Updated: Apr 8
"We want to back companies that are not just building great tech, but building great tech that solves a real business problem in the world. " - Rob Toews
This episode features Rob Toews, a venture capitalist at Highland Capital Partners. Rob covers many topics such as AI Industry trends, focus areas for VC’s when considering AI Capital investment, and the impact of deep learning applications such as radiology on AI Commercialization.
Rob Toews is a venture capitalist at Highland Capital Partners focusing on investment opportunities in machine learning. He writes a regular column for Forbes about the business implications of AI. Previously, Rob worked on strategy at the autonomous vehicle startup Zoox and management consulting at Bain & Company. Rob received a joint JD/MBA degree from Harvard University.
Robin.ly is a content platform dedicated to helping engineers and researchers develop leadership, entrepreneurship, and AI insights to scale their impacts in the new tech era.
Subscribe to our newsletter to stay updated for more inspiring talks:
1. Why Interest in AI?
Margaret Laffan: I'm here with Rob Toews, a venture capitalist at Highland Capital Partners. Rob, good morning. Thank you for joining us.
Thanks for having me.
Margaret Laffan: We're going to dive a bit into your background first. Your academic background includes a BA from Stanford, and you've also spent time doing your law degree and MBA at Harvard. You are an investor at Highland Capital Partners joining in 2019, and you spent time with Zoox doing strategy. You're a contributor to Forbes, and you write about the big picture in AI. Can you share what triggered your interest in AI and a bit more around your background in this space?
I really got interested and drawn into the AI and machine learning field through the realm of autonomous vehicles. Prior to joining Highland Capital full-time last year, I spent the past several years pretty deeply immersed in the world of autonomous vehicles.
I started on the policy side, and I worked in the White House under the previous administration in the very early days of Obama's administration, figuring out what our regulatory framework could and should look like for autonomous vehicles(AV). This was in 2015. Frankly, the government is still grappling with that question. But it was really interesting to be there in the early days and think the big picture around how much autonomous vehicles are going to change so much about the economy and society.
Since then, I've worked on AV from a number of different angles. I worked on a part-time basis with Highland for a couple of years, helping lead their autonomous vehicle investment efforts - looking at autonomous vehicle companies across the spectrum from the core AV stack, core autonomous vehicle technology, computer vision, simulation, etc. I spent a couple of years helping lead the strategy team at Zoox, which is an AV startup here in the Bay Area. And then most recently I joined Highland.
When I joined Highland, my decision as I was thinking through next career steps was, as passionate as I am about autonomous vehicles, I came to appreciate more and more that machine learning was a topic that was going to disrupt so many different industries, and venture capital gives you that broader aperture to look across the economy and see where different opportunities for disruption are going to emerge.
2. Highland Capital Partners' Investment Focus
Margaret Laffan: How did you come into the VC world back then?
I’d like to think of it as depth versus breadth. When I was at Zoox, I was very, very deep in the world of autonomous vehicles and all of the business implications, technology implications, and regulatory concerns. It’s a fascinating place to be, and I was drawn. The thing about VC that drew me more was the opportunity to have that broader perspective and to look not just in transportation, but in agriculture, manufacturing, law and construction, and a bunch of different industries, and think holistically around how is this technology - which really is a transformative general-purpose technology - going to change across so many different categories.
Margaret Laffan: Highland Capital Partners was founded in 1987. It’s a 4 billion venture capital firm. You have offices in New York, Boston, Bay Area, in Palo Alto, and you recently opened another in San Francisco. Your track record includes 46 IPOs and over 125 acquisitions. Notable investments include Auris, Carbon Black, nuTonomy, Gigamon and Rent the Runway. And more recently, you invested also in Vecna Robotics. Can you tell us a bit more about Highland’s overall investing strategy?
What we focus on at Highland in terms of identifying new investment opportunities are companies that have early signs of strong product-market fit. Typically, that means a company that has a product in the market, and then there's a very strong pull from customers, there's a clear market need, and customers are reacting very positively. Generally, that means at least some amount of real revenue - not just pilot revenue, but real recurring, committed revenue - that can be at the series A, at the series B, sometimes even the series C. I think the names of rounds can sometimes be less helpful than the actual stage the company is at. But that's where we'd like to get involved, the things that we look for.
Highland is a generalist firm. So we invest across all categories on the consumer side, and on the enterprise side. I spend most of my time in the world of machine learning, but as a team, we look across sectors. And what we like to look for across sectors is companies that have first and foremost, amazing teams, and specifically teams for whom there's great founder-market fit - people who are not just really impressive in the abstract, but whose background positions them to be the best people in the world to build this particular company. We like businesses that have recurring revenues; we like businesses that are capital-efficient; we like businesses that are disrupting industries in fundamental ways, as opposed to building an incremental solution on top of what already exists.
