Amplify Partners: AI Investing and the Startup Landscape with General Partner Sarah Catanzaro

Yeah, but so, I mean, for some background, I focus on investing in data and ML tools and platforms, as you said. And I spent maybe a year, year and a half doing generalist investing. I think I realized over that year that I loved investing. I loved seeing different things.

I loved seeing founders come into themselves, to see companies grow from products to platforms. But I just didn't have the same passion for all domains. My heart was really in data and ML tools and platforms, which makes a lot of sense because that's where I've spent most of my career. So certainly, I think given my upbringing, I was always very interested in the sciences. I'm going to embarrass my mother again. She's a psychiatrist. My father's a molecular biologist. But I tended to gravitate more towards things related to data.

So I spent kind of the first couple of years of my career studying insurgencies using techniques primarily from symbolic systems to better understand Somali pirate groups, radical Islamic groups, and things of that nature. Not shockingly, after a couple of years, I found my way to Palantir. And maybe circa 2014, 2015, I wanted to move out to the Bay Area. My then-boyfriend kind of had given me an ultimatum. And so I called a friend and basically said, like, I know how to do things with data.

Where should I go work? And he connected me to a woman, Danielle Morrill, who was then the CEO of Mattermark. Mattermark was collecting data on other startups and selling it to investors. I'm not sure if FirstMark was a Mattermark user back in the day. Probably. And really, it was at Mattermark that I think I learned both about the startup ecosystem, but also about machine learning. Up until then, I had done various things in kind of statistical learning, but deep learning was kind of new to me.

We had a small team, though, and we didn't have an army of analysts who could go out and collect information or manually classify things. And I kind of knew enough math and statistics to figure it out. So the rest is kind of history. Ultimately, though, one of our customers ended up recruiting me, and so that's how I made it to the dark side.

Yeah. I think one thing that we probably share is a belief in focus. I have tremendous respect for generalist VCs. Frankly, I don't know how they do what they do. I don't know how you can excel at picking companies in travel tech, GPU-accelerated databases, and LLM infrastructure, and so on and so forth. So for some background on Amplify, as I kind of alluded to before, we focus on investing in technical tools and platforms. Generally speaking, I would categorize those as data and ML tools and platforms, developer tools, and enterprise infrastructure.

We also look at vertical applications of some of those underlying technologies, whether it's databases or distributed systems or machine intelligence. I mean, just commenting on focus, I think one thing that we see, and I'm sure you see the same thing in your portfolio, is that the same conversations happen often. I think the challenges that technical tool builders face actually tend to be fairly homogeneous, whether it's thinking about open-source strategy, thinking about top-down versus bottoms-up execution, or what that even means, thinking about multi-stakeholder sales.

But these challenges are also quite different from those that companies in other subsets of the market face. And so, instead of trying to be good at not just picking companies in all categories, but also helping companies in all categories, we wanted to just really focus on getting great at helping technical tool builders realize their visions. Very good.

But it's also going to take over the world.

Yeah, absolutely. I think it's important to think about both what we learned, but also what might be different this time around. I think what we learned back in 2017, 2018, when there was a lot of hype around reinforcement learning, GANs, things like AlphaGo, things like Alexa, was that applying AI is not a substitute for building great products. The same principles of delivering value to users, of managing their attention, all of those principles hold. AI is just a technique.

It's just an underlying technology that you can use to build great products, but you still need to build a compelling product that solves an urgent problem for somebody and has the opportunity to expand into perhaps other parts of the workflow, other markets, et cetera. I hope that's a lesson that we learned. I think time will tell. One thing that makes me nervous about what I see today is that I see a lot of startups that I think recognize that AI is not a silver bullet.

But instead of focusing on delivering concrete value to people, and almost delivering mundane value to them, they seem to be more focused on competing for mindshare. It almost feels like a lot of AI startups today kind of think about themselves as being embroiled in this battle for attention, where they're just releasing a bunch of things that they think will capture people's imagination, but have yet to really figure out a formula for, again, delivering quotidian value to their users. So I'd say that that's kind of the primary lesson that I think about.

Yeah, I mean, I think excitement is the right word. Just like the same principles around product development, product management hold, so too do I expect the same principles around venture capital to hold. There will be a lot of startups that come to market. There will be a lot of startups that fail. I think right now we are in kind of peak hype, and I think for a lot of these AI products that exist today, the first experience with them, it's magical.

It's so radically different. There's something kind of very visceral and emotional about ChatGPT writing you a sonnet in the style of Quentin Tarantino. But the magic starts to fade again as you think about, how am I going to use this as a tool every day, every week? How am I going to extract value from these systems? I was admittedly, I think, a little bit more cynical about opportunities related to LLMs even a couple of months ago.

