So you know the job.Um
I'm talking my book a little.I am, too.
It's good for us.But uh yeah.We are seeing So, first of all, it is a great spec at the very least.It's like this is what I want.And yes, they are they make they have some bad ideas.I saw it.There's many bad ideas.Um technologists also have many bad ideas.And what's amazing about this technology is that you can cycle through those ideas so much faster.It's dizzying for me to look at the uh pace of our our MRs or PRs that we're pushing through on our in our more experimental platform.Um and figuring out how to uh how to blur the line between the engineers and the investors is is part of my goal because I see such great outcomes when we're able to do that.But it's difficult to actually make happen.
Your comment about dizzying, it's exactly the word because obviously AI, whether whether it's coding or whether it's it's any any kind of task, uh I don't know whether humans can actually keep up with what AI is doing.So, obviously there's the human in the loop kind of comfort factor, but what is your opinion?Does Do humans actually become the bottleneck because you can have two extremes, right?Either the human is a bottleneck because AI is generating so many things which humans cannot review in that much time, or humans also get into that thing of, you know, rubber stamping, okay, I approve because I I II can't really process all all this information.
Yeah.Humans are definitely a bottleneck.
Um Are the bottleneck?
Are the bottleneck.And then it depends on what type of system and what area you're in on whether that's okay or not.For our trading systems, yes, we absolutely want the human to be the bottleneck.In fact, the humans are still bottleneck and then we add on top one more bottleneck that's very helpful, which is an AI, you know, red team basically on ideas.So, that's that's one case.In cases though where you can where the cost of being wrong or a mistake is significantly lower, removing that bottleneck can be very freeing and can can change the uh velocity of the team.
Yeah, I think this validation step, you know, go from hypothesis validation to production, right?That that's a source of if people can figure that out, that's a big source of competitive advantage.I think organizationally, certain cases I think people over trust AI when they shouldn't and sometimes they under trust AI when they shouldn't.And it's trying to create the right incentive structures and the right kind of procedures around when do we want to do it?Sort of the the problems you want to apply it to or the what's the risk if if the machine is wrong?But ultimately, we have to be really careful because we can't go to our clients and say, you know, the machine told me to do that.That's why we lost the money for you.So, you still I think even even in those cases where there's the machine is making the majority of the decisions, you're going to need some kind of explanatory layer that you can then give to the clients to say,
okay, this is why the, you know, relationship between momentum and volatility is this random squiggly line, right?That comes out of this this fit that I can point to you why we have that that that relationship that you might not be too intuitive to a human.So, you got to spend as much on explaining to yourselves and to your clients as you are just letting the system generate good investment ideas.We see the in the example I talked about before about the the ops and clearing and settlement piece, as the engineers were getting so excited, it's like, oh, it's 93% accurate, 94, 95.Okay, now we're starting to get into the danger zone where it's 96, 97.The human is completely desensitized and they're going to like very risky of like the Homer Simpson press the thing like over and over because, you know, it's starting to be so good, but it's not good enough.
So, we're thinking like, do we have to introduce false failures to like keep the team on their toes and you know, you could see on the investment side of things, I feel like we're in a in a uniquespot with the kind of computer-assisted humans that are the drivers of much of our investment on the fundamental side that, you know, we're a little bit less worried about hallucination and things because they're used to taking in baddata.You know, they read a lot of research reports that are, you know, let's say not predictive of what's going to happen.So, they're very good at assimilating that and so we can be a little more risky on those tools because they're already very skeptical of everything they consume.So, kind of depends on the case, you know.
My question is and you said it about getting ready fast at hypothesis generation, prototyping, coding, all of those things.Does it actually go to faster end decisions because this this the front end is is going faster, but you're not pushing it to trading.There's validation, there's all these things.Is is the whole chain getting faster at the same time?It doesn't sound like it.
creates a chemical reaction, right?The rate limiting step is still the people in the middle, right?So, that's again culture to figure out.I think the other thing you do worry about, you know, cognitive surrender.Are your people really getting into the data or are they just relying on AI?So, part of me feels like we should just get rid of PowerPoints and people should just sit in a room and talk cuz then you'll actually figure out what if the person really knows what they're talking about cuzyou can make some really really nice PowerPoint presentations.I It is It is It is sort of embarrassing if you ask a very open-ended question to a group of four or five people and they all come back with sort of the same answer in the same formatwith the same graphs.So, you're like, "Oh, yeah, that's I wonder what you did for that."
