Lan Xuezhao has spent the last few months pulling together $136 million for her new machine intelligence-focused venture capital fund, Basis Set Ventures. I met Xuezhao for tea on a park bench in Potrero Hill earlier this week to chat about her strategy for the fund.
I spend a good portion of my time meeting with investors, but if you don’t know a lot about the scene, Potrero Hill is not a place you go to meet VCs. Hot spots for meetings generally range from opulent coffee shops in San Francisco to opulent offices on Sand Hill Road. So a park bench in a fairly residential, low-profile neighborhood stands out.
But even more than that, Xuezhao has a surprisingly laid-back demeanor and an apparent academic appreciation for technology. With a PhD in quantitative psychology, the former head of mergers and acquisitions for Dropbox can do something that most other investors cannot — relate to the incredibly talented founders of highly technical startups.
Breaking rank with increasingly flashy, services-focused AI studios like Yoshua Bengio’s Element AI, Xuezhao wants Basis Set to be the anti-VC. Everything blindly promised by AI-focused VCs gets a layer of realism. Data sets: What data, why and does it actually exist anywhere? Technical mentors: How about I just sit down and we both start by being honest with each other — then if we can’t come up with it, let’s text someone who can.
We spent about an hour talking about the state of AI startups and how Basis Set Ventures aims to capture the windfall from the burgeoning space. I’ve edited all remarks for brevity.
TechCrunch: Why did you feel $136 million was the right number to start with?
Lan Xuezhao: The number is more strategic than anything else. I feel like there’s a gap between Series A and smaller seed deals. There are a lot of smaller seed funds and it’s hard to compete with them because there are so many.
At Series A there are a lot of bigger names who do a very good job with those. But in-between, there’s a sweet spot for checks ranging in size between one and three million dollars. And not that many funds are able to do that.
TC: Can an AI focus still be a differentiator in a market that now seems saturated with AI-focused funds? What do you think is the real value a VC can add to a machine intelligence startup?
LX: Given my experience, I think go to market is the most important because algorithms are less defensible. Being able to help startups close larger clients is something I spend a lot of time on. Startups value me as a thought partner. You don’t have to be very formal with me in terms of presentation or reporting numbers.
I’ll sit down with a founder and we will go through an Excel spreadsheet and figure things out. I’ll help startups recruit people. Those are the resources that people want. I’m very pragmatic; I want to help founders get this stuff done.
The fund is very focused in terms of thesis and size. We do a lot of inbound leads, but we also do a lot of research to ensure our leads are not biased. Every Friday we talk to customers, the real people who actually use these products, and we try to figure out what works best and what doesn’t work at all. A lot of times the products people are using are from companies not based in California. These end up being very helpful conversations.
TC: Is the AI studio model overhyped?
LX: There is value in technical talent. I have advisors and their perspective is very valuable to me. Even product managers and designers, their perspectives are really valuable to a fund. But you want to make sure that these people are involved enough to actually help uncover blind spots.
Some incubators try to provide data which will help companies build early products. I think that’s a little tricky because the data needs to be very targeted. There is a lot of potential for value, but it depends on exactly what a company needs.
TC: Are machine learning APIs and developer tools defensible as investments in the long run?
LX: I’ve seen this approach working for some companies, but I’m a bit torn. I don’t have a strong opinion. It’s truly a case by case basis. I have invested in one company that fits this profile and things are going great for them, but I’ve also heard cases where it’s not working out so well.
I like when companies develop their own technology. The integrations need to be good and the experience needs to be native in order for this to be useful. Developers need to have very strong incentives to make this work. It’s not that easy to get all three, but if you can, companies are in a pretty good position.
TC: Do you agree with the majority of folks opting to invest in verticalized AI over horizontal platforms?
LX: I believe in vertically integrated full-stack solutions. Algorithms are getting more and more commoditized and big companies are trying to do a lot of the horizontal plays. It’s hard to do well there.
TC: Are you OK with startups using off-the-shelf AI tech early on?
LX: You need to be building something that actually solves a problem versus working on tech for three years and building something that people will not use. AI is a path to solving a problem versus the solution. AI is not the goal, it’s something that solves a problem. Having a real product that people will actually use sometimes means using off-the-shelf tech. Then, in the future, when the product actually takes off, you can make the tech more robust.
TC: You’ve been investing in building a quantitative sourcing engine; what’s the real value that it brings? Is this a natural application of AI inside Basis Set?
LX: Quantitative sourcing is a great way to cover blind spots. Each person’s network is limited and biased. It’s a great tool to supplement people’s own network so that you have a shot at seeing something you otherwise might not see. When doing CorpDev for Dropbox, I first hired a PhD from MIT who did a lot of work building us a quantitative sourcing engine. Together we found a lot of interesting companies that we definitely wouldn’t have seen if we didn’t use that engine. This approach won’t replace traditional sourcing, but it’s a really strong tool and I plan to build one for Basis Set Ventures.
A lot of the challenge is finding the right signal. The algorithms themselves don’t actually need to be that complicated. There will be some curve smoothing when we look at growth etc., but most of it is understanding the problem and finding the right signal so that you can get the right trigger set up when something happens. It requires domain expertise in the same way as AI, though.