78. Artificial Intelligence Investing, Part 2 (Nathan Benaich)

Download_v2Listen to The Full Ratchet on SoundCloudNick Moran Angel List

Today we cover Part 2 of Artificial Intelligence Investing with Nathan Benaich of Playfair Capital. In this segment we address:

  • Can you provide an overview of the funding landscape for AI in recent years and the major things that have changed?
  • What are the primary sectors wherein AI-based startups are receiving the most funding?
  • Can you describe where you see some of the most interesting applications both in the enterprise and consumer spaces?
  • Who are the VC firms that are most active, investing the most capital in AI/Machine learning, both in the U.S. and outside?
  • You conducted research on exits for AI vs. that of greater tech. Can you highlight some of the key learnings from your research?
  • What, in your opinion, is most exciting about advances in AI?
  • What, in your opinion, is the biggest threat regarding AI?
  • What advice would you have to early-stage investors that have an interest in AI?
  • Anything else that we didn’t address that you’d like to touch on for the listeners?

Guest Links:

Key Takeaways:


1- The Pace of Progress

First, we started with a definition of AI, which in Nathan’s estimation is: The field of building computer systems that can understand and learn from observations without the need to be explicitly programmed. And that these systems can perform functions that are increasingly human-like. Nathan referred to AI as the umbrella term that includes all the disciplines and technologies required to build those systems.

Next we discussed the concepts that illustrate technological progress, contributing to the advancement of AI. Those two concepts were the Law of Accelerating Returns and Moore’s Law.

Law of accelerating returns: This was first discussed by Ray Kurzweil and says that as a society becomes more advanced, the rate at which it can progress increases. This is anything but linear. The rate at which technology has advanced in the past decade is much greater than the rate of advancement the decade before it, which is much greater than the decade before that. Here Nathan cited compute capacity and cost, which related to Moore’s Law; that maximum computing power doubles rougly every two years. Both of these ‘Laws’ have proven their merit as we measure progress and compute capacity over the past century. And they have significant implications on AI as it moves through different phases of development.

Those phases and types of AI discussed, included:
ANI, Artifical Narrow Intelligence: Intelligence around a very specific subtask. The intelligence system is narrow, focused on a specific problem and is unable to do anything else. Here is where Nathan cited the example of IBM DeepBlue which beat Garry Kasparov in chess. There are many examples of existing narrow intelligence systems.
AGI, Artificial General Intelligence: This is the goal of Google DeepMind. Intelligence from a non-human system is on-par with tasks that can be completed by humans. This could include reasoning, planning, communicating and even exhibiting emotion. And, this can be accomplished without the need for re-engineering, reporgramming or rebuilding. The experts from the science community generally believe that AGI will be reached within the next one to two decades.
ASI, Artifical Super Intelligence: As Nathan descirbed, this is essentially the point of singularity; the point at which AI is more intelligent than humans. This encompasses every field or activity that humans can accomplish and things beyond human capability. Creativity, wisdom, social skills, physical representation. An ASI system is one that is more preformant in every single task than the smartest humans on earth. And the experts predict this to be achieved by 2040.

2- Waves of AI
1st Wave: Rules Based Expert Systems. Effectively a series of if this, then that statements. Great for knowledge workers that have activity trees and tasks that can be preformed by software in a faster and more accurate fashion than humans.
2nd Wave: This is the Machine learning wave, where a raw training data set and a test data set is required; with one being tested against the other. Here Nathan talked about Supervised vs. Unsupervised approaches, the former providing examples upon which the data is tested against.
3rd Wave: Deep Learning. This wave requires massive data sets and significant processing powerWe also discussed the technology categories within AI including but not limited to:
-machine learning / deep learning
-predictive analytics
-natural language processing and semantic analysis
-speech recognition
-computer vision
-and others
And Nathan suggested that one considers which technology is best suited for the type of problem that being dealt with. Examples of problem domains included:

3- The VC AI Landscape

Nathan sees a future of human-machine collaboration instead of just machine replacement. Knowledge work, information retrieval, task automation, and business automation are all areas that he cited which are ripe for AI application. Especially as more SaaS services have become available that can acquire data and talk to each other.

Nathan mentioned that the majority of VC funding in AI started in 2013. Last year, $2B invested in AI… which is about 4% of all VC-backed technology and AI accounted for 1% of all technology exits. And the primary sectors receiving funding at this point includes business intelligence, predictive analytics, adtech, fintech. And Nathan also pointed out the health sector and predictive medicine. While most are focused on the threat of AI they may be overlooking the incredible benefits, not only to our efficiency, but to our well being.