But we're open to looking across different sectors. And Highland’s philosophy is to be really high conviction, low volume investors. In that sense, we will do maybe six or eight or ten investments a year, but for every single investment we make, we're really excited about it. We typically take a board seat and we plan to stay with these companies through their entire lifecycle.
Margaret Laffan: Still focusing on HCP, you've opened your new location here in San Francisco. What has been the decision to have a geographic presence in SF?
Highland was founded in Boston in 1987 and remains our headquarters. We've had a presence in Silicon Valley since the early 2000s, so close to 20 years now. We've been based down in the South Bay in Menlo Park and then in Palo Alto. Like a lot of VC firms, just last year, we as a firm realized that there's so much startup activity happening in San Francisco, it would be good to have a presence up here. More and more young founders want to live in the city. And if we want to be as close to the startups and as close to the action as possible, it just makes a lot of sense to have a presence here. So we opened this office, we also opened a small office in New York last year as well.
3. Future of Autonomous Vehicles
Margaret Laffan: Your investment area is machine learning and autonomous driving. You’ve had work experience with Bain and Zoox, which gives you unique insights into the autonomous vehicle space. We've seen a lot of consolidation here in the last few years and much disruption in terms of the business models, ability to scale and regulatory impact. From your study, where do you see the AV space currently?
At the highest level, I remain as excited about autonomous vehicles and the potential to transform society as I ever was. The big thing that has been an important learning for me, and frankly for the entire industry, is the timeframe that it's going to take for it to fully come to fruition. I think long-term autonomous vehicles will be a huge deal. They’re being underhyped.
But I think we're in this interesting phase where, at least, if you're talking about the core AI technology for robo-taxis, like autonomous vehicles on the street, there are a handful of companies who are extremely well funded and who are putting years into development efforts like Waymo, Cruise, I would mention Zoox, of course, and Aurora. These companies are honestly not really startups anymore in most cases, they have billions of dollars and years of work behind them. I think it would be formidably challenging for any new real startup, like a series A startup or two guys with a laptop to start now and try to develop the technology for robo-taxi.
But even though, from a startup point of view, the landscape is fairly mature. We're still not commercial. None of those companies that I mentioned have actually gone to market with a fully autonomous vehicle, except in very controlled limited environments, like a few miles’ areas in Phoenix, for instance, in the case of Waymo. I think we're in this weird intermediate phase where there's not very much VC activity on the core robo-taxi side, but you still don't see this product in the market. And as a result, it's leading to this trough of disillusionment. I think there are less excitement and momentum around autonomous vehicles today.
Just to reiterate, I think in the long term, the impact will be absolutely tremendous across many different sectors, the way cities are designed, the way people get around. But today, I think there's probably less interesting investment opportunities for traditional VC, series A, series B investments.
Margaret Laffan: From an investor's perspective, where do you see the vision of AV going in the next two to five years?
One area that's fruitful for VCs to look at in the near term is the concept of applying autonomous vehicles in fields other than automotive. There are a number of industries who involve the use of vehicles driving around in more constrained structured environments, less complicated technologically - agriculture, construction, mining. These are some of the biggest industries in the world. But they're historically under-digitized. So there's a huge opportunity to automate some of those more simple driving tasks. Labor is one of the biggest parts of the cost structure of these industries, so there's a huge opportunity for value creation, and you are seeing some startups emerge that are addressing those areas. I think the go-to-market will be tough, precisely because those industries are historically under-digitized. And they're not used to buying newfangled technology. There are very valid concerns around job loss as it relates to automation. So I think there will be plenty of challenges.
But in terms of where autonomous vehicles will be deployed in the next two to five years and have a big impact, I think it’s probably going to happen in those industries we just mentioned first, even something like long-distance trucking is another example of an application that's less technologically complex. You don't have pedestrians on highways, you're basically driving a straight line. I think those are the ones you will probably see coming to the market before you see a truly level 5, urban robo-taxi.
4. Application of Deep Learning: Radiology
Margaret Laffan: Let’s jump a small bit deeper into one of these areas, which is deep learning. In your Forbes article, “Deep learning has limits, but its commercial impact has just begun”. You specifically mentioned radiology, and Geoffrey Hinton has stated that radiology is an ideal use case for deep learning. However, AI adoption in healthcare, for instance, is relatively slow compared to other industries, such as transportation, FinTech, or what we see across the enterprise today. Why do you think this is the case?