So, there's just a lot of Let's go back to human interaction, I think, is really really important for this.
The speakers examine which strategies AI will affect most, from high-frequency trading to fundamental and mid-frequency approaches. They discuss alpha discovery, alpha decay, and whether wider AI adoption will make markets more competitive.
Um switching gears a bit.So, we spoke about the investment process.I want to talk now about the investment edge.So, and we are entering speculative territory here, but very curious on your thoughts on which type of strategies will AI impact or will either make it faster or more alpha.So, from the HFT which always traditionally use a lot of ML are we going to see acceleration there or are we is going to be more in the mid-frequency space or it's going to be more in the fundamental,you know, fundamental equities, fundamental credit, macro where there's a lot of messy information, fragmented data which was not not used that much at scale.So, do we have any views on which parts could be impacted most?
So, yes.All the parts.You know, for us it's about um how can we continue to deliver returns when we know the rules to the market.So, if we know the rules to the market, our job is to try and win within the rules.And, you know, we had all novel data sets.What is it?2026?So, call it 11 years ago, 12 years ago, there was all these novel data sets.Credit card data started coming out.And so, certainly the team that got credit card data first it was like, you know, very easy to pick up money there cuz you're oh my gosh, I understand the revenue of these companies almost perfectly.And then, everyone started getting credit card credit card data, but people still buy it today, right?So, that data is not worthless.I don't know, you know, I think edge is a cultural thing and about can you continue to push yourself and continue to evolve?
And you know, toss this is just another tool.We were misusing machine learning for a long time.Um and it'll evolve how we use it.It'll evolve where we use it.I do think PMs and and analysts will become completely independent from software engineers at some point.And so, yes, it's we're just constantly trying to say can we continue to make use of this to stay at the at the forefront.
Yeah, I think for us our edge is can we it's more on the cultural side.Can we get the investors, technologists, and scientists to work together to unlock new capabilities that only Bridgewater has.And it's it is definitely a challenge.
Yeah, and I I for me I think, you know, technology alone is not a is not a strategy.I think it's, you know, having a good being operationally effective, having the latest technology is like a necessary condition for success, but it's not a sufficient condition for success.And I think it ultimately comes back to your investment philosophy.How can I get the technology to fit my philosophy rather than the other way around?And And I think what we've what I've witnessed a little bit of is you have And we're getting there.I think maybe we'll see some convergence, but on the discretionary side, they're becoming a little more systematic cuz they can about back test stuff pretty easily.And on the systematic side, rather than getting super excited about the technical implementations, you can actually start to be creative, have some, you know, interesting hypotheses that sound maybe like a discretionary manager in some way.
So, again, I'm not saying that's going to happen tomorrow, but it is the the organization is starting to converge a little bit using these tools, which I It was a lot of fun for me.
Yeah, the systematic, obviously, they now have access to more information, which traditionally they didn't they didn't use.And the discretionary side almost are becoming more systematic, right?They're they're they're coding using cloud code.They're back testing.Yeah, that I definitely also see that that a kind of convergence.
the things are hard to back test, though.The controversy of the current quarter is hard to back test if you've never Like you said, financial crisis, okay, that's not in the data set, but, you know, we we I joke with one of the the equity businessheads about like the spicy chicken sandwich wars that happened a few years ago about like Popeyes and all these So, like, someone is going to benefit from this, but how?We've never had these.You can come up with proxies for them, but I don't know that I would trust the T-stat on back testing, you know, that type of thing.And so, there's opportunity, I think, for the fundamental analyst in using the the, you know, these tools to extract some of the insights, but um I don't know.Some of these The humans will still be at the middle of it.
Yes, it's a yes.