Unfortunately the majority of the companies doing acquisitions in AI are not performing well in the public markets, which may restrict their ability to continue making acquisitions.

And upon wrapping up the interview, Nathan questioned if the current, broader VC category is well-suited to fund capital-intensive, long time-horizon technologies. It’s certainly a legitimate question as costs, risk-profiles, and return expectations have changed over the past couple decades.

Tip of the Week:   Changing the Asset Value Equation

*Please excuse any errors in the below transcript

Welcome to the podcast about investing in startups, where existing investors can learn how to get the best deal possible, and those that have —

Nick: Yeah what we’re talking about, sort of the investment landscape a little bit. Can you give us an overview of the major things that have changed in AI within the venture capital landscape?

Nathan: Yeah. So, you know, there’s been a pretty significant growth I’d say since 2013 onwards. Before then there really weren’t that many venture investments. Of course there were in the US in the 1950s a couple of government, of different government funding initiatives, whether that’s in the UK or in Japan or with #Darpa in the US which has been, you know, a very important source of funding innovation and machine learning and the US military. But really from a venture standpoint, it’s mostly from 2013. Actually 2015 is, has been the biggest year to date relatively speaking in this space where, depends on how you, how you define your search for the companies within AI. But, but the way I look at it is more from a, from a core technology perspective that’s important for, for achieving the goal. So I consider a company that’s, that’s using natural language processing and natural language understanding or computer vision as a company that is the AI space. And if you take that approach, then you see that in 2015 there’s around 2 billion dollars invested in the AI landscape. And before that in 2014 something broadly similar. But that’s, but that’s only about 4% of all VC backed technology unfortunately. So it’s still, it’s still a jump in the water, so to speak. But, but the growth since since 2013. It’s just that as a guy I think mostly motivated by the fact that a lot of companies are sort of realizing that they have droves of information about their customers and users and essentially want to able to extract more value from that and grow when, you know, large companies around them have started to adopt to machine learning technologies in their own products. And you start to see accelerating returns once you, once you do.

Nick: Well, more deal flow for #Playfair, at least for now, right?

Nathan: Yeah

Nick: So you already talked about some of the sectors in the area of focus and you put together a couple group reports on the funding landscape in different sectors that are receiving the most funding.

Nathan: Yeah

Nick: Can you talk about what those sectors that are receiving the most funding in the AI sort of landscape?

Nathan: Yeah. So from an enterprise perspective, again business intelligence and analytics and ad-tech and fin-tech are, are really taking up the lion’s share of financing. Fin-tech is kind of, kind of an outlier because what are these lending companies? They originally start as, as credit scoring companies, like # ZestFinance or #Avanse in the consumer space. But you know, very quickly once they get a methodology to more accurately score the credit risk of their customers they, they move towards being balance sheet lenders, so they raise, you know, hundreds of millions to be able to, to lend from their own balance sheet versus just to supply their score to a company that does have the balance sheet. And so, so when you look at it from a couple investor perspective fin-tech it often biases the distribution. But what w’ve seen in 2015 is also an emergence of just, you know, engineering companies that are trying to advance the technological landscape before they start really developing products. And these companies are, you know, more, more like the #DeepMinds of this world, but they often don’t have products that are, that are alive yet because the R&D cycle is particularly long. They often times become you know interesting acquisition targets for larger incumbents that want to buy in that know how. But, you know, they, they take a little bit, bit longer to, to mature. But certainly it’s, it’s you know, predictive analytics and DI you mentioned, you know, one big one that raised 55 million, and see an uptake in the robotics space raised 45 million, and then in tech and fin-tech. But yeah there’s still, there’s still a lot of white space to be created, you know. I think in the autonomous vehicle area where computer vision was absolutely seminal to enabling that to work and, and also the re-enforcement learning space for having robots to, to learn how to behave without, with, by just seeing examples. That’s going to be important. And in healthcare as well there’s been a few financings but, but still only, only a few ways that you know, potential benefit you can generate in that space is far outweighing the interest so far.

Nick: And your example of companies using credit score to acquire leads and then make smarter decisions about lending, it sort of reminded me a little analogous to #Zenefits, that’s using free HR software basically as leads for insurance. Do you know if they are utilizing any machine learning or any AI in order to better assess insurance rates for customers?

Nathan: I don’t know, I don’t know if they’re doing it actually. But I know that oscar insurance in New York has a team of data scientists that’s working on that task for their own purposes. So, so for sure I mean, as you try and span the market beyond those customers who have the requisite traditional, you know, data file to make a decision. And then you’re going to have to, you know, mine beyond that and mine in untraditional or non traditional sources of data that might have better coalitions with credit worthiness or with worthiness for any other service. So that’s where, that’s where companies in this AI space can really, you know, have an advantage I think.