But the technology, it is radically better than 2017. These models are incredibly powerful. And so I think people will figure it out. I think a subset of those who are founding AI companies today will figure it out. But I think we're going to learn a lot in the next couple of months, probably the next couple of years. And I think there will be startups that don't make it. Kevin and I were actually starting to talk about this.

We were saying there's so many people who are founding companies right now, and you get the sense that there's kind of FOMO that exists in the market, that people feel like now is the moment for me to start an AI company, and if I don't start it now, I'm going to miss out on the opportunity. I actually disagree with that. I think we are learning so much literally day by day. I mean, GPT-4 was released yesterday. And there is possibly a second-mover advantage.

There will be lessons that we learn again from seeing this first wave of startups emerge, build, grow, and fail.

Yeah, it's a great question and perhaps a perfect illustration of what I was saying earlier about new information unveiling new patterns about the market and the delivery of AI. I think even a month ago, there seemed to be consensus that startups had an advantage over incumbents when it came to leveraging LLMs and integrating them into their own products. People expected that existing tech companies, including some of the FAANG and Fortune 500 companies, would be slow to release new systems.

The quintessential example is Google versus ChatGPT. I think the past month has demonstrated that that is not true, that a lot of existing companies can actually act pretty swiftly to integrate LLMs into their applications. I actually think this is one of the most powerful things, one of the things that has changed the most in the ML landscape. It is so much easier to prototype LLM-driven applications. You no longer need to hire an ML team before you can even run an A/B test or experiment to determine if this is something that users like.

That said, I think there will be an opportunity to redefine workflows around LLMs. People often draw comparisons to mobile. I think some of those comparisons are good. Others are a little bit misguided. But mobile did change the way that we conduct certain workflows and behaviors. It unlocked opportunities to do things differently. And I think that is the promise of LLMs and startups. It's not to build better versions of things that exist today, but rather to reimagine some of the tools and systems that we use in an LLM-centric way.

Yeah. Interesting is an ambiguous word. I would say I love tools and infra. So, if you're asking me what is more intellectually interesting, I just love tools and infra. And I will find opportunity there because it's what I love. It's what I'm most passionate about.

I don't think we'll see that value is accruing exclusively at the infra layer or the app layer. We haven't seen that in any other paradigm shift. Again, if you think about mobile, if you think about social networks, et cetera, there was value at the app layer, there was value at the infra layer, and I don't expect this to be any different. It feels like right now some of the value at the app layer can be captured immediately, whereas the value at the infra layer will expose itself over time.

Because so many of these companies who are trying to integrate LLMs into their existing products started in the past two to three months. So they haven't hit that brick wall yet. My colleague Sunil calls it the stubbed-your-toe problem. Before tools and infra can really get breakout adoption, people need to try to do things on their own. They need to try to do things without adequate and sufficient tools, stub their toe, and then they'll come back looking for better tools and solutions.

Yeah. So I will admit that personally, I started to get a little bit bored with the modern data stack and the data ecosystem, and as a result, shifted more of my attention to the LLM ecosystem. But I'm beginning to revisit some of the assumptions that I had about the so-called modern data stack that helps companies transform data into insights. I will say that I think that the tools that data analysts, data engineers, analytics engineers—the tools that they have to produce reports, dashboards, insights—they're very good.

They can do their jobs effectively today. But what we've seen over the course of history, what we see time and time again, is that very good things can be replaced with things that are even better. And I expect the same to happen with regard to the modern data stack. I think one of the key things that I see today is that a lot of the tooling around the modern data stack seems to really be based on the notion that storage is really cheap.

And so five years ago, we saw this transition where companies previously might try to transform and prepare their data before it landed in their analytical database to one where they just shoved everything into their data warehouse and then did the work of preparing, curating data there. I think people are starting to recognize that storage is cheap, but compute is not. And if you are preparing your data in a data warehouse, you're using compute, and your bill is going to get pretty high.

And so it's funny because I feel like we had this big transition from ETL to ELT, and now it feels like the market is beginning to rediscover ETL, is beginning to rethink where data transformation happens. I do think that that transition has been enabled by a lot of fundamentally new technologies, things like DuckDB, which enable you to transform data at the edge, thereby saving significant costs that might otherwise be associated with data movement, data storage, and compute.

Yeah, absolutely. So the MotherDuck team hailed from Google, and I think one of the phenomena that they saw there is that basically big data is a lie. Most companies do not, in fact, have big data. Perhaps there are a few companies in genomics or astrophysics that actually have petabytes of data, but most datasets are not that big. And similarly, most datasets need not be that big. Again, if you think about this pattern of ELT and this trend towards storing everything in the data warehouse, if you store only what you need, if you use only what you need, then your dataset also starts to look a lot smaller.

