Alexandr Wang - A Primer on AI - [Invest Like the Best, EP. 272]

My guest today is Alexandr Wang, the CEO and founder of Scale AI. Alexandr founded Scale in 2016, having been inspired to accelerate the development of AI through his work at Quora and his studies at MIT. Specifically, Alexandr realized there was a lack of infrastructure solutions for producing high quality data, the lifeblood for AI models. Today, Scale provides data solutions to leading AI teams at Meta, Microsoft, OpenAI, Flexport, the US Air Force, and many others. This time last year, the business was valued at over $7 billion. Our conversation is a primer on AI. We discuss the building blocks beneath successful artificial intelligence, AI’s role in both the public and private sector, and why data is the new code. We also cover the similarities and differences between AI and software from an investing perspective and what inspiration Scale takes from AWS. Please enjoy my great discussion with Alexandr Wang. For the full show notes, transcript, and links to mentioned content, check out the episode page here. ----- This episode is brought to you by Canalyst. Canalyst is the leading destination for public company data and analysis. If you're a professional equity investor and haven't talked to Canalyst recently, you should give them a shout. Learn more and try Canalyst for yourself at canalyst.com/Patrick. ----- This episode is brought to you by Lemon.io. The team at Lemon.io has built a network of Eastern European developers ready to pair with fast-growing startups. We have faced challenges hiring engineering talent for various projects and Lemon.io offered developers for one-off projects, developers for full start to finish product development, or developers that could be add-ons to the existing team. Check out lemon.io/patrick to learn more. ----- Invest Like the Best is a property of Colossus, LLC. For more episodes of Invest Like the Best, visit joincolossus.com/episodes. Past guests include Tobi Lutke, Kevin Systrom, Mike Krieger, John Collison, Kat Cole, Marc Andreessen, Matthew Ball, Bill Gurley, Anu Hariharan, Ben Thompson, and many more. Stay up to date on all our podcasts by signing up to Colossus Weekly, our quick dive every Sunday highlighting the top business and investing concepts from our podcasts and the best of what we read that week. Sign up here. Follow us on Twitter: @patrick_oshag | @JoinColossus Show Notes [00:03:04] - [First question] - The role that AI and data play in geopolitics and foreign policy [00:07:21] - The end state of a digital arms race akin to nuclear weapons [00:08:53] - Current state of things writ large and how the public and private sectors differ [00:11:33] - The flow and importance of talent when scaling AI and whether it’s more important than software [00:14:29] - His thoughts on how to communicate categories of what AI can do well and what is still a ways out [00:20:18] - The process of creating an AI model and the stages of development [00:27:16] - Principles of building a great engine for gathering data [00:29:04] - The state of technology around annotating data writ large [00:31:31] - What Scale does as a business and their product lineup [00:35:08] - The Storage and Compute equivalents in the AI space [00:37:08] - How Scale fills the gap in producing better and cleaner data [00:39:52] - What Scale will look like in 10 years if their vision is fully realized [00:41:11] - Where AI is in the S curve of acceleration and where AI and software intersect [00:44:32] - Questions to ask about how to incorporate AI and data sets in your business [00:46:23] - What worries him about the proliferation of technology that makes AI more accessible to the masses [00:48:27] - The most interesting AI model he’s ever come across and collapsing the friction between human intent and programmable outcomes [00:51:51] - The kindest thing anyone has ever done for him

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Patrick O'Shaughnessy?

  I know firsthand how complex the tech stack is for asset management firms. And seemingly every new tool and data source makes the problem even worse, adding more complexity, more headcount and more risk. RidgeLine offers a better way forward. One unified platform that automates away the complexity across portfolio accounting, reconciliation, reporting, trading, compliance and more, all at scale. RidgeLine is revolutionizing investment management, helping ambitious firms scale faster, operate smarter and stay ahead of the curve. See what RidgeLine can unlock for your firm. Schedule a demo at ridgeline.ai. This episode of Invest Like the Best is sponsored by Canalyst. Canalyst is the leading destination for public company data and analysis. Founded by a former buy-side analyst who encountered friction sourcing, building and

• 0:43Patrick O'Shaughnessy?

  updating models, Canalyst is now used by over 400 institutions, including the largest money managers globally and by a number of guests on the show. With detailed company-specific models and data on virtually every public company, Canalyst's clients are able to ramp up faster, update models instantly and incorporate the highest quality fundamental data into any workflow. If you're a professional equity investor and haven't talked to Canalyst recently, you should give them a shout. Learn more and try Canalyst for yourself at canalyst.com/patrick. That's C-A-N-A-L-Y-S-T.com/patrick. If you're curious to hear more about Canalyst, stay tuned at the end of the episode where I talk to Canalyst's customer, Brandon Weir, from BWCP to discuss Canalyst in more detail. This episode is brought to you by lemon.io.

• 1:27Patrick O'Shaughnessy?

  The team at lemon.io has built a network of Eastern European developers, ready to pair with fast-growing startups. We have faced challenges hiring engineering talent for various projects and lemon.io offered developers for one-off projects, developers for full start-to-finish product development, or developers that could be add-ons to an existing team. Check out lemon.io/patrick to learn more. Hello and welcome, everyone. I'm Patrick O'Shaughnessy and this is Invest Like the Best. This show is an open-ended exploration of markets, ideas, stories and strategies that will help you better invest both your time and your money. Invest Like the Best is part of the Colossus family of podcasts and you can access all

• 2:08Patrick O'Shaughnessy?

  our podcasts, including edited transcripts, show notes and other resources to keep learning at joincolossus.com.

• 2:18Speaker 01

  Patrick O'Shaughnessy is the CEO of O'Shaughnessy Asset Management. All opinions expressed by Patrick and podcast guests are solely their own opinions and do not reflect the opinion of O'Shaughnessy Asset Management. This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions. Clients of O'Shaughnessy Asset Management may maintain positions in the securities discussed in this podcast.

• 2:43Patrick O'Shaughnessy?

  My guest today is Alexander Wang, the CEO and founder of Scale AI. Alexander founded Scale in 2016, having been inspired to accelerate the development of AI through his work at Quora and his studies at MIT. Specifically, Alexander realized there was a lack of infrastructure solutions for producing high-quality data, the lifeblood for AI models. Today, Scale provides data solutions to leading AI teams at Meta, Microsoft, OpenAI, Flexport, the US Air Force, and many others. This time last year, the business was valued at over $7 billion. Our conversation is a primer on AI. We discuss the building blocks beneath successful artificial intelligence, AI's role in both the public and private sector, and why data is the new code. We also cover the similarities and differences between AI and software from an investing

• 3:28Patrick O'Shaughnessy?

  perspective, and what inspiration Scale takes from AWS. Please enjoy my great discussion with Alexander Wang.

• 3:37Patrick O'Shaughnessy?

  So Alex, we're going to talk about every dimension, I think, of AI, artificial intelligence, machine learning, all the things that it's going to impact. I'd like to structure our conversation from sort of the widest angle down to the most specific, which will probably be around your specific business and product. But given what's going on in the world today, I think it's probably an appropriate place to start with the role that AI, data, machine learning, et cetera, play in geopolitics and foreign policy. We're sitting here in the US, and everyone's mind is on Ukraine. I think everyone understands that cyber is a sort of plane of conflict that exists, but we don't know a whole lot about, to be honest, even I don't know a whole lot about. So I'd love you to give us your overview, given you've got a sort of insider view and seed and understanding of all this.

• 4:20Patrick O'Shaughnessy?

  What role does AI play in the global theater today, in your view?

• 4:25Speaker 02

  A few assumptions to talk through. There are a few things to talk through that I think are really relevant here. I think one is that generally speaking, deterrence has been a incredibly positive thing for world peace and the stable world order over the course of the past forty, fifty years. I grew up in Los Alamos, New Mexico, which is the birthplace of the atomic bomb. The deterrence from the atomic bomb has been incredible forcing function for a stable and peaceful world order. And I think where we're at today in a broad sense is that there's an entirely new set of technologies that are either in early development, being developed right now, or have been recently been developed that are significantly shifting what conflict looks like in general.