Nick: Can you describe where you see some of the most interesting applications, both in the enterprise and in consumer spaces?

Nathan: Yeah, I mean, broadly broadly speaking I think, I think the most interesting opportunity is the one where there’s, where there’s an existing requirement to generate loads of high dimensionality, complex data to be able to complete that task. But where as a function of that requirement, this is way too much of it to be able to process it manually. And therefore also that the, that the value chain of processing and coming to a decision or coming to a result is, is very complex and can be automated to a certain extent. And, and another parameter I think makes, makes an opportunity, interesting or not, is one where the, the solution can be fed back on by the end user. So, you know, we talked on the underdog versus cats classification example that, you know, users can, can essentially see whether the responses were good or bad. I think there is a huge opportunity to be taken with designing experiences around specific problem that actually involves the end user in some way to, you know to extract that extra knowledge from that person’s mind which has been working on a specific issue for far longer probably than the engineer that built the system. So then really we can, we can create a, a product that has more performance than what machine learning can give on it’s own or what humans can deliver on their own. And so, so I like to think of those active learning situations where you have human machine collaboration versus just machine replacement. And so, so again I think knowledge work and information retrieval, task automation, business automation, those areas are, are really ripe for this. Specially as we see, you know, us using more SaaS services to get about our business every day. And part of these services can now talk to each other and pass data between each other. So we’ve already seen a generation of, of SaaS connectors where you can apply ‘if this then that’ rules. But, but over time systems should be able to observe the types of behaviors as we do with those different systems once they’re connected. And then start to, to automate and suggest new, you know, new things that we can do to make our jobs more efficient. And, and I touched briefly on the, on the medicine healthcare angle. But, but I really do think that we’ve come from a background where genetics and sequencing and studying all the, you know, hereditary components of disease have been, and health have been pretty well studied over the last, you know 10, 20, 30 years. But now really we’re at this point where we’re all, you know, carrying around super computers in our pockets and we’re interacting with digital products which, you know, leave this cookie trail of our behaviors and our knowledge and what we’re exposed to in the environment. And, and from that we can start to, you know, extract quantitative data that describes our behaviors. And I think with that you can go back to, to contributors to help the disease which is durture, so all the hereditary components which is sold by cheap sequencing for example. And nurture which is the behavioral aspects and the environmental aspects that feed into dictating whether we stay healthy or whether we fall , we fall ill to disease. And so I think in, in a couple years to come, there’s going to be this unique opportunity to assemble the two, you know, machine learning and , and other approaches have a huge role to play there. And to start to, you know, predict in what situations our, our bodies are in and whether we should seek medical care, we should modify our diets or change our daily routines to reduce the vulnerability with which we’ll develop a condition going forward. Because, you know, it’s all very clear that the health systems are completely drowning in costs and just can’t, can’t rely really or, or efficiently care for the body of patients that need it. And so coupled with the fact that we talked at the beginning of the podcast, you know, it takes billions of dollars to develop therapeutics and you’ve got all sorts of regulation and it takes loads of time that I think really like the future of medicine is predictive medicine and trying to catch things early versus trying to cure damage that’s already occurred. And I think AI is really really core to that becoming, becoming real.

Nick: #Nathan, who are some of the VC firms that are most active investing the most capital in the AI space, both here in the US and also outside of the US?

Nathan: Yeah. So you know this data is going to be, going to be biased towards the US because it’s something like 70% of deals last year and about 80-85% of the capital invested were in US companies

Nick: Yeah

Nathan: And you know, which is fine, it’s just a reality. But I think it poses, you know, very important considerations for where one should build the company and what markets one should get exposed to in the space. But in the US, you know, the top investors are #Andreesson, #New Enterprise Associates , # Data Collective , # Khosla Ventures, #Google Ventures, and # Sequoia. Essentially all those guys who’ve taken a part in at least, at least 5 deals last year. And then outside the US, I mean the most popular geographies are, are the UK and Israel. And there, at least from public data that we know of in the most active and participated in three deals each and , and #Playfair, us, we were one of them. There’s a trans fund called #Zen fund, and one in Israel called # Giza Venture Capital. So really like the, the playing field is, is still pretty barren outside the US and, you know, we’re quite happy to be, to be one of the investors who, who are taking, taking a head start there.

Nick: You did a bunch of research on exits for AI versus greater tech. Can you highlight some of the key learnings from that research?

