Is deep learning a disruptive or sustaining innovation?
Dissecting the deep learning trend with Clayton Christensen’s disruption theory
Disruption is an overused term today. It is the favorite word in startup pitches and is used to describe any hot new technology. One can no longer make meaningful prediction using the theory if almost everything in startups is disruptive. I’m optimistic about deep learning but somehow, I feel that the current buzz with deep learning (or AI) and framing deep learning as disruptive is misguided. For the purpose of this article, let us go back to the narrower definition of disruption and examine the recent trends through the lens of disruption.
Clayton Christensen’s disruption theory and value creation in startups
Clayton Christensen’s disruption theory describes a scenario where a smaller company with little resource is able win a large incumbent in a market where the incumbent has established a strong position. The smaller company leverages disruptive innovation, which is technological innovation that give rise to products that are inferior to the mainstream product on traditional performance metric, but typically with a smaller price tag or more convenience to use. The smaller company typically gain a foothold in small emerging markets with disruptive innovation and then move upmarket to challenge the incumbent when its product become good enough for mainstream customers. Disruption is complete when mainstream customers start to accept the smaller company’s product.
The disruption theory put innovations into two buckets: sustaining innovations and disruptive innovation. Incumbent has no problem adopting sustaining innovation that improves its product and make it even more attractive to their most profitable customers. Disruptive innovation on the other hand give rise to products that initially under-performs its current offering on traditional metrics, and does not satisfy the needs of its most profitable customers. Incumbent is also not efficient at discovering small emerging markets as small markets cannot generate the growth and profit needed at large company. They are often unable to adapt to lower profit margin offered by disruptive innovation due the cost structure developed at the company over the years. For example, an incumbent cannot survive with a new product with 10% margin when 20% of the price is used to support the sales. (read more about disruption theory here and Christensen’s The Innovator’s Dilemma).
A common misconception is to think of “disruptive innovation” as “radical innovation” and “sustaining innovation” as “incremental innovation”. The fact is that the magnitude of technological improvement is not the point in disruption theory. A technology can improve a process by hundred times and is still a sustaining innovation because it helps the incumbent in sustaining its market.
Disruption theory predicts that companies who leverage disruptive innovation have relatively higher chance of success against large incumbent companies. A startup is the perfect vehicle to take advantage of disruptive innovation. Disruption is not the only way a startup can succeed. Successful startups at some point must defy theories and accepted wisdom, including disruption theory. Uber is a great example of successful startup that did not start out with disruptive innovation. Clayton Christensen also wrongly predicted with disruption theory that the iPhone will not be successful.
Deep learning is a set of machine learning techniques that are behind recent breakthroughs in many domains such as natural language processing, image recognition, speech recognition, game/chess playing, etc. Deep learning has broad impact across different areas, to examine it with disruption, we need to break down the different contexts where deep learning is applied.
Deep learning as a better machine learning is a sustaining innovation
Machine learning (ML) has been around for quite a while. Before deep learning, ML were already used for machine translation task in Google translate, speech recognition in Apple Siri and Google assistant, face recognition in photos albums (Facebook, Google). ML is also closely related to the rise of data science where analytics are applied to and promised to improve business processes everywhere. If we consider the two defining characteristics of disruptive innovation: worse initial performance and gaining a foothold in small emerging market, deep learning has neither of them. In existing ML markets, deep learning delivers superior result compared to traditional ML algorithms across many domains.
Indeed, deep learning has replaced traditional ML techniques in Siri, Google and Facebook photos, and Google translate, and many other areas of ML. It is a sustaining innovation as incumbent readily adopt deep learning to improves their customers’ experience. In business analytics which typically employ ML methods, deep learning is not widely adopted yet because it often the businesses does not have to use deep learning or simpler ML models are good enough.
Deep learning applied to verticals
Vertical AI startups are companies that solve problems for a specific industry using AI. Deep learning has enabled new AI capabilities where previous ML methods failed. Vertical AI startups are likely to be using deep learning as they tackle hard industry-related problem. An example in agriculture tech is Blue River technology where computer vision is used to target pesticide to individual plant in the field. Blue River technology was recently acquired by John Deere, an incumbent in agriculture.
I would argue deep learning tends to be sustaining innovation in the context of vertical AI startups. The defining attribute of vertical AI startups is full stack solution for a specific industry. Disruption begins with a technological innovation looking for emerging applications because it does not fit well in current mainstream market. On the other hand, vertical AI startups often have a set of specific industry problems and employ the best AI (deep learning) methods available to tackle them. In practice, vertical startups may use a mix of both approaches: picking the concrete problem and the technology at the same time. The disruptiveness of deep learning depends on which approach the startup bias to. Starting with fixed industry-specific problem is sustaining innovation in nature because the lack of searching for a new market implies it likely to be serving the incumbent’s mainstream customers.
The disruption theory predicts that incumbents are very efficient at adopting sustaining innovations. It might be attempting to think the defensibility of deep learning vertical startup comes from the deep learning expertise that is inaccessible to the incumbents because they are not in operating in the AI space. The reality is they are good at deploying their vast resource to acquiring the talent and embark on research efforts necessary to improve their products along current performance metrics. Increasingly, they acquire startups to be good at deep learning.
This trend of acquisition of deep learning startups by incumbent is in line with Clayton Christensen’s prediction that the most efficient outcome for startups working on sustaining innovation is to be acquired by an incumbent.
Another often cited defensibility of vertical AI startups is the depth of the domain knowledge and proprietary data integrated into the AI product. It is true that another startup cannot easily replicate the domain knowledge and data, but the incumbent often have plenty of those. The vertical AI startup is competing with the incumbent internal research department. It should be take an unique angle to the problem to make itself valuable (as an acquisition target) in the vertical.
Then when is deep learning truly disruptive?
At this point in the essay, I feel obligated to make predictions about the next disruptive deep learning trend. But I have no such prediction and it is not the goal of this essay. This essay hopes to encourage startups founders, especially those dealing with deep learning to see through the hype and think through the lens of disruption theory.
Deep learning startups can create lots of value in different ways using deep learning as both disruptive or sustaining innovation. It helps to be discerning on which trajectory you are aiming for and formulate your strategy around it. Being sustaining means that market discovery is straightforward but competing on traditional performance metric with incumbent is hard. Being disruptive means taking the risk not finding any market but the startup enjoys a head start against incumbent once it finds a product-market fit. Both trajectories are scary in its own right. Startups are scary.
Disruptive or not, the bottom line is to build something truly unique and interesting. Because, without that there is neither a basis for disruptive competition with incumbent nor value to sell out to incumbent in the department of sustaining innovation.