• 5:08Speaker 02

  One of these is cybersecurity, which you mentioned, which is a very challenging general attack plane because it's generally speaking, almost impossible to defend against and the attacks are numerous in nature. There's sort of like these shots that can produce a lot of attack vectors. I think AI and machine learning is a very, very significant one. Its ability to intersect with many of these attack planes, whether that is intelligence and understanding what's happening in these various theaters, whether it's through the actual warfare component or it's through cybersecurity, a cross-sectional technology across each of these different fronts. And one general thing that is important to think about is that versus the prior paradigm of conflict,

• 5:50Speaker 02

  where a lot of conflict would happen in literal, physical conflict in wars, in theaters. I think we're moving much more to a paradigm where hopefully ninety percent of conflicts are actually resolved or battled in a purely digital context, long before an actual conflict arises. And so this broad shift, which is, I think is much more than just AI and machine learning, but this broad shift of sort of the physical battleground to a primarily digital battleground is probably the greatest macro shift. And then on a micro level, the ability to have great AI technology is probably one of the greatest enablers of the greatest power contributors in terms of which countries are going to be able to be

• 6:31Speaker 02

  defend the most effectively, as well as to enable the greatest level of deterrence, the broad world order. That's the most practical view from the conflict perspective. And then I think from a geopolitical plane, I think that we're entering this period of what's called great power competition or strategic competition, where there is a small set of highly technically capable countries who are great power competitors, that are strategic competitors on a number of fronts, from economic competition to ideological competition, to obviously a lot of on the field competitions. And so this shift from the paradigm of the past 20 years being counterterrorism, this shift from counterterrorism to great power competition means that the technology itself,

• 7:15Speaker 02

  and in particular being exceptionally good at technology, is really, really critical. And artificial intelligence is just one of those technologies that happens to have compounding order of magnitude effect on any of the other technologies that we could possibly develop. And so that, I think, contextualizes why all this stuff matters. And I think that something that I think is incredibly important, because the way that these world events happen, and I think the Russia-Ukraine conflict is one great example of this, is that they do not happen in predictable ways. There's not a predictable, logical way in which events will play out over time. It is one of these great tail risks, so to speak, to the good and well-being of humanity.

• 7:54Patrick O'Shaughnessy?

  If we think about the analogues between physical kinetic conflict culminating in nuclear deterrence, it's hard to imagine something more powerful than whatever the largest strategic nuclear weapon is against a city or something. Is something similar being developed in the digital sphere? And what pops to mind is the worry that another country or another power reaches some sort of level of general artificial intelligence, to use the common term, first, and that that thing represents the equivalent of being the first and only to have nukes, and maybe it's even harder to copy than a nuclear weapon might be. If we are shifting to digital, what do you think the arms race equivalent is towards? What is the end state that the different powers might be trying to achieve in that digital realm?

• 8:37Speaker 02

  First of all, generally speaking, for artificial intelligence, it's a very multifaceted technology. It's more similar, you would say, to software or the internet than it is to like any specific technology that will be built, a very particular weapon or whatnot. And so I don't quite fully agree with this analogy that an AGI is the equivalent end state to a nuclear weapon. That being said, I do think that AI and AGI, the best way to think about is it enables maybe 10x better strategic decision-making and 10x better ability to actually take action against various tactics that you want to enact. Maybe the best analogy is like playing chess versus a human versus playing chess versus an incredibly talented algorithm or incredibly talented computer. The way in which it'll

• 9:20Speaker 02

  play out strategically dominant behavior in extended periods of conflict or extended periods of time.

• 9:27Patrick O'Shaughnessy?

  How would you describe, based on what you know, the current state of things amongst these great powers? What's the status of the United States, I guess, is the first question? How does that relate to the status or digital capabilities of other countries? And how does public versus private sector look? Is the US private sector far better equipped than the public sector or something like this? What's the state of things?

• 9:46Speaker 02

  I'll talk about AI technology, which is the area that I understand the best. So fundamentally, AI technology, modern deep learning, was initially developed primarily in the private sector through a lot of incredible research that was done at Google and DeepMind and OpenAI and Facebook and these incredible firms, as well as a lot of contemporary research done in China, the clear other major location where a lot of great artificial intelligence research has been done. And a lot of that research, there's been this very fast integration of that into very particular use cases in China. Facial recognition technology, I think, is one of the primary examples where this company is like SenseTime, Face++ etc. that are primarily contractors to the Chinese government, have built world-class computer

• 10:30Speaker 02

  vision algorithms for facial recognition for use by the Chinese government in furthering a bunch of national objectives that they have. We may have questions from humanitarian perspective, like the work being done with the Uyghurs. That all has happened in China. Then at the same time, I think this is starting to change now, but over the past, call it five years in the United States, there's been this very unclear relationship between the private sector and the public sector, or the private sector and the government around the use of AI for applications like defense and intelligence. The Google Project Maven, Google Cloud Project Maven conflict was probably the most visible, most clear example of this. But as a general rule, I think that there's not the same level of partnership between the best and class technologists in the United States as well as the government.

• 11:15Speaker 02

  That is a significant cultural problem that needs to be solved in the United States for us to really be well positioned in the future with respect to artificial intelligence. I think broadly speaking, we're in the very early days with all this technology. The speed of innovation in artificial intelligence incredibly, incredibly fast. I think what we do over the next five years, do over the next 10 years, is probably 10x as important as what was done over the past five years, just given the pace and acceleration in the technology of artificial intelligence. I think the real question for policymakers in the United States, global leaders, for technologists in the United States is what are we going to do over the next five to 10 years? And how does that compare to what other great power competitors will do over the next five to 10 years?

• 11:59Speaker 02

  And that, I think, is what will really set the stage for the next phase of this great power competition.

• 12:06Patrick O'Shaughnessy?

  In the world of software, you always hear about these 10x engineers. And one wonders if in this realm, it's even greater than that. Is the talent at the tippy top of this world going to be accessible to great powers and governments? Like how, this is so important, and there's this multiplier effect of like the absolute best talent. It seems like that's the thing that matters, like competition for that talent, not just between great powers, but also between the top engineer, AI person at Facebook versus going to work for the NSA or something. What's been your sense of the flow of talent and the importance of talent? Do I have this roughly right that it's even more important than it might have been in software?

• 12:43Speaker 02

  There's three critical components to broad scaling of artificial intelligence. One is certainly talent, which I'll get to in a second. The second is compute and computational power, which is impossible to ignore. And especially as we think about a lot of potential disruptions to the supply chain of large data centers or large shipping manufacturing. That's a really important one for us to consider. And then the last one is data and data scale. And that is obviously near and dear to my heart, just running scale, but is also something I think is also impossible to ignore. So I think those are, generally speaking, the three vectors that matter. And there's been a bunch of research that shows that more or less power of these AI systems, both quantitatively and qualitatively, scales on these three dimensions. Then if you think about talent in particular,

• 13:27Speaker 02

  the interesting thing is there probably is a greater scaling effect, but not because the technology itself is one that is more inclined for a brilliant mind than software per se. But actually, it's probably much more a function of the power law impact of various AI systems that is actually even greater than that of software systems. Like, I think that a lot of software systems, because they have to operate within fixed paradigms, they have more capped total impact or total utility that they can deliver. These AI systems, like if you think about the large language models that have been developed recently by OpenAI and Google and others, those algorithms have an incredible ability to adapt to new domains. The cap on the level of impact or the cap on the amount of utility to be delivered is much, much higher.

• 14:12Speaker 02

  And so what that means is that the set of people who can develop these technologies, OpenAI, I don't know what the latest count is, but I think their total headcount is 250 people as a corporate. That's an incredibly small number of people to generate the broad scaled impact that I think they will deliver over a multi-decade timeframe. So I think that's a very accurate point. And this is one of the things that I think, by the way, matters a lot, is high skilled immigration into the United States. America is an amazing place. And I think that one of the things that is differentiated about America is a lot of people want to move to America and want to do their lives work in America. They want to build families in America. And I think we need to do everything we can to enable and encourage that as much as possible. If you think about inputs or you think about the big needle movers

• 14:55Speaker 02

  that are actually maybe easy to change, this is one of the ones that has, again, multi-decade consequences.

• 15:02Patrick O'Shaughnessy?

