I did get a fair amount of feedback from my previous post. That, combined with some learning I’ve done in the past couple of months, leads me to continue on this topic.
I want to start by making a distinction between speculators and investors. My personal definition is related to time: the less time an investment is for, the more speculative it is. A day trader speculates that there will be someone buying what she just bought for a higher price, quickly, not because there aren’t any changes in the fundamentals of the business.
For an example at the other end of the spectrum, take a look at Warren Buffett's investment in Coke. He is getting about 40% yield from dividends now. He doesn’t need to sell, ever. A truly great investment, totally tied to how the business performs.
The nature of Venture Capital is closer to a speculative investment, rather than long term, as they tend to operate on a rather short (2 to 4 year) time frame. The best outcome for a VC is selling their investment to a large company for a fat premium. With this in mind, VCs buy shares of startups thinking of selling. Will another company buy this startup? If the answer is “yes,” invest. (An alternative exit would be an IPO. But I don’t know of any IPOs in the AI industry.)
The elapsed time to exit for AI companies is relatively short. They take about 5 years, compared with almost double for SaaS companies something which sparks the speculative instincts of the VCs: the shorter the better.
The problem is that there are few buyers, mostly the tech giants (although TD recently bought an AI company). And the tech giants don’t buy AI businesses, they buy people, (these are called acquihires, buying a company just to hire its staff). This is strange, since once a company has been bought, 50% of the staff leaves to start a new one and try their luck again (think of Apple buying VocalIQ. Shortly thereafter, the team left to start Prowler.io).
Sensical or not, the tech giants have an appetite for acquihires, so in the age-old cycle of supply and demand, VCs will supply them with what they want in the interest of making money on their investment.
In my previous post I question why VCs invest in companies that are clearly not viable, just because they have many PhDs in their team. This has been simply answered by the above: because that is what would sell. By one account, on average, each PhD is sold for 8 million dollars. Can you blame VCs for investing a million or two on a company that can attract 10 PhDs and sell for 80 million? It’s like blaming a dog for doing tricks to get a treat.
VCs' understanding of AI
Something else I have learned lately is that VCs have a narrow understanding of what AI is. For them, it seems “AI” means going over a lot of data and using it to draw some insight. For example, this can mean going over a few million credit card transactions labeled as ‘fraud’, and ‘not fraud’ and training a neural network to predict whether future transactions are fraud or not.
Since the algorithms used for this are well-known and available as open source, then the code has no value. It is actually not too hard to scramble together an app that recognizes handwriting with very few lines of code. So the value must be in the data. Whoever has the data then has the value. But if the data is easy to generate, then it is not good. Thus the term ‘Defensible Data’, which is what VCs like to see in AI companies.
This type of Machine Learning has real value and lots of exciting applications, but I’m thinking the best is beyond learning from data. Learning from interacting with an environment leads more easily to ‘AI that makes decisions’, instead of ‘AI that classifies’.
Decision making is incredibly valuable. Here’s where you have self driving cars, which is arguably the most valuable use of AI in the near future. But think of robots helping with all sort of chores and tasks, while learning more skills. AI for decision making is the holy grail! (Let’s not think of Terminator and Skynet for the moment!)
Learning from the environment has been studied for the last 20 or 30 years and it is called Reinforcement Learning (RL). Canadian legend Richard Sutton is the father (ok, grandfather) of it. A few weeks ago I was talking to a leading researcher of Deep RL (which is RL which uses multi-layer neural networks as a way to generalize a function) and I asked “Is there a book you can recommend? I already read the one from Sutton.” He answered: “No, that’s it. Other than his book, there are just some papers published from time to time”.
My point is that the algorithms are far from known, let alone implemented, let alone open sourced when it comes to the most interesting part of AI: reinforcement learning.
VCs focus on the number of PhDs and the quality and quantity of ‘defensible data’ for their investments, ignoring the most exciting areas of AI. To be fair, it is very hard to find a company that will actually turn RL into a viable business - there is little precedent of this - so VCs are acting in a predictable way. I just wish there was more domain knowledge from VCs, more creativity, and more innovation, rather than simply following trends.
Vertical AI vs. Full Stack
Another interesting aspect of AI and investing came up while attending an excellent presentation given by Narbe Alexandrian from Omers Ventures. He was explaining that there are Innovation AI companies that help incumbent companies compete and improve. For example, Zestfinance helps Ford assess credit scores, optimizing lending for leases and car loans. These are called ‘Vertical AI’ companies. VCs don’t like them a lot, they think that the customers (in the example, Ford or banks) will develop their own machine learning teams, since they own the data, sending the Vertical AI partner packing. Remember, they see no value on the algorithms, only on the defensible data.
I disagree with this. In general, companies tend to specialize on what they do well and are happy to buy from 3rd party experts what they need. Otherwise, banks would have developed their own database software, hardware and even their own chairs and desks a long time ago. IBM and Staples are hardly under threat of banks doing this, rather, they make a lot of money off these organizations. To me, the argument of ‘AI is too important to outsource’ is funny. I think chairs and desks are very important too.
On the other hand there are the ‘Full Stack AI’ companies that want to displace incumbents by the sheer competitive advantage that having AI gives them. An example of this is Kreditech, which is basically a bank that was started with AI at its core. In this approach, rather than helping existing banks use machine learning, one starts their own bank. They then compete against the incumbents with the edge of using AI while their competitors are stuck with their old mainframes and struggle to hire Cobol programmers. Full Stack AI companies are super cool and exciting, because they accelerates disruption with a clear benefit to customers and society.
I can think of a fun example of disrupting incumbents with software. Remember Blockbuster? It is now known by being disrupted out of existence by Netflix. But it started as the disruptor.
Blockbuster was a single video rental store that actually adopted software for its operations. A single PC could be used to track all the inventory and customers’ juicy late fees. This was so successful compared with hard work of keeping track with paper files that Blockbuster expanded worldwide, crushing the competition.
Despite all this, one wonders how realistic is for AI companies to go around starting car companies, banks, supermarkets and farms. In the long run, I believe more in Vertical AI that specializes in AI for a specific industry, rather than the Full Stack.
From experience here at Decisive, I can see the appeal of Full Stack. The video game industry is particularly dynamic, tech savvy and innovative. And yet it is hard to explain to them the value of using AI instead of hardcoded decision trees for their Non-Player Characters. I can only imagine how hard is for Vertical AI companies to talk to banks, insurance, hospitals, etc.
The crux of the matter when it comes to investing and AI is, how can AI be used to provide value to customers and actually get paid for it? The only ones doing this successfully are companies like Google and Facebook (I actually see those two as pure AI companies). Others will, but there will be a lot of trial and error, which makes investing very risky and difficult.
But fun too.