Back when I was an EE graduate student, the Deep Learning (DL) excitement was very much in its infancy. Neural nets, which is the core class of models that DL is focused on, was taught as a subset of machine learning, as it was one of the many powerful techniques an engineer would use to make predictions using data. After graduating, I started applying these techniques by building machine learning applications for derivatives trading. I quickly learned something that my partner Rich Boyle already knew from his decades in the AI trenches: DL was going to be the catalyst for the next new wave of interest and innovation for companies to use Artificial Intelligence (AI) within their organization.
I spent most of last week at CVPR (Computer Vision Perception Research), one of the top academic conferences for computer vision. I am convinced that we are at another inflection point in the development of AI. Skynet? Not yet (phew!). But the collaboration between industry and academia is at an all time high, and the infrastructure (from Nvidia’s GPUs, Google’s TPUs, and frameworks like PyTorch and Tensorflow) has made huge strides in democratizing AI. As we look to invest in the next generation of successful startups, I believe that they will not only use DL and other AI techniques to build best in-class solutions, but we will also begin to see the emergence of “AI-first” companies, where AI moves from a feature to becoming central to the overall value and vision of the company.
As I listened to presentations on the next wave of computer vision models, and spoke with people from the largest enterprises to the next generation of early stage startups in autonomous vehicles, robotics, and AR/VR, I was fascinated by the technical depth of problems that they are tackling. The low hanging fruit of using a Convolutional Neural Net and a large scale training set to classify an image is over, as that is now a basic feature of their technology stack. Rather, the problems are now much more in-depth. Autonomous vehicle companies have enough data for a vehicle on a highway, and are now focused on the long tail of edge cases that infrequently happen (e.g. how do you deal with an angry pedestrian, or even a Bird scooter along on the road). Media companies are looking at the evolution of automated caption generation from still images to video. And robotics companies are looking at transfer models to increase the breadth of applications their robot can tackle. There were many proposed solutions at CVPR for each of these problems, ranging from building context-driven models that observe an object and estimate their intent, to building Generative Adversarial Networks (GANs) to create realistic training data for better simulation. The next wave of “AI-first” startups are going to be rooted in Deep Learning, but we are starting to see newer branches of AI emerge as the problems and challenges these companies are solving continue to evolve in scope and complexity.
As the applications and subsequent AI techniques improve, startups are also innovating on their business strategy in the rapidly changing ecosystem. I believe that as the stack (sensors, perception, planning, mapping, and control, and UI) for each vertical within AI companies matures, startups need to think more about their long-term defensibility, and how they can extract the most value for their innovation. The deep tech ecosystem is at an interesting point in time, where the leading companies in autonomous vehicles, mobility, and robotics are beginning to take shape. This is either through a lot of venture investment (Uber, Didi), or through deep partnerships with the legacy players (Cruise / GM, Argo / Ford). However, the strategy to solve the long tail of use cases are still ill-defined, and that’s where I think the the next AI gold rush of opportunity lies. The best startups are thinking about new ways of charging for their AI models, ownership rights to data between partners, and long-term defensibility in a world where “open source” and “free” is the new normal. The new wave of AI startups are not only defining the solutions with their state-of-the-art algorithms, but also the strategy that maximizes their value in the ecosystem.
It’s been 6 years since the 2012 CVPR when ConvNets smashed prior ImageNet benchmarks, sparking a whirlwind of Deep Learning and AI activity. A lot of the advancements since then have been focused on particular solutions, where companies used AI for very specific tasks within their organization. I believe that phase is ending and we are beginning to scratch the surface on the next evolution of AI, not only in the algorithms, data, and models used, but in way that startups extract defensible value for their innovation. At Canaan, we are excited for the partnering with early stage startups and founders in deep tech that are pushing towards being “AI-first” companies. If that’s you, let’s chat!