And how does the the the new world look like where not just your firms, but actually every firm is is using AI, is using kind of trying to accelerate, and um I think investing is is kind of anunusual area where wider adoption of technology typically makes the technology more valuable but with in alpha wider adoption actually can erode the opportunity.So, do you see AI increasing of course the rate at which alpha is discovered?That becomes faster but also perhaps that that it decays.Is that the world we are going into?
That's a tough one.Now, I I think it depends on the domain.I mean ultimately I ate higher frequency trading other things where there's a lot of data in there.You know, the correlations are very stable over over short periods of time.That could be quite competitive, right?And and having more volume through that but it goes back on the at the more mid frequency side again there's only so much data as I've talked about.So, again it's how do you use the technology?I'm not sure it'll necessarily speed alpha decay faster than it's already been in some way.Um something you're cognizant of but I think you just have to figure out your use case.
And I think Andrew you already mentioned this enough differences between how people interpret data and use tools and
Yeah, and I think if scarcity is your strategy, you're just collecting rent and so eventually that will go away.And so if you can't keep replenishing your alpha in whatever it is, fundamental, systematic, uh you know, I don't think that's new like you said that the these things decay and the decay rate has accelerated over time umfor some class of strategies and so replenishment is the you know, not new for most of most of us.
Yeah, our goal is to uh be in the years two to five x our rate of improvement.And we're not at five x but we're also not Well, I think we'll get there.Um and so definitely the the rate of alpha discovery for firms that are leveraging AI to the max is is going to increase but also that's what makes this fun is like driverless cars the roads don't get harderto drive when you've figured out how to automate driving but in markets it gets it likely does get harder to to extract alpha at that point.
The conversation closes with speculation about what the cutting-edge investment firm will look like in five years. The speakers argue that data, culture, and AI integration will remain central, while technology and investing increasingly merge.
Interesting world we are moving into.Um my final question is is kind of super speculative.It's about um the investment firm of the future.So, uh we we've talked about models, architecture, research process, alpha, all of these things.I want to finish by asking whether over time actually all of this changes the firm itself.And I think both of you have mentioned culture.You also mentioned that similarly.In your view, say 5 years from now, what would be the cutting-edge investment firm of the future?I mean, are you thinking that this will be And 5 years is a long time.Are you thinking this will be completely like AI-native?And and more importantly, would it be materially or structurally different from your firms today?Like it's hard when everything is changing, but I'm curious what your thoughts are on on the firm of the future, cutting-edge firm of the future.
Okay, another another hard one.I think for me ultimately, even 5 years from now, the data is important.So, I think that will always be there, and I think that does provide a bit of a competitive advantage remote from other entrants coming in.So, the AI firm of the future, I think will still need the data.It will I know that the economies of scale and the spending of technology, you could argue there'll be even more consolidation rather than less.That's an argument to be made rather you won't have a bunch of startup hedge funds, I don't think, given the data and and tech spend, but maybe I could be wrong.Um that I don't So, I think the future I hopefully if your your organization in the future looks a lot like your organization today from a philosophical perspective, but they're just faster, more nimble.
You're not getting too hung up on data mapping, security masters, all of that, data bringing in data, which is not fun, but it's all you can do more of the fun stuff.
So, AI takes away all the the pain and the the ugly stuff, but essentially the same firm.
hope.But just for my sake, it's more fun.
I I'm not sure AI-native exists as a word in 5 years.I think it is there's going to be AI-native and dead.Like I don't think there's going to be another path.Uh and it to me it's going to be like spreadsheets.Like companies that were using spreadsheets or not, and then using computers or not, and uh you know, even for the fundamental analyst, are you using computers to analyze their performance and give them feedback?It is just sort of the label doesn't make that much sense to me, but I think it is going to be everywhere.And then I'm curious on I wish we had more time on the startups.But we're seeing all these startup hedge funds popping up that are so AI focused and AI native.I don't know will they survive or not, but it is probably an area that is going to like expand and explode, I would think.
I'm happy to remain off the record on this one, but but but I I do think that the roles will change as I said earlier.I think technology and investment and possibly science as well is going to start to merge a bit.