It definitely is slow in healthcare. Honestly, I think it's slow in those other industries as well. It's hard to bring an AI-based product to the market. I think certainly in healthcare, also in transportation - the difficulty of commercializing autonomous vehicles, and in financial services as well.
I think a key point here is that there's such a huge gap between building the algorithm that will work in a research study that can, for instance, helps a radiologist in identifying breast cancer, and operationalizing a company around that. Things like health care and transportation are particularly hard because they're heavily regulated industries. They involve a lot of safety-critical situations, lives on the line. So taking that algorithm you develop that can perform well and turning it into a company involves developing a business model, developing a go-to-market strategy, navigating the regulatory concerns, and a lot of cases selling into industries that are really slow-moving, like healthcare systems, navigating those sales cycles. The whole thing takes years.
So I think radiology is a great case study of technology from a purely technical point of view. Deep learning algorithms probably are better than radiologists at spotting breast cancer overall. And yet, we still haven't seen an AI startup built that's deploying AI at scale for diagnosis purposes. So there's just so much difficulty in the commercialization process and the operationalization process, which I think represents a huge opportunity for entrepreneurs that are really willing to dig in and put the effort in.
5. Different Point of View against the AI Community
Margaret Laffan: What is one commonly held view in the AI community today that you disagree with?
Good question. I think there are too much focus and fixation on an accumulation of massive data sets and specifically, massive labeled data sets. I think there's this high-level motif like the more data you have, the better your models are, and whoever has the most data is going to win, and in general, more data does help you train better models.
But there are so many signs that this quest for bigger and bigger data sets, bigger and bigger models is unsustainable from an environmental point of view, from a cost point of view in terms of compute - there are very few companies in the world that have the resources to actually train these world-class models - and also from an efficiency point of view.
There are techniques that you see emerging that are in some cases still in the research community, or in very early days of commercialization that will let companies build AI models using way less data, at least way less labeled real-world data. There are things like synthetic data where you can generate your own data set to your own needs, things like few-shot learning, small data techniques. These are essential if we want to be able to build AI that's more flexible and also less resource-intensive. So I think there's too much of a focus now on labeled data, big data, supervised learning, and you'll see that shift to more efficient models as we go forward.
Margaret Laffan: Rolling off that and we're looking at now, it's early 2020. What are you most excited about regarding AI startups this year?
This is a Highland-specific answer, but they're a couple of areas that are particularly interesting for us. One is computer vision. Highland’s focus is on companies that have a product in the market that is starting to build a real business, and computer vision has a lot of appeal for us in that regard because it's a little bit more mature of an underlying technology compared to something like language. There are a lot of companies that are building real products that are creating real value in the market already, leveraging computer vision.
In particular, we really like vertically-focused computer vision companies, with computer vision solutions applied to property insurance or computer vision solutions applied to agriculture or construction. There's a lot of ways to automate things that would otherwise require a lot of humans and manual labor and make it cheaper, make it faster and make it more accurate. So computer vision is one area we're spending a lot of time in.
Another one is machine learning developer tools. The thesis here briefly is that a lot of massive multi-billion dollar businesses have been built over the decades, providing developer tools for traditional software engineerings, like Atlassian or GitHub. Developing a machine learning model is a fundamentally different task, a different workflow than writing traditional software code. It’s much less focused on the writing of code and crafting hand rules, it's much more about handling the data, and then set the parameters properly to train the model. So it's really a totally different skillset and demands a very different type of tool. Right now there are not really mature offerings for developing machine learning models. You see a lot of data scientists and ML engineers doing this by hand or ad-hoc, not in a really streamlined way.
I think huge businesses will be built, providing tools that streamline that process of collaborating, and error checking, version control, quality control, and managing data and so forth. That's another area where there have been tremendous startup activities in the past 12 to 18 months, and that's an area that we're monitoring really closely.
6. Changing Focus of VC
Margaret Laffan: For our listeners here who may have founded an AI company and are in the early stages, or are thinking about starting an AI company, what are the top three things that you would look for when considering an AI startup for investment?
I think talent is a big one. It's critical to have people that really have credibility, that will help both with building a product, and also with recruiting other top folks.
Margaret Laffan: What do you like to see here when you look at teams for talent?
Rob Toews: It's less on extensive professional experience. I think having an academic background in machine learning is helpful, sometimes it’s the more recent of a grad you are, the more up to date your background and skill set is. So I think talent is a big one.