  I couldn't agree more on that point. Maddening that we don't become just the perfect beacon for all the most talented people. The interesting analogy that I've heard before, just to wrap our minds around the sorts of things or tasks or functions or whatever that constantly improving AI ML models can accomplish. One model that was funny and interesting was like anything that an intern could do for you, you might be able to scale up through one of these models. It's complicated enough that a person's on it now, but it's simple enough that you give it to an intern and it's sort of repetitive. And I always kind of liked that conception. What's your way of thinking about how to communicate to your audience, you know, other businesses using your tooling and just people in general? Like what categories of things AI can do well?

• 15:46Patrick O'Shaughnessy?

  And maybe what categories of things we're excited about, but might be a very long time until AI can do well?

• 15:52Speaker 02

  I think this is one of the general misconceptions about AI and machine learning, which I think causes a great deal of fud, which is that the intuitive belief is that the things that are easy for humans to do are going to be the things that are easy for AI and machine learning to do, which is absolutely not the case. And the things that are easy for algorithms to do are relatively orthogonal, frankly, to things that are easy for humans to do. One simple example here is I think that it's going to be a very, very long time before we have home robots that can do things like fold your laundry and do your dishes, but a much shorter time span slash, you know, I think this is already today where you can have artificial intelligence systems that are world-class copywriters and can write better rhetoric, better words than most people ever could. There's probably a few frameworks I would assign to this.

• 16:37Speaker 02

  I think in a broad, general sense, one way to think about the potential impact or lower bound the potential impact of artificial intelligence is kind of, as you mentioned, which is the ability to scale repetitive human tasks. So take repetitive human tasks, instead of going from zero to one, go from one to N human work. And I think that this is a generally amazing thing to happen because I think that humans, for the most part, don't enjoy repetitive tasks or generally find those relatively unpleasant and find it much more exciting to be creative and to constantly be creating. This ability to scale human tasks from one to N is going to be this incredible, not only economic good or economic enabler for the world, but also going to be a significant enabler for humans to be more leveraged, more happy, more creative, et cetera.

• 17:22Speaker 02

  So I think that's one way to sort of contextualize the broad impact that AI can have. And there's a bunch of other nuances, which I'm sure we'll get to. If you think about what tasks humans are good at versus what tasks algorithms are good at, generally that more or less boils down to data availability, which is that where there are large pools of digital data, an algorithm can learn from, and those pools of digital data either have been collected in the past or easy to collect in the future, those are going to be the problems which algorithms can do effectively and can learn to do effectively. And then areas where there does not exist digital data and is expensive to collect this digital data, those are going to be the last things to be automated. So a great example is if you look at GPT-3 and these large language models, the real

• 18:06Speaker 02

  secret behind it is that it leverages two decades of Reddit data, which is two decades of humans using the internet and basically typing language into the internet in various forms for decades and decades and decades. And that is the pool of digital data that it used to be able to do all these incredible things and writing long-form text. Then if you think about this parallel that I mentioned around home robot, there's so little data about actual capture data of let's someone folding a shirt or somebody folding a towel or going around and doing chores. The ability to actually collect and produce that level of digital data necessary to produce algorithms that can understand that and actually perform those tasks is an incredibly,

• 18:50Speaker 02

  incredibly, incredibly hard road. This extends, by the way, to things that are like really unintuitive. So for example, DeepMind and OpenAI very recently released algorithms, some of which are very good at, DeepMind released an algorithm that's very good at competitive programming, OpenAI released an algorithm that can prove very difficult math problems or math theorems. And these are both things which are very, very challenging for humans to do, very, very premium skillsets as far as humans go, but there are incredible pools of digital data as well as abilities to verify or simulate the outcomes here, which allow these algorithms to perform actually incredibly, incredibly well. There's this very interesting process by which artificial intelligence will slowly

• 19:33Speaker 02

  automate or meaningfully change what human jobs that are primarily digital in context will look like. And then a lot of the physical work will, I think, be generally untouched for a very long time.

• 19:44Patrick O'Shaughnessy?

  In many ways, if you're right, the whole idea of blue collar work, let's call it, or something like that, being in jeopardy is maybe completely wrong, that the white collar work, the sort of knowledge work that's mostly digital in its form today is more susceptible. As you said, this should be a good thing, right? It should open up human leverage to higher order, more creative, more interesting tasks, non repetitive tasks. I think as of this morning, jobless claims are at their like all time low or since the 1960s or something. So all this great technology doesn't seem to have affected things that badly. But is that right? That some of the jobs that we thought might be automated first by robots or something actually might be the last things to get automated by this model and the more digital knowledge work stuff maybe first?

• 20:26Speaker 02

  It's a strange thing, but I think that probably is the case. Maybe one simple mental model is if you spend all day in a word processor or all day in Excel, the actions that you take in those products are some of the most likely to be automated, frankly. And for the most part, I think a lot of that work is probably not very inspiring. And so it will be this incredible enabler of leverage to think bigger about those kinds of jobs.

• 20:50Patrick O'Shaughnessy?

  Maybe it makes sense to help people understand the process of creating one of these models in the first place. I think the discrete steps, let's say the outcome is a model that makes a useful prediction. Ultimately, this is all predictions. That's sort of what's being generated by the models in the first place. I don't know where to start, whether it's with raw data or annotation of data. And we're starting to get into what scale now provides for companies. But how do you think about explaining the discrete or the important stages of building one of these models in the first place? I think just understanding that architecture will let us dig into each piece a little bit more.

• 21:24Speaker 02

  Again, everything starts with the data. You know, I often will analogize the data for these algorithms as the ingredients that you would make a dish with, or the ingredients that you make something that you'd eat with. It is incredibly, incredibly important. We often say this thing, which is data is the new code. If you compare traditional software versus AI software, in traditional software, the lifeblood is really the code. It's the thing that informs a system what to do. In artificial intelligence and machine learning, the lifeblood is really the data. That is certainly like one major change. That's really important. The lifecycle for most of these algorithms is a fewfold. So first is this process of collecting large amounts of data. By collecting, it could be data that is already sitting there. There's a lot of software processes that already collect a bunch of data.

• 22:09Speaker 02

  There's a lot of cameras in the world that already collect a bunch of data, but you need to like get the raw data in the first place. Then it goes through this process of annotation, which is the conversion of this large pools of unstructured data to structured data that algorithms can actually learn from. And so this could be, for example, in imagery or video from a self-driving car, marking where the cars and pedestrians and signs and road markings and bicyclists and whatnot are, so that an algorithm can actually learn from those things. It could be, for example, in large snippets of text, actually summarizing that text so that an algorithm can understand and learn what it means to actually summarize text. So whatever that translation is from unstructured data to a structured format that these algorithms

• 22:53Speaker 02

  can learn from, then it goes through a training process. So these algorithms basically look through these reams and reams of data, learn patterns, and slowly train themselves, so to speak, to be able to do whatever task is necessary on top of the data. And then you launch one of these algorithms in production and you run them on real-world data and they're constantly producing, as you mentioned, these predictions. The very important piece is this is not a sort of like one-way process. This is actually a loop. And if you look at almost every algorithm that is launched out there in production, it is not a sort of you build the algorithm and then you're done, because these algorithms are generally very brittle. And unless you're constantly updating them and maintaining them, they will eventually do things that you don't want them to do, or they'll eventually perform poorly.

• 23:37Speaker 02

  There's this critical process by which you are constantly then replenishing them. You're constantly going and recollecting new data, annotating it, training the algorithm, launching that new algorithm out to production. You have to constantly undergo this process to create very high-quality algorithms.

• 23:52Patrick O'Shaughnessy?

  I want to make sure that this interesting point you made about data being the new code really hits home for people and maybe even put that in a business context. So if the IP or the mote of a software company is this code base that takes a very long time to develop, has all sorts of dimensions to it, maybe it's microservices, maybe it's some code monolith, it's unquestionably like an incredibly valuable asset. It's digital, but it's an incredibly valuable asset to the company. And you're talking about, I think, a transition where it's something different, where maybe, I don't know, maybe Google's data repository or something is this unbelievable advantage that they have because no one else has access to all of their data. Is that kind of what you mean that ultimately maybe something like Google, their data is

• 24:34Patrick O'Shaughnessy?

  worth a lot more than their code base and that that will become a trend that we see sort of across industries?