Nathan: Yeah. So, so actually AI companies constituted about 1% of, of all tech exits in 2015. So there were 41 m&A transactions in one IPO vs the broader tech market which saw 3350 acquisitions in 61 IPOs. So yeah. Doesn’t sound too great. But again, it, you know, and important geographical, I guess reality is two thirds of the acquisitions were done by US companies. And most of the acquisitions, 70% of them were of US companies. But actually the, the part that’s a little bit worrying is that the US public, publicly traded companies who acquired businesses, and this includes, you know, #Twitter, #LinkedIn, #IBM, #Apple, #Credio and one or two others, one of which is #Google, have all performed really poorly this year in the public markets, like they’ve all lost market capital, raised billions of dollars of value. And the only exception is #Google. So I think it will, it will just pose a lot of questions as to whether those companies will be back in the market to make acquisitions this year. You know I, I hazard that, that these businesses will, will definitely take their foot off the pedal in acquiring companies, you know, more generally and I think, you know, this industry will probably suffer also because the, the network of, of acquiring companies is not that big. But the rosier side of things is that there were a few, you know, large outcomes, the biggest of which was a marketing company called #TellApart which was bought by #Twitter. And, and that one yielded, you know, a really solid return on capital. So exit prices of function of capital raise, that was 30 times. There was, you know, a few other exits that yielded 6 to 7, one of which was 10. So you, you do see, you do see good exits as a function of capital raised. So it shows you that large outcomes are possible. And the frequency of those outcomes is roughly similar to what we saw last year where the biggest acquisition was #DeepMind and after that was #RelateIQ for, for $380M, and then a few smaller ones, one of which was #The Echo Nest by #Spotify. So, so yeah, the conclusions are US acquires public companies not doing so well and most of the companies exiting were US based. But there’s still a jump in the water with respect to, to overall tech, tech exits. 

Nick: I’ll be curious to follow #GM after seeing their recent acquisition of #Cruise Automation.

Nathan: Yeah, yeah exactly. But you know, I think that the important thing to bear in mind is still a very, very nascent field. You know most of the money that we talked about, you know, came in since 2013 and you know with the average duration to exit of, of 6 years or more, I think we’re still to, to really realize the true value of these companies And those that were bought were still very early on in their product life cycles and probably haven’t even, you know, properly demonstrated the value they can bring. So I’m overall, still very bullish on this space.

Nick: Yeah. What in your opinion is most exciting about advances in AI?

Nathan: Just the, the far reaching potential that, that they have and to basically any discipline where you have, you know, a wealth of, a wealth of data and ability to properly and reliably extract value and complete tasks. And I think this stance, you know, everything from, you know, enterprise consumer ones, you know, basic things like stamp filtering, and we talked about information retrieval and search, you know, changing our behaviors with interacting with physical devices through and even gesture control and then moving towards autonomous vehicles and robotics and manufacturing. But it’s just, it’s just so far reaching and we’re still, we’re still really at the, at the early days of, of translating a lot of you know seminal research that’s been, you know, in the works for the last decade or two decades into commercial applications.

Nick: Unfortunately fund cycles are a pretty long horizon, right?

Nathan: Yeah.

Nick: And then, you know, many people think of AI and conjure up images of robots taking over the world. We’ve seen that in the entertainment space for many years. But what in your opinion is the biggest threat regarding AI?

Nathan: So, yeah, I think it is, it’s definitely important to , you know, consider and explore the way that the, the systems that we’re, we’re seeking to build today and the results that we’ve had could be implemented to ends that weren’t originally envisioned, you know. And it’s a part and parcel of, of developing some of the, you know, fundamental building blocks of the future. And we have a duty towards that. And I think having, you know, informed and rational discussion with technical leaders who actually know what it takes to build this future is definitely important. I guess my, my fear is sometimes that it’s, it’s still misperceived in the, in the public domain and I just wouldn’t want to come into a situation where, where this, you know, there is significant uproar and movement against the space because it’s been misinterpreted by, by the public and by people who, who don’t want to see a future that’s empowered by these approaches. Because I do think that it’s generally speaking positive for the world, and will, you know, will liberate a lot of us in various functions to focus on the areas that we’re most interested in and we can excel the most in by, by removing, you know, responsibilities and tasks that, that are otherwise, you know, quite boring and that can be better performed by software.

Nick: Yeah. And what advice would you have for early stage investors that have an interest in AI?

