The Light Cone Strategy for AI
The pace of innovation in AI is accelerating. Deep learning is being fueled by exponentially scaling compute, sensor proliferation and massive inflows of funding, driving research into better and more efficient architectures and algorithms.
For anyone building in AI, selecting problems, setting strategy and creating a product roadmap in an environment where the underlying technology foundation itself is evolving so quickly requires thinking differently.
Let’s borrow and adapt a concept that helps physicists think about relativity – the light cone.
A light cone maps out the circle of influence that an event can have on the universe at some future time. The speed of light is fast, but finite – it determines the slope of the cone. Things that fall inside the cone in the future can be affected by the event, while things outside the cone can’t, because light – and therefore information – can’t reach them yet.
To adapt the light cone for AI strategy, we need to tweak it a bit. Let’s replace the speed of light with the speed of AI research. The slope is steep and the cone is wide because AI research is moving quickly.
The cross-sectional circle that it maps out on the present is the application boundary.
Let’s rotate this and look down at the plane of the present day in 2D. AI market maps, like this massive one from Matt Turck, show this view. Things outside the circle – beyond the application boundary – are current AI research topics. Things inside the circle are AI applications that can be commercialized now. As you go deeper into the center of the circle, it gets more competitive, applications are more obvious and technology quickly becomes commoditized. At the boundary, applications are less obvious and the technology more difficult and differentiated.
The last few decades were characterized by innovation in the form of big step function technology shifts – the web in the ‘90s, the smartphone in 2007 and the cloud at around the same time. Each step changed a paradigm. But, the new paradigm then held fairly stable for the next decade, usually more. Building in the current paradigm and executing was sufficient to win.
AI is more than a step change. It’s a phase change. It’s like we’re moving from a solid to a liquid. New laws of physics apply.
I believe large transformer-based foundation models triggered this phase change. With transformers, AI has its first, scalable general-purpose brain. Read more about why I believe this in my post on transformers here. I expect the pace of AI innovation to only speed up from here.
Traditional business strategy asks for a scalar. You pick a point inside the application boundary, draw out a circle and conquer it. That circle dictates problem selection and shapes your roadmap. Over the arc of time, your company carves out a cylinder.
Light cone strategies ask for a vector. Can we find not just a point, but a radial line pointing out to the application boundary that will put us on the natural glide path of the cone? If we can grow along the cone, then our circle of opportunity naturally expands, and our technological differentiation doesn’t erode as the application boundary pushes outward with upcoming AI innovation.
Light cone thinking is contrarian. It might mean choosing to take on problems that aren’t obviously connected, when looking at the plane of the present, but are likely connected when you look along the cone, toward the future.
For example, at Arena, our view of the light cone directs how we select and where we see commonality in problems. We select problems that not only create 10x impact for our customers today, but naturally push our technology outward, along the cone in a specific direction e.g. more active learning and the integration of new industry-specific modalities.
While some of our applications are at the boundary today, we operate on the belief that we’re creating a compounding vector advantage as the AI innovation slope continues.
Thanks for reading Pratap’s Substack! Subscribe for free to receive new posts and support my work.