But so many of the data tools and systems that people use today, they're optimized for big data. They're optimized for running a point query across petabytes of data in milliseconds, which is not the work that most data analysts and analytics engineers are doing on a regular basis. So MotherDuck is kind of rethinking this paradigm and thinking about, if we were to build world-class data tools with the principle that most data is small, and around this open-source technology called DuckDB, in a way that is fundamentally cloud-native and serverless, what could we build?

What would we do differently? I think the other promise around DuckDB and MotherDuck, which I just alluded to, is that we can now rethink compute and leverage compute where it exists today, on your laptop, in the browser, thereby delivering much more performance and cost efficiency.

Yeah, that's such a tough question. I've been investing for seven, eight years now, and it's something that I think about a lot. But frankly, I don't see that many patterns related to successful founders or those who don't succeed. We've had founders who have run small businesses and recognized a problem in their environment, realized that that could be potentially scaled to other businesses, who've been successful. We've had founders who come from Google. We've had founders who are commercializing the research that they did at MIT.

I think what matters is that the founders deeply understand and care about the problems that they're solving. No startup is ever easy. Even the ones that grow the fastest are often not just incredibly hard, but incredibly emotionally taxing, incredibly physically exhausting. And if you don't really care about the thing that you're building, it's hard to survive that roller coaster. But it feels kind of trite to be like, well, you need to know about and care about your problem. But that, to date, has really been the only thing I've been able to discern.

Yeah. I do think that some things have changed in the past decade that have made it easier for technical founders to build massive companies. Certainly, it felt like maybe 20 years ago, companies would have to do a lot more market research and user research before even launching a product. And so the sales and marketing planning prior to launch could take years. And I think for that reason, you tended to see more business-focused founders. Now, with more agile development, it's easier to respond to the market and just collect information that you then act on, which I think has created a simpler path for technical founders.

But that doesn't mean you don't need to build sales. It doesn't mean you don't need to build marketing. And I think one of the most important things that we can do is have empathy for technical founders. Just earlier this week, I was speaking with someone who started a company after leaving a position at an industrial research lab. And we were talking about how he might prepare for a Series A or Series B round in thinking about his capital needs. And he said to me, "Okay, I think we're doing enterprise sales, so I'm probably going to need to hire an enterprise salesperson."

I was like, yeah, you might want to also consider hiring a sales engineer. You might also need to consider hiring a sales development representative. And I got a blank stare. And there was a moment where I was like, oh my gosh, he's preparing for his next round and he doesn't know about the basic unit of enterprise sales. But that's our job. Our job is to ensure that founders know what to expect. Technical founders know what to expect as they pave their way into these new territories and domains where they don't have experience.

I think the other part of the job is just to make sure that they have access and connections to people who can help them think about maybe how to build an enterprise sales team or how to approach developer marketing. I find myself constantly reminding technical founders, including those within the Amplify portfolio, that they don't need to learn things from first principles or even from experimenting. They can just ask people who know how to do it. And we're there to facilitate some of those connections.

Yeah. I hope this doesn't come across as whiny, but I think one of the things that is really hard in venture is that as you become more senior, it's actually more work, not less. I always thought, okay, once I'm a GP, I'm just going to sit in Hawaii sipping Mai Tais and other people are going to source for me.

Exactly. Or sit in Miami and sip whatever they sip there. But that is not the case. Structurally, as you mature as an investor, you do more deals, which means you have more portfolio companies, which means that you have more obligations. And I have definitely found that as my portfolio has grown, finding time to learn can be very hard. To the point about how we help technical founders learn, I think it's important that we're constantly learning from the technical founders in our portfolio.

We're constantly analyzing our interactions with them, how they build, what they build, what works, et cetera, and think scientifically about our own craft. But I've also made a significant effort to find that time to do research, to read papers, to write Projects to Know, which is a newsletter where I typically highlight a few interesting research papers, open-source projects, company blog posts outlining internal work.

I should know this. Maybe ask ChatGPT. Sorry. I'm sorry to Malia, our marketing director, but if you Google Amplify Projects to Know, it'll probably pop up somewhere.

I mean, the truth is that I think at this point in my career, I just need to say no to pitches more. I try not to see more than two or three pitches a week. Now, I think the danger there is that it tends to entrench your biases about where there are opportunities because people don't have the opportunity to change your mind. And so, when I do see pitches, I do try to think about this question of, is there anything that the founder could say to me that would make me love this, that would make me think about this differently?

But there's only 24 hours in a day. Like, you need to sleep, you've got family, et cetera. I've certainly found that in order to make time for research, in order to make sure that I'm constantly talking to practitioners too and hearing about what they're seeing on the ground, the thing that's had to give is listening to pitches.

Thank you for having me. Thanks for listening to The MAD Podcast. If you liked this episode, be sure to leave us a review. com/events/data-driven.