• 24:40Speaker 02

  If you look at the highest performing algorithms across a variety of different domains, image recognition and speech recognition and summarizing text and answering questions of text, so these very different cognitive tasks, look under the hood, they actually all use the exact same code base. That's been this very meaningful shift that's happened over the past few years in artificial intelligence. And so we're at this point where the code has become effectively the same and more or less a commodity, so to speak, when it comes to artificial intelligence and machine learning. And the thing that enables the differentiation is really the data and the data sets that are used to power these algorithms. To your point, if you think about, one of the ways that we talk about in a business context is if you think about what is your strategic asset?

• 25:25Speaker 02

  In general, in business, your strategic assets are the things that allow you to differentiate yourself against your competition. In a world where 99.99% of the software in the world is sort of traditional software and then only 0.01% is AI software, then you care the most about your code. Your code is what will differentiate your product versus your competitor's product or your processes versus your competitor's processes, et cetera. But then as more and more of the software in the world's written infused with AI, using AI, or over time, the interfaces shift to AI interfaces, an Alexa-like interface, for example, as that shift happens, as you go from Nine nine put nine nine to ninety ten or eighty twenty or even fifty fifty over

• 26:07Speaker 02

  time, the vector of differentiation totally shifts to data and the data sets that you have access to. And so that means is that your strategic differentiator, to your point, as a firm, is going to be primarily based off of what are my existing data assets? And then what is the engine by which I'm constantly producing new insightful differential data to power these core algorithms that are effectively powering my business? And these algorithms at the core that will power the future of business, I think, are relatively core. I think there's definitely algorithms around automating business processes that are going to result in significantly more profitable firms over time. There's going to be algorithms that are based around customer recommendations and customer lifecycle,

• 26:49Speaker 02

  which is a lot of the algorithms that we've seen to date. Imagine TikTok recommendation algorithm, but for like every economic interaction or every economic transaction in your life that is constantly identifying the perfect next thing that you may want to transact with. And that is going to exist across every firm or every industry is basically going to have to build their version of that. And that's going to result in like significantly more efficient trade. The long-term impacts of that you could think of as like a general reduction in marketing expenses or sales and marketing expenses, because the algorithm just does a better job at knowing what the user wants to do next than having to do all this marketing and all this very active sales. There's a lot of very real changes to, I think, the physics of what the best businesses will look like in, let's say, a decade or

• 27:34Speaker 02

  two decades or three decades that come from artificial intelligence. And if you think about what will allow me to do these things better than someone else, it's the quality, efficacy, and volume of the data that is used to power these algorithms.

• 27:48Patrick O'Shaughnessy?

  If getting to that state of some sort of differentiated data advantage is the goal, then that engine that you referenced becomes a critical idea for any company, a piece of infrastructure. What principles are there about building a good engine for data gathering that may be cut across different types that you've observed? What's behind the great engines that you've seen for gathering data?

• 28:11Speaker 02

  There's a few tenants. I think first, quantity is a quality of its own. You want to just have lots and lots of data coming in. That will be a differentiator, no matter what. Not all data is created equal. And so you want, in general, the way these algorithms work is they're sort of like needles in the haystack, almost in most of your data, that end up being differentially valuable. You'll want to develop a process by which you can actually, what's called curate, but identify and understand what are the really valuable pieces of data in general data inflow that we're getting. You'll want strategy for data diversity, or you'll want a strategy by which you're not just going to keep collecting the same old data, by which you're going to constantly be expanding the domains, or the kinds of data, or the diversity of data that you're going to be collecting over time. And that's a really interesting one.

• 28:55Speaker 02

  Like, I think that one of the more intuitive examples that I think people understand the best is that Tesla's autonomous vehicle strategy, one core part of their strategy is that they have, all of the Teslas in the world are, in some sense, collection vehicles, they're all encountering very rare situations, but then improve and help the machine learning algorithm. There is a truth to that idea, which is that because not only they have a greater volume, but they also have great systems to identify these needles in the haystack, that's a strategy by obtaining an incredibly diverse dataset over time, that's very beneficial to the long-term differentiation of their machine learning stack. By the way, Tesla has other challenges as well, so this is not a clearly dominant strategy, but there is truth to some of these ideas.

• 29:36Patrick O'Shaughnessy?

  What should we know about the state of technologies around annotating data? That I know, obviously, Scale is deeply involved with. The example in Ukraine around like satellite imagery, there's this interesting jump from like raw data or raw input into some sort of refined signal or piece of information or whatever. I don't know what the right terminology is. What is the state of that in the world today? Like, has that changed a lot? Has that been pretty constant over the last few years? Where's it going?

• 30:03Speaker 02

  I think it's meaningfully changed. It's one of these very interesting problem domains and one that we find incredibly exciting because it's undeniably a quote unquote charity problem, or it's undeniably a very complex, multifaceted, operational challenging problem. Historically speaking, this has been done in an incredibly low tech way and in incredibly very manually intensive way for the AI industry to date. And a lot of that's been because it hasn't been a problem that technologists have wanted to dedicate themselves to. For the most part, most people who go into computer science, most people who go into artificial intelligence, by definition, they go into computer science because it's a significantly more constrained problem that doesn't involve these complexities around human operations.

• 30:47Speaker 02

  But I'm proud of the scale team for having really done a lot of this work that I think will be seminal, is starting with a very manually intensive or manually challenging process that is operationally very intensive and then adding meaningful amounts of automation, whether that's algorithms that will do a lot of the work before it gets to people, or algorithms that are able to identify most of the errors that people might make, or in general, adding, injecting a lot of operational efficiency processing into this process to enable significantly better outcomes and either more efficient processes or more higher quality data. That is really the core engine behind what we've built at scale that has enabled some of the biggest platforms in the world, such as Meta, or Microsoft, or many of these other large tech firms, to significantly scale a lot of their machine learning efforts.

• 31:31Speaker 02

  One of the companies that we really look to for inspiration, we've been deeply inspired by, is what Amazon has done in two different spots in their business. They've done it not only with logistics for e-commerce, but they've also done it with AWS. They've taken both of these very large, complex, half operational, half technological problems, and they've added significant amounts of technology and automation and operational efficacy to result in orders of magnitude better performance than you could have accomplished in the past without sort of this embrace of the messiness of the problem.

• 32:03Patrick O'Shaughnessy?

  A good excuse to talk about literally what scale does more than an hour in. And it got into the business itself, a whole bunch of amazing context and lessons so far. What are the building blocks? I'm struck that if you go to the little products tab on Scale's website, it looks not dissimilar from the AWS tabs of old, where there's lots of individual functions that are part of this value chain, if you will, of building one of these things that you can quote unquote higher scale to do. Just talk us through the business itself, what it does for customers, and how that product lineup fits together. If you are the retail operation infrastructure equivalent, or the AWS equivalent for the AI future, what does that look like today?

• 32:44Speaker 02

  Where we started was really this problem around data annotation. Because what we noticed is that it was this massive bottleneck for the overall progress of machine learning, is that most firms could simply not get high enough quality data sets for them to actually build great algorithms. That was very much the initial problem that we set out to solve, which is how do we enable companies to get better data to fuel better AI at its core? And we started by doing that with some of the most technologically sophisticated firms in the world. You know, we work with, as I mentioned, Meta, Microsoft. We work with large automakers in building autonomous vehicles like Toyota and General Motors. We work with OpenAI on a lot of their cutting edge research. Work with large internet companies like Instacart, or Etsy,

• 33:28Speaker 02

  or Flexport on a large scale machine learning problems. And then through doing that, we've built what we really believe to be the best engine in the business, or the best engine in the industry, around producing these very high quality data sets. And a lot of our view in expanding beyond that has been, that primitive is incredibly powerful. And then integrating that primitive with other components of broader machine learning lifecycle that I was referencing before, enables us to produce really powerful products. Just like how for AWS, you know, they started with the primitives of storage and compute. And then they sort of combined those primitives in many interesting ways to produce like these incredibly powerful products like RDS, or their DNS products.

