Defensible Deep Learning Ventures
Venture capitalists (VCs) have a new term “Systematic Intelligence” that is worth paying attention to. Gil Dibner wrote an excellent article (“Systems of Intelligence: Is this the VC meta-thesis we’ve been looking for?” ) that breaks down the current thinking on defensible businesses. This is an important matter for ventures that focus on Deep Learning. Jeff Bezos once remarked “Invention is not disruptive, only customer adoption is. At Amazon, we’ve invented many things, most were not at all disruptive!” So, it is not enough to invent, one needs to be able to be disruptive. However, disruption is not enough, one needs businesses that pass the test of time. That is, businesses that are defensible.
I’ve discussed this earlier about the need to focus on a platformization strategy. That is, we would like to build businesses that avoid being disrupted. To do so, we need to build platforms that are able to leverage networking effects. There are four kinds of platforms; these are marketplaces, social networks, supply networks and creation networks. However, as Gil Dibner argues, the opportunities to discover new kinds of platforms are dwindling and therefore we need to discover new ways for building defensible businesses. Please read Dibner’s article before proceeding. His article sketches out the line of thinking that leads to the current situation.
Dibner enumerates several true barriers of entry that a Deep Learning venture needs to attempt to erect:
Vertical Domain Expertise
There many businesses that require unique knowledge and expertise to perform competitively. These are hard to find because businesses will keep this information very close to their chest. However, finding these businesses and enhancing their processes with advanced Deep Learning methods can be a route to something that will continue to be defensible.
There are many systems that are extremely complex where there are a multitude of edge cases that need to be addressed. This can be a very treacherous space because many times the complexity is not really intrinsic to the problem but rather due to a lot of technical debt being accumulated over a long period of time. A venture thus has to be extremely careful, otherwise they can find themselves in a hamster’s thread mill. That is, performing a lot of work but not really going anywhere.
Hybrid Human Machine Interaction
This is an evolution of the idea of a killer user interface. Many of the deployments of Deep Learning are likely to be hybrid systems that will involve humans-in-the-loop. Therefore, to ensure adoption of these new systems, a lot of innovation needs to happen with how users interacts with the AI within a complex workflow. A lot of user interface tends to focus on the individual user, however companies like Atlassian understand a very different aspect in that there is extreme value in supporting complex workflows.
Here’s a perfect example of Deep Learning and how it can be used to interact with humans: https://aiexperiments.withgoogle.com/objectifier-spatial-programming
This is an extension of the previous ideas but extending it across multiple organizations. As an example, when intelligence is embedded across a supply chain, what sort of new kinds of efficiencies can we discover. This is an extremely compelling strategy. In fact, we will be unveiling our technology in this area that addresses this kind of a business model.
This is where one makes a heavy investment in a specific unique technology such that alternatives are simply not good enough to be considered. For smaller ventures without the outsized resources, this is very difficult to achieve without gambling on an entirely unique and novel approach. We actually see this in the AI and AGI spaces where you actually see a vast variety of approaches to AI (see article on Tribes of AI).
My opinion on this is that unless the approach leverages current advances in Deep Learning it would be a huge gamble. That’s because Deep Learning is the only AI approach that has shown to deliver exceptional learning capabilities. Progress therefore will be most probably in the ‘adjacent possible’. Of course that does not mean that something else can emerge out of ‘left-field’.
Deep Learning still has many flaws and I characterize the work in the field as being similar to what we find in Biotechnology. That is, much of the work is experimental and that just like Biotechnology, a revolutionary advance may be discovered by accident. I think many VCs miss this characterization entirely. They cannot imagine that software development can be so haphazard. But that’s the reality (see article on Black Magic).
This is the most common explanation of how to achieve a defensible strategy. There have been many startups in AI that have been acquired, not because of unique technology, but rather by having a head start in collecting a massive set of data. There definitely is a long tail here since its literally inexhaustible as to what kind of data we can collect.
These are six strategies that we should be continually exploring as we search for new business models in Deep Learning. A lot of work in Deep Learning is done in the trenches of research labs and they continue to astonish both researchers and observers. However, what is sorely lacking is the examination of how to build products using Deep Learning.