Nathan: Yeah. So, so I think you should, I think you should get into it for sure. But it’s certainly worth, you know, spending some time to, to read about, you know, the various boom and busts in the history since 1950s to, to kind of appreciate, you know, it’s heritage and what’s, what’s limited. The realization of converses in the past and, and what, you know, what considerations it has for us, for us moving forward. And it’s important to also dive into, you know, these various technologies that we talked about, you know, how they work and what task they’re best suited to solve and understand the requirements and limitations. And then I think the best way to, to really approach this is you know think about your chosen area of focus and, and seeing whether the problems that exist in that area of focus look like tractable machine learning problems where, where really that technology could be the best solution for it. But it’s always important to , to remember that, that you know technology differentiators work very quickly and become commoditized and essentially , you know, fit into this ever growing developer tool kit that’s, you know, publicly available. And I reckon that AI will largely follow the same, the same fate in broad strokes because, you know, over a series of a few months last year we saw a nudge in companies like Google, Facebook, Baidu, IBM, Microsoft, and others all open sourcing elements of their technology. So while it’s important to understand how these things work and how they fit together and how they can be tasked to solve the problems in your domain of interest as an investor, it’s important to also look beyond that and consider how to build compelling experiential products using those technologies so we can start to harvest data that others don’t have. And of course if you don’t want to do any of this then, you know, we’re happy to invest with you.

Nick: Yeah, fundamentally this is a technology that enables other things to be accomplished. It’s not, we’re not really at the ASI stage yet, right?

Nathan: Yeah, yeah.

Nick: And then, anything else that we didn’t address with regards to AI that you’d like to touch on for the listeners?

Nathan: You know, I’m , I think we touched on a lot of things. I’m generally really interested in engaging with people from all sorts of disciplines, because I think as we talked with AI techniques getting commoditized, it just means that more people can come in to play in this field and use productized solutions that have been developed by engineers to fulfill their own visions for how products should work to solve problems using machine learning approaches without necessarily to become machine learning engineers. So yeah, if there’s, if there’s you know, areas that investors think should be tackled by or could be appropriately tacked by machine learning and various technologies within that umbrella then, then I’m all ears I would love to fuel those thoughts. I think that engaging in discussions like this are some of the best ways to, to learn what’s next and meet some, some brilliant people.

Nick: #Nathan, if we could address any topic in venture, what topic do you think should be addressed and who would you like to hear speak about it?

Nathan: Yeah, I think the general question of whether venture is really still a risk taking business as it used to be in it’s inception. You know, I think that the nature of, limited partners and, and GPs and the macro environment has, has changed a lot in the last, you know, 40  or so years. And that’s certainly had an impact on the types of companies that get created, and the types that venture investors want to seek the fund as driven by pressures from LPs and from the broader macro environment. And so, that coupled to the fact that, you know, a lot of innovation that used to happen in universities is now, you know, going back into private companies or probably even public companies with large balance sheets of it. Like how #Bell labs and #AT&T were, you know, some of the bedrock of innovation back in the day. And so I, I’m quite interested in, interested in getting a topic as to whether venture investors still have the right risk appetite as they used to. And in terms of like who’d be the best person for that, I think it’s, you know, an individual who’s, who’s seen a variety of boom and busts and who has been there from the start. So perhaps somebody who is on the, something venture documentary on #Netflix.

Nick: You know we had #Adam Draper on the program and

Nathan: Yeah

Nick: And he was talking about his father and his father’s father. And just how a couple decades ago you needed a team of PhDs and you needed much more significant capital just to get to product market fit. And

Nathan: Yeah

Nick: And now you don’t. And so, I think it speaks to your point of has the risk appetites of VCs changed to try and invest in things that have already achieved product market fit and have very easy scaling versus the more ambitious core science heavy capital endeavors.

Nathan: Yeah , exactly. Well it’s probably , you know, the latter category which is, is clearly more risky but clearly the one that this asset class was created to support, and one that’s really going to build the better arc of our future. It’s not social media, analytics companies and predictive marketing and sales companies, right? But, but unfortunately the latter category is the one that generates traction and one that generates revenue. Which means that, you know, companies actually were something and, you know, they generate liquidity which is what LPs want. And so I just wonder whether that timescale is sort of liquidity events still getting compressed, which then means that GPs are funding certain types of companies because those are the ones that can reasonably generate that return, which then has a knock on effect on what interests founders have and what not.

Nick: And finally, just to wrap up, #Nathan, what’s the best way for listeners to connect with you?

Nathan: I mean, the best way is if we, if we have a mutual connection. Otherwise I’m pretty accessible on Twitter and all over the web. So I usually respond on Twitter if you ping me there.

Nick: Well, I’ve heard great things about you Nathan, and very much enjoyed the time, the couple times that we’ve had a chance to chat. So, thank you so much for carving out this time to cover the topic today. And hope we get a chance to connect in person soon.

Nathan: Yeah, of course, it was a real pleasure