• 34:13Speaker 02

  Or the list kind of goes on and on to solve customer business problems. Really look at the same way. We have built this core primitive around data annotation and the production of high quality artificial intelligence data sets. We can combine that with other primitives that we've built such as data management products or model monitoring products or algorithmic development products. And we combine all these primitives and producing these products that solve really core customer problems, whether that's in the government space, working with defense and intelligence agencies and building really high quality artificial intelligence algorithms to solve core national security problems, or that's in solving, working with large, complex Fortune 500s who have

• 34:52Speaker 02

  incredible amounts of unstructured data, such as documents flowing into their business, and we build automation to support a lot of that document processing. Or we work with large map builders and we use artificial intelligence and human processing to enable this map creation. One of the great lessons from Amazon as a business is that parallel execution is incredibly powerful. The traditional business logic or the traditional business tourism is around, hey, you should really just focus on a few things and do those things really, really well. And that's your sort of like strategic edge. Amazon really threw that out and became very focused on parallel execution while focusing on what are the primitives that are going to enable them to build a lot of products very successfully. And we take more of that approach, which is we've built incredible primitives. We're going to make those primitives better and better and better.

• 35:36Speaker 02

  And then we're combining these primitives into building best in class products for our customers.

• 35:40Patrick O'Shaughnessy?

  One of the things you'll hear about Amazon Web Services specifically is that at the end of the day, what they're really trying to drive people towards is more S3 and EC2, storage and compute spend. Those are like the most basic building blocks and a lot of the other services that AWS provides effectively maps back on to increasing the volume of storage and compute as a business model. Do you think that that's true for you all too? And if so, like what are the S3 and EC2 equivalents per scale?

• 36:08Speaker 02

  We think that's true to an extent in that, again, we're almost religious in this belief that better data results in better AI and that data is a new code. And so we really think about is like, how does scale become the long-term infrastructure provider to powering more or less the global expansion of data to meet the needs of AI globally? That I think is certainly one very large driving factor. And a lot of our goal is to build the most data-centric AI infrastructure platform out there. But at the same time, I think that maybe one of the things that we really index on as a company, and I think that some Silicon Valley companies do this, some don't, but we would certainly take this approach, is to really index on the customer value creation, which I think is really important in the space of artificial intelligence, machine learning, because frankly speaking, this has been a technology that people have been

• 36:53Speaker 02

  talking about for a very long time. And then the actual value creating use cases are few and far between. There's a lot of distrust of AI systems. There's a lot of belief AI isn't good, or AI is so bad, or AI is just this snake oil technology. That is absolutely not true. Artificial intelligence is incredibly powerful technology, but there's been an incredible dearth of firms that are focused on how do you index against the actual customer value and the customer outcomes they're able to generate using these technologies? On the one hand, we probably do believe this large-scale growth in the primitive belief around large-scale growth of powering these data sets for machine learning. We also believe on the other end of working very hard to power as big or as meaningful

• 37:36Speaker 02

  of customer outcomes as possible, and that being a primary metric on which we value ourselves.

• 37:41Patrick O'Shaughnessy?

  If we zoom in on the original use case, helpful annotation of data, what does that mean in a literal sense? If the outcome is well-labeled, clean, large data set, and the input is whatever, like, what is the gap that scale was filling? What is the function that it was actually doing to help the company produce more cleaner, better data?

• 38:04Speaker 02

  This is sort of the state of the market that we saw when we entered it. Even when we had started, there were a large number of very talented machine learning teams that were solving varied problems from autonomous vehicles to building AI for e-commerce, to building sort of voice assistants and systems. There were a lot of interesting use cases of the technology. Each time we'd go to one of these firms, we would ask them, what are your biggest problems? Always in the top three, data quality and data volume is one of them. And you dig into it and it's because they have this incredible challenge of motivating their incredibly brilliant, smart ML engineers and ML scientists and working on dirty problems around data and data quality. And so these gaps are technological, some of these gaps are cultural. I think for us, we really view as some mix of both.

• 38:48Speaker 02

  But what we did is we went into all of these firms. We took incredible religion and care to build great technology to solve data quality, data volume bottleneck. Those really limiting the efficacy and power of their machine learning models at the end of the day. And so we built, we took, I don't know, maybe the Stripe playbook is what you would say it was, or the AWS playbook before them. We built great developer APIs and a great developer platform that enabled the machine learning engineers and machine learning scientists to basically hit an API, get as much high quality data as they needed or they wanted. And behind the scenes, we would do incredible amount of dirty work to make this possible and actually enable them to build these great algorithms.

• 39:27Patrick O'Shaughnessy?

  Maybe just to nail the point home, like a couple of examples of these dirty jobs might be the categorization of images or subcomponents of an image or something like that. Would that be like a good example?

• 39:36Speaker 02

  Yeah, great examples are like given a huge amount of imagery or a huge amount of data off of a robot, whether that's a self-driving car or some other robot marking and all of that imagery. What are the things that are really important for that robot to see? So enabling the robots to see in the first place. Other similar quote unquote dirty jobs are for financial services firms. They have incredible amounts of transaction data. A lot of this transaction is fundamentally unstructured. So understanding who are the vendors in these transactions? What are the locations? What are the other pieces of insight that you can pull out of these transactions? And other ones around other forms of audio data, for example, and pulling out the relevant intent and understanding from audio data. So all these dirty jobs, they fundamentally start with some data format that machines can't read today

• 40:21Speaker 02

  and making them effectively legible for these machine systems.

• 40:25Patrick O'Shaughnessy?

  If I fast forward and think about scales, future five, 10 years down the line, how would you describe the absolute best case of success? So if the mission is accelerate the development of AI, make it easy for companies to build more AI models effectively, which I think is kind of the core mission of the firm. What is the best version of that look like? Do you think five or 10 years from now?

• 40:46Speaker 02

  I mean, I think we really take the view that it just comes down to the customer outcomes that we're enabling. So rather than thinking about what scale exactly looks like in 10 years, I think the right thing to look at is in 10 years, are the majority of companies in the world able to effectively use artificial intelligence in some very high impact application to solve a business need? And is that application in some way or another powered through scale products? And we take a very open minded view is we're going to build lots of products over time that can power these products in different ways. And the delivery mechanism of the algorithm, we may not build the exact product that a Fortune 500 uses.

• 41:27Speaker 02

  We may be powering some SaaS company that is integrating AI into their product that the Fortune 500 uses. We basically want to take this very long term platform view that as long as we're there powering decade long massive expansion of AI and machine learning, we're going to be really pleased and happy.

• 41:44Patrick O'Shaughnessy?

  If we zoom out and go to the more market side of things and put our my investor hat on and think about what drives enterprise value, value creation, the things that investors ultimately care about when they're putting money into a business, they want to get a lot more money out. The world of software has obviously been a center stage for seven, 10 years now because they've tended to be very scalable, fairly high margin, incredibly fast growing businesses. And the word that you never want to hear as an investor is deceleration in the growth world where maybe they're reaching saturation points and software is no longer a new thing. It's a fairly mature thing. How do you think about, you mentioned this concept of thinking about like an S curve? And maybe we're for software approaching the diminishing part of that S curve. Where is AI in that same thing?

• 42:29Patrick O'Shaughnessy?

  And how might these two things intersect to form lots of new enterprise value in the future if software becomes overly saturated?

• 42:37Speaker 02

  One thing to think about software for a moment, the sort of alchemy or the magic of software is that A, you're able to collect very large scale data sets in a very coordinated way. B, that you're able to build simple workflow tooling on top of these data sets. Think about your traditional CRM or frankly, the majority of SaaS tooling is workflows on top of these data sets that enable business value. And then three is basically infinite scalability of a lot of these systems. These are some of the technological primitives that have enabled SaaS, broadly speaking, or software in general, to produce a lot of value for most enterprises. But these primitives or these forms of alchemy have some cap. That's sort of the saturation of software that you're mentioning. Well, then if you think about AI technology and you use this mental model that I mentioned

• 43:21Speaker 02

  before, which is the fundamental promise of AI technology is that you can take repetitive tasks that people are doing. You go from one to N with those repetitive tasks. So you can automate the Nth repetitive task rather than relying on humans for that. Well, if you look at the majority of Fortune 500 business or the majority of largest enterprises in the world, there are an incredible number of parts of their business where they spend enormous amounts of money on large teams of people to do repetitive tasks. The alchemy that is possible there is not only the automation of meaningful parts of that work, but also the ability to even go further than even the best trained humans could do in many of those tasks.