Data strategy is another big one, how you think about developing your data assets in a way that will let you build a real product.
The third one is basically a business model and a go-to-market strategy that is robust. Basically, a company that wants to build a product that's addressing a real concrete, well-defined use case, being vertically focused, for instance, is one way to do that - if you say I'm building a product for the construction industry that does X, Y, and Z, as opposed to, I developed this cool algorithm, let me go out and see how am I able to find a business use for it. Because at the end of the day, we want to back companies that are not just building great tech, but building great tech that solves a real business problem in the world. So I think having that go-to-market focus is really helpful.
Margaret Laffan: Especially when you think about it from a buyer's perspective. Buying AI tech is still something that's been explored in a lot of companies in terms of the amount of investment that they're going to make. A lot of them are experimenting initially before they make those bigger purchases. We have to ensure there's a balance on the startup side that they can go the distance and understand that the selling cycle is going to perhaps be longer in these cases too.
Margaret Laffan: We want to look a bit more at the big picture now. When we think about what's happening at a global level, and we think about things this year, for example, in the US we are in an election cycle, we have COVID-19 unfortunately spreading outside of China. You've got facial recognition legislation, other legislation coming up in Washington around AI technology. And then there's a lot of global discourse on ethics in AI and so forth. What do you see as being the biggest macro challenges and opportunities for AI globally?
In terms of global macro trends, I think there are two that I would call out as being particularly worth tracking. The first is the push for AI regulations. I think there's no question that governments around the world are going to start putting in place regulatory frameworks to address AI. The European Union just recently proposed something this week or last week, more and more members of the US Congress are starting to talk about it. You are going to see regulations put in place, similar to GDPR(EU General Data Protection Regulation) in the data privacy context.
It's a good thing. I think it's an important step for governments and regulators to take. It will just be really important to track: Are they crafted in a way that doesn't stifle innovation too much? Is it too restrictive in a way that it won't be able to evolve over time? It's just really important to be aware of as an AI entrepreneur or investor, because it means that this era of technology development will be different from, for instance, the.com era or the SaaS era more recently, where companies were building software products, and we're really thinking about, how might this run afoul of regulatory concerns? How might this impact society more broadly? That's going to be for sure another factor that is going to be really important to consider.
The second macro trend that I would point out is something that's been widely reported on and as a favorite talking point, the dynamic in the tension between the US and China in the field of AI. It very much does feel like a new space race, maybe minus the intensity of the Cold War geopolitical situation. But both countries see AI as a key geopolitical asset, especially China. China is investing tremendously to develop its prowess and supremacy, and its government has really launched an entire country effort to achieve its stated goal, as to be the world leader in AI by 2030. And the US government is starting to think about it more, starting to talk about it more. There's a tension between them more probably in AI, and I think it's going to become one major dimension of that, and it will have concrete implications for people both in the US and China working on the topic in terms of cross border investment, in terms of cross border recruiting and partnerships. Hopefully, it's a dynamic that can be competitive, but also constructive and positive and peaceful. But it's definitely dynamic that's worth being aware of.
7. Key Challenges for AI Companies
Margaret Laffan: Let’s look more at a micro level now and look at AI companies themselves. We talked to a lot of AI startups regarding the dynamics of their go-to-market strategies, their pricing model, how they attract top talents and so forth. What do you see as being the key challenges for AI companies this year?
I think another hard piece that every AI company is still figuring out, is around scalability - once you've built a model that works in one context, being able to build technology and a business model that you can scale.
There are a few challenges specific to ML that companies are grappling with. One is being able to build a model that's generalizable. For the radiology example, you have a model that does well at identifying breast cancer in this hospital, in this population of patients. But if you looked at scale into a big business, you can also use those algorithms in different communities and the technique still works. A lot of times that requires collecting new data sets from different contexts, which can take a long time, you have to label the data and everything. So it’s building models that can scale and generalize.
And also, AI services still don't work off the shelf in a seamless way the way as typical SaaS products do. Oftentimes, there's still a pretty heavy service component with humans in the loop. That naturally limits how quickly you can scale, for instance, for machine translation, product deployment that you have, you have to have actual humans in the background, the customer success team. Those are a drag on unit economics and also a drag on how quickly you can scale. So I think naturally as the tech gets better, those issues will be more resolvable, but it's something that you really want to see that companies grab on.
Margaret Laffan: And then of course, it adds into the cost of the product versus what a buyer is going to pay, the cost of maintaining your product and then deploying it and so forth as well.
Margaret Laffan: Rob, it's been a pleasure. Thank you very much for joining us today.
Thanks for having me.