• 44:01Speaker 02

  The sum value of that potential economic value, or this TAM, so to speak, of AI and machine learning is just absolutely astronomical. And I think that that is at minimum 10x, probably 100x, the total business value that has been generated by SaaS systems or software historically. I think if you think about it, you have this one S curve of the saturation of software. And then there's this very, very early S curve that is being developing right now around the privatization and productionization of large scale AI systems, let's say in the enterprise or let's say across businesses. And the real question is, okay, what's the pacing of that S curve versus the pacing of the saturation and deceleration of the current software S curve? And I'm an optimist.

• 44:46Speaker 02

  In not too long, we're going to have a massive proliferation of AI use cases within the enterprise that are going to be way more impactful than the use cases of software in the past. And the way you'll see that, the business ROI's generated from high quality AI systems are going to be 10x more than the business value generated by, let's say, deploying a CRM or deploying ERP system.

• 45:06Patrick O'Shaughnessy?

  How would you advise those listening who maybe run businesses and are nodding their heads and think, yeah, this all makes sense. This is a new competitive plane or competitive frontier. And I want to make sure I'm not left behind. But I'm not a data scientist. I don't have this in my background strategically or tactically. What would be the questions that you would have them ask themselves to start incorporating this thinking into their business?

• 45:30Speaker 02

  I think the big questions are really around understanding the data that powers your business and understanding at a meaningful level, what are these potential data sets within your business that can fuel a lot of this future wave of artificial intelligence and machine learning? And then I think it's really thinking about, and I think this is probably in partnership with partner or in partnership with advisors, is thinking through, what are the highest value business problems that artificial intelligence could solve that would meaningfully move physics of my business? I kind of mentioned a few of them, but one that I think applies to nearly every business is this problem around customer recommendations or basically building better recommendations to empower customer life cycles. Another one that I think powers almost every business is taking some of the

• 46:14Speaker 02

  most expensive repetitive process internally and figuring out ways to automate or make those more efficient. And I think it's thinking through what are these core business use cases. One very tactical piece of advice that I would give is, one of the challenges of AI is we're going to have this huge shortage of human capital that is trained in AI and machine learning and deep learning for a very long time. I think that's going to be big bottleneck for many decades, frankly. And so I think it's important that you identify, for most business owners, you identify partners who have some of that human capital, have a lot of expertise and experience to help guide you through that process, whether that's scale or whether that's another company. It is important to identify these business partners who can help accelerate journeys here.

• 46:56Patrick O'Shaughnessy?

  Yeah, I love the human capital angle. I mean, we're still short traditional software developers and engineers, and we're into that cycle a long way. So can imagine this is that on steroids. What if anything worries you about the proliferation of all of this technology, maybe even the technology that you yourselves are developing that allows faster development of AI models? I'm of the view always that technology is sort of like a morally neutral. It can be used for good or bad. I like to think that it's mostly used for good. It seems that way, based on outcomes over time. But what is concerning to you about the potential or the leverage that this might give people that shouldn't have it?

• 47:34Speaker 02

  One major one that actually keeps me up at night is this thing that we talked about at the very beginning, which is AI in the context of geopolitics and great power competition. I think that there is an incredible potential for AI to either rapidly accelerate or result in bad outcomes for some of the conflicts for the next, call it 50 years. And so I think that that is one very pointed, specific use case of the technology that does keep me up at night. In addition to that, I think there's these two competing curves in AI. It's interesting to see how they'll play out, but one of the things that I think generally technology, concerned about any sort of technology accelerating is inequality in the world. And I think that it's not totally clear with AI what

• 48:14Speaker 02

  the exact effect is. Intrinsicly, the technology benefits from scale. The bigger and bigger your models are, the bigger and bigger your data sets are, that results in you building better AI systems. Generally, technologies that favor scale do result in greater levels of inequality. And I think that's something that we really need to watch out for. At the same time, that's why the democratization of the technology matters a lot. And that's also why the building it into use cases that affect a lot of people and enable a lot of people to live better lives is also really critical. And so, again, there's sort of these two competing curves in the technology. One is the benefits to scale and the other is lowering of fixed costs and the democratization of the technology. And I think we need to be very mindful about how these curves intersect over time or what the directions look like and the relative speeds.

• 49:00Patrick O'Shaughnessy?

  What is the most interesting AI model you've ever seen?

• 49:03Speaker 02

  I'll nerd out a little bit here. I'm predispositioned to like these scientific applications, but both AlphaCode and AlphaFold are deeply interesting machine learning algorithms. I think that AlphaCode is, it's almost like brain breaking in terms of thinking that you have these machine learning systems, they can solve these competitive programming problems better than the median competitor. And the competitors who compete in these things are already incredibly high up on the distribution of humans who can do some of this work. Beating the median competitor is probably beating 99.9% of humans at these programming tasks. And that's just insane. You know, if you'd asked me even five years ago, if I thought that was going to be possible, I would have said, no, I don't think that AI systems are going to be able to reason and think through these complex, creative

• 49:48Speaker 02

  problems well enough. But here we are, again, the availability of the digital data, we have systems that can perform that well.

• 49:55Patrick O'Shaughnessy?

  What's an example of that? What would be something that AlphaCode is producing that a very talented human competitor is also producing so that we can compare them?

• 50:04Speaker 02

  There's these classic programming questions, almost like the interview questions that a software engineer would be asked in an interview process. So classic brainteasery algorithm questions that for decades we have used in Silicon Valley to test incoming engineers for their talent. That's something that the machine learning system, AlphaCode, can perform just as well as some of the best engineers out there. So another way to look at this, if your Turing test is a set of programming interviews to get a job at Google, you probably have a machine learning system now that can pass that Turing test. It's a pretty shocking thing. And I think that this is more philosophical, but I think that myself as a programmer, I think

• 50:44Speaker 02

  a lot of programmers take pride in their role as sort of the master of the machine, so to speak, or the person who tells the machine what to do and programs the machine. It's this funny role reversal, but the machines are now better at a lot of that work than humans.

• 50:58Patrick O'Shaughnessy?

  What are the implications of that? As you said, it's kind of hard to wrap your mind around. Does that start to mean that we just continue to collapse the frictions between, I'll call it like, human intent and programmable outcomes? It's almost like our imaginations become the limit rather than the elegant work?

• 51:16Speaker 02

  Yeah, that's exactly right. This is like a broad trend within software. But I think, to your point, the beautiful end state is that we as people, we're just going to be able to dictate into a machine learning system what kind of software we want it to build, and it'll build something that basically accomplishes that exactly. So we can all be product managers effectively in our own mini fiefdoms. That future state, I think, is like reasonably far away, maybe not infinitely far away, but reasonably far away. But we're already seeing this where, take these GitHub Copilot systems, they meaningfully accelerate the programming task. There's this thing that's going to happen where programming and product management, in some sense, are going to keep converging, where differential task or the differential skill is in understanding what to build rather than actually being able to build it in the first place.

• 52:01Patrick O'Shaughnessy?

  It's sort of like the classic idea in machine learning that it's the label. It's the outcome that you're targeting that requires all the imagination and creativity, not the features that you're using to predict that outcome. Questions become more valuable than answers, I guess, is the other way of saying what you said, which is just, is very cool. It unlocks the core of, I guess, what makes us human, which is pretty exciting to imagine. I hope I get to live to see a lot of that happen. Alex, this has been so much fun. I love this topic. I am so interested in these technologies. I think, you know, I asked the same traditional closing question of everybody. What is the kindest thing that anyone's ever done for you?

• 52:35Speaker 02

  It's not one specific action, but I think that my first violin teacher, her name was Kay Nunam, I don't know if she'll ever listen to this, but maybe she will. Over the course of many years of working with her, I think she really instilled a few things in me. I think first she really created a love and joy from art and creative activity. And she herself was just so incredibly joyous and very inspired and clearly loved very deeply the art and process of the creation of great music and great performance, which was just incredibly infectious. And I think one learning is just that passion is incredibly infectious thing. There's very specific moments that I remember really quite vividly, which is, there's this period where I wasn't practicing very much and I thought I could sort of like skate by

• 53:20Speaker 02

  or she wouldn't notice. And obviously it's plainly obvious to her. At one point, she just ended the lesson, you know, I remember two minutes in, and she said, if you're not practicing at all, we should stop wasting your parents' money and just stop on lessons. It was like this shocking thing. I think I was in sixth or seventh grade at the time, and so I was shocked to have this something so direct, be confronted with something so direct. But I think what it really showed me, if you're going to do something, you have to do it with full passion and full force. And if you're not going to do something with full passion, full force, then it probably just isn't worth doing in the first place. And this has been really a driving belief behind a lot of my life. I wrote this blog post a while ago called Hire People Who Give a Shit.

• 54:02Patrick O'Shaughnessy?

  I remember reading it, yeah.

• 54:04Speaker 02

  Yeah, it just becomes so core to, I think, how even higher at scale or a lot of what we do at scales, don't half-ass. You gotta do things with full force and full passion.

• 54:12Patrick O'Shaughnessy?

  Absolutely love this story. One follow-up question, I remember the post well and it was excellent. How do you do that? What is the most effective way to understand not having yet worked with somebody if they're someone that gives a shit?

• 54:24Speaker 02

  One of the funny things about giving a shit, you notice it along the edges, so to speak. You notice it in maybe the depth of thought, the depth of obsession, the attention to detail along the edges of something they really care about. I think one great example, let's take music, for example. When you're preparing a great performance, you will start noticing just the smallest fricking things and you will just stop yourself and like keep practicing until you're like impeccable. All the mini flourishes and all the mini components. And the way that you notice a great performance is not, do they get the notes right? But it's like, how like impeccable and how clearly perfect is each little detail? That's one general thing is like, I think you notice it along the edges more so than just looking full frontal at the thing itself. You can notice it by talking to people about like,

• 55:10Speaker 02

  what were the little details that you really paid attention to? Or what are the things that really bothered you about how something was being done? Or what, what were the things that took 80% of the time that other people wouldn't notice? That's one big thing. You'll notice it just from like the passion in people's voice and how they care. And some people express that more outwardly. Some people are maybe more reserved, but you'll notice in just the care and sort of the attention by which they express those things. And then the other thing I think is from my time meeting lots of people and talking to lots of people and learning about humans in general. There are people who are generally very inspired and there are people who are generally less inspired. And the people who are more inspired, you'll notice this like theme in their life. They find things that inspire them and they find the next thing that inspires them. They find the next thing that inspires them and inspires their full force and will.

• 55:55Speaker 02

  And I think you can notice that pattern a lot when people just talk through their lives and what they've done in the past. And sometimes it's a very wide set of things. People can be inspired by punk music when they're a kid and then they become inspired by DCFs and investing when they're an adult. But I think that just noticing that strength of will in these instances is really critical.

• 56:18Patrick O'Shaughnessy?

  I think it's such a wonderful, interesting place to close that giving a shit manifests in the details and at the edges. I love that concept. I think it's so true. If you ask yourself the things you're proud of producing, it's very easy to answer that question, like what were the details that took all the time? And when you can't answer that question, it's probably indicative that you did a just okay job at it or didn't care that much. I just love it as a closing thought. Alex, this has been a total blast. I've learned a ton. Thank you so much for your time.

• 56:43Speaker 02

  Yeah, thank you so much. As I mentioned when I first chatted, I'm such a fan of the podcast. So I'm excited to be on here.

• 56:49Patrick O'Shaughnessy?

  Pleasure's all mine. This episode was brought to you by Canalist. In this three-part mini-series, I sit down with Brandon Weir, founder and portfolio manager of BWCP, a fundamental-oriented TMT consumer hedge fund manager based in Dallas, Texas. Here, Brandon's biggest lessons from launching his fund, their unique blended investment strategy, and how BWCP has integrated Canalist into the investment process since day one.

• 57:14Patrick O'Shaughnessy?

  So Brandon, the best place to begin is with your personal origin story and background. What were you doing prior to BWCP? And then we'll start to talk about how that experience led to the founding of the firm.

• 57:24Speaker 03

  We can take it all the way back. I got into investing in a strange way. I'm from Kansas City, middle-class background. My parents made me have jobs, and the only one I could stomach at the time was mowing yards. And I started a business when I was probably 14. It turned into a passion for investing. And the reason that was is I noticed early on, when I was able to drive, you should drive to the nicest areas you can find. They had the best profit margins on mowing. And I got to know some of my clients that I was mowing yards for, and they were young guys that were obviously doing very well. And I asked them a couple of times what they did, and they told me they worked for American Century. And I didn't know what American Century was. I'd never heard of it. They put me onto some books, and it was simple. This was 94, 95.

• 58:08Speaker 03

  It was Peter Lynch, one up on Wall Street. They said, "Find things you're passionate about, what products people are passionate about, and do that, and make it your own." And that's what we do. We figure out how the world works. So I started investing in 1996. I opened these accounts. And what I saw then was phones. Phones were starting. You know, it was the beginning of the cellphone era. And so the first three stocks I bought, which really kind of ingrained a passion for investing, were things that were evolved with cellphones. And so I go to college, and I'm basically wondering if I ever am going to have a job. These are 10-bagger-type stocks at the time, and it was the late 90s. And I really wanted to figure out how to turn this into a real career, not just the day-trading stuff.

• 58:50Speaker 03

  And I was lucky enough in college to have a great financial professor. My finance professor helped me get a job with Goldman Sachs. So I went and spent a couple of years in New York doing investment banking. And I was really fortunate. My first job came with a firm called Highside Capital. Highside, this is why BWCP's roots are in Dallas now. Highside was a Tiger Cubs spinoff. I guess it spun out of Mavericks. It was a Tiger Grand Cub. I spent eight years there doing technology, media and telecom. Phenomenal firm. We were a few billion dollars. I got an opportunity. I was, I was actually kind of thinking of forming this exact firm back in 2012.

• 59:30Speaker 03

  But Citadel came along and Citadel was expanding into Dallas under the name Surveyor Capital. And they hired me and I ran a portfolio for Citadel for six years. And it was a wonderful thing. And I tell you, what Citadel does so well is risk management and portfolio construction. What Tiger does so well is how do you find great ideas and uncover things that people aren't really thinking about? And use time as your advantage. And what I wanted to do, and this is what BWCP is, is kind of a bringing together of those two things. How do you use great fundamental stock picking? But understand that the world's different. When I started in 2004, 70% of the dollars traded were done so every day by fundamental managers.

• 1:00:15Speaker 03

  When you fast forward that to today, it's kind of the opposite, the inverse. And today, only 30 or 40% of the dollars are traded by fundamental managers. And so you need to be aware of what's going on. And so we use some risk tools, some of the factor awareness, some of the things that were from Citadel. But really just doing it around that framework of the Tiger Cub style of stock picking.

• 1:00:35Patrick O'Shaughnessy?

  I love the fusing of those two disciplines of the rigorous portfolio construction quantitative that Citadel obviously is famous for with the deeper fundamental insight on businesses, which is a great bridge into the key things that you found setting up a new investment firm. There's a lot you need to do. It's a lot of effort. I've seen it done. I've done it myself. I know that it's difficult. And I think companies and research is one aspect that gets talked about less. Like most of the focus on legal and systems and team and all this kind of stuff. How was Catalyst a part of that early setup? What did the company models allow for relative to your prior experience? That let BWCP get up and going faster and sort of how were they used in the early days?

• 1:01:17Speaker 03

  Let's paint a picture. You're, you're running a two plus billion dollar hedge fund inside of Citadel. You have a team of six. You don't even know what back office means because every single thing is taken care of. And you have a lot of resources. And you have a lot of young analysts that need to spend a lot of time doing the Citadel way or whatever that is. Roll that forward. I have a couple analysts, myself, small back office, friends and family capital, and my money, basically, our internal capital. And we've got to figure out how do we divvy up resources? Because right now, there's two things that are our most limited thing, time and money.

• 1:01:56Speaker 03

  And we have to find, across the board, where can we use our analysts' time to do the things that are most impactful to the research process? And where can we outsource with great partners to leverage both that time and money? And that's where Canlis came in. And we used some of our friends, we asked them about best, you know, practices, how they utilize their limited resources upfront, and Canlis came up time and time again. So what Canlis does for us, think about the amount of time that it takes to set up a model, to build a proper model, to update that model each quarter. We probably, with the number of stocks that we look at, we're saving one and a half to two analysts a year.

• 1:02:40Speaker 03

  It sounds like a huge number, but that's probably only looking at, you know, maybe 100 to 200 companies. But to do that right, that's probably what it would take from a manpower point of view. Now, if you think about the research process, what is it about the model that's so important? Models are the output of a lot of inputs. And the value add that we need to spend our time with our team on is, what are the inputs? What do the numbers mean? It's not necessarily the building blocks of the nuts and bolts of the model, but it's what it really comes down to is, what are your assumptions? How good are your assumptions? And what do your assumptions mean? So whether or not you're a super long-term investor, and you're going to make a decision once every 10 years,

• 1:03:24Speaker 03

  you still need to know, then what are the cash flows over those 10 years? If you're a super short-term investor, where you want to know everything about every margin, about every top line item, about every revenue trigger, you obviously need to make those assumptions somewhere, put them in models. So. It's a basic building block and it's the foundation of everything that we do, but it's not something that we want to spend all of our time doing the things that other people can do. So we can then turn around and use our time more valuably. And that has been a huge, huge help in how we allocate our resources.

• 1:03:56Patrick O'Shaughnessy?

  How do you think about the uniqueness of company models that Canilist provides sector to sector? So you mentioned TMT being a focus. I think GAAP accounting is great, but it also isn't perfect as a way to interpret or measure or investigate how a specific company works. Say a little bit about like company specific metrics, KPIs, you know, things that matter for one business that might not matter for another and how that factors into modeling.

• 1:04:21Speaker 03

  So that's the other important part. I've used services in the past that give you sort of a starting place. There's two problems with those usually. One, they don't have hardly anything above sales. And to us, usually in the TMT space in particular, the top line is one of the more important variables. Margins have less controversy, but are obviously important and then which brings us to free cash flow. So when you're using a generic service, sometimes you can get the basic building blocks of what's reported and maybe a queue with no context above for what's maybe reported on conference calls. Canilist adds that extra layer, which is important. But the most important part of it is the models live with us.

• 1:05:06Speaker 03

  So everything that we change or that we adapt to the things that we need, even if it's not already preloaded, we can add that and the model just simply updates. I used to have to send models away and when they come back, they're stripped. Anything that I've added, any sales that I've done, anything has been a problem. Now the models sit here with us. And the reason it's important is because those models lead to our investment templates, our investment spreadsheets. Those investment spreadsheets are the things in which we pitch our stocks off of in our internal meetings, which then flow into our rankings. Our rankings are what create discussions about what in the world's going on.

• 1:05:46Speaker 03

  So you put a number into a model with Camelist and it stays on my system and goes all the way through to price target changes. I do it somewhere else and it has to go send off. The links break, it comes back and all the work that we've done updating is gone. So how do we focus our attention on what is the most valuable stuff that we can do with the least amount of friction? And that's the important part. When you're using the Camelist system, there's no friction and it allows us to maintain our process. And almost like we just added a team member.

• 1:06:19Patrick O'Shaughnessy?

  Where do you see the most opportunity today in the coverage universe that you care about? And what's unique about how those businesses look in terms of their financial statements? And the levers that you think are important to underwrite and understand as an investor?

• 1:06:34Speaker 03

  I'll take that a slightly different way, but I'll get back to the original question. The thing that's most interesting to us, and this is kind of reshaped versus when we started. Again, we started really small and have been fortunate enough to grow to a few hundred million. But what has changed in the last 12 to 18 months is the number of companies that have come public. Whether that's IPOs, direct listings, all of these types of things are a big tax on the sell side. And they don't have enough time and energy to cover these things. And so if you think about, if your goal is to find something that is less well-known, undiscovered, something that has massive amounts of innovation.

• 1:07:12Speaker 03

  And your variant perspective could be simply knowing more about that company and actually being able to model it on a timely basis. That's where you have to have these types of partners. I would say the world from two years ago has changed dramatically in the sense that there are so many more companies. The sell side has not gotten bigger. If anything, it's getting smaller. The resources that they have to be able to help managers is not good, but. It has gone down somewhat. So what I think is going to happen, if you think about active management and you think about the number of ideas, we'll probably shift more of our capital into small and mid caps over time, where we have a balance between small, medium, and large. I didn't think that was something that was going to happen. You know,

• 1:07:54Speaker 03

  two, three years ago, people would say, you know, the number of companies are just disappearing. What is the role of active management? You can buy a few indexes. I think, given the valuation, given that the stock market just continues to grind higher, some of the best places to go out and look are some of these more undiscovered things that are newer, both on the long and the short side. There's a lot of stuff that came to the market that has no real meaning and no real value. So that, I would say, has been a real big change in the number of types of companies that we're going to have to use for financial modeling. And I will tell you, when you're looking at things that have limited history, that's

• 1:08:35Speaker 03

  where a place like Kales can come in. They'll be able to scrape that, go back, find historical stuff. And really, what we're trying to do to sift through all of these, we have to get some baseline level of knowledge. And you can't do that building out your own models. It would do, it would simply be prohibitive.

• 1:08:53Patrick O'Shaughnessy?

  In closing, thinking about other emerging managers, new managers that are hanging up their own shingle and setting up shop, what other key lessons did you learn about that process relative to your expectations? What was harder? What was easier? What did you have to do that you didn't even think about?

• 1:09:08Speaker 03

  There's a lot of lessons. Look, I will say a few things. One, starting a fund can be a very humbling experience. I had a manager and allocator tell me early on, if I could tell all new managers one thing, it would be that the world doesn't need another hedge fund. What makes you different and why do you need to exist? And that's a tough pill to swallow when you're, when you're out on the road marketing. I think people really wanted Citadel. So when we left, they loved Citadel and they wanted us to replicate Citadel. I will tell you, some people have been able to do it, but replicating Citadel is a very tall thing. So the advice I would give is find what you want to do. And what we wanted to do was slightly different. And if the thing that what you want to do is slightly

• 1:09:50Speaker 03

  different, in our case, it was a bringing together of the Tiger Cub and the High Side model with the Citadel model, be prepared to take a few steps on your own before somebody will underwrite you. And I don't think that It's a negative thing. I think we all get convinced that one or two people, you know, raising $5 billion day one means that the fundraising environment is very good. And I think we all have survivorship bias towards all of the people we look up to, whether it be Ken Griffin, or Steve Mandell, or all these people who have come before us. Those funds started really, really small, and they proved it. And they they worked themselves out for a few years. And I think that's probably the biggest lesson. The other one is getting

• 1:10:34Speaker 03

  off to a great start. You can never underestimate the power that that has. And that was something that, you know, we were very fortunate to be able to do. And then the last thing I would say, and I've had a lot of mentors over the past, and they've always said, well, if you want to do something, you're going to have to find something that you love to do. You can't do something that you don't love to do. And I used to think that that was just something that people said that were super successful, and that was a nice tagline. But investing is a hard business. And what those allocators mean, when they don't think everybody needs to be a hedge fund manager, they mean that. And so if you don't, investing can have its ups and downs. And if you don't love what you do on a day to day basis, and love building something, you know, from the ground up, you're really

• 1:11:18Speaker 03

  going to have trouble. And I think, don't underestimate the amount of time it's going to take you to build it. Don't underestimate the amount of money that it's going to cost to build, and find the most talented people you can possibly find. I think sometimes, new managers fall into a trap of not wanting to hire people they think are a lot smarter than they are. I tell everybody that we hire, I hope that I'm the dumbest person in the room. I want the smartest people around, because that's how we'll become more successful.

• 1:11:47Patrick O'Shaughnessy?

  Well, Brenna, I love your story. Love the fusing of styles, and also love the fact that the leverage point for Catalyst existed in the early days, that it let you do more faster. And I think you've made great points about sort of why repetitive processes that are nonetheless important and detail oriented matter a lot. But the one and a half to two analysts sort of stuck per a couple hundred companies per year, that's a real thing. So I think we've done a great job of overviewing your story, and also the ways in which it intersects with the Catalyst product. Really appreciate your time, and wish you the best of luck in your brand new firm.

• 1:12:21Speaker 03

  Thank you, Patrick. Appreciate it.

• 1:12:25Patrick O'Shaughnessy?

  If you enjoyed this episode, check out joincolossus.com. There you'll find every episode of this podcast, complete with transcripts, show notes, and resources to keep learning. You can also sign up for our newsletter, Colossus Weekly, where we condense episodes to the big ideas, quotations, and more, as well as share the best content we find on the internet every week.