Unfolding Origami Labs
After months of introspection and talking, we’re finally launching our very own company. At the risk of being self-indulgent, we want to put a few words down to capture why we’re doing this (and why now in these rosy economic times) and what we want to achieve. We’ll be using this space to discuss ideas and share research, but for now, here’s some words from us.
Chris: The growth in artificial intelligence from a niche research field to a global agent of change has been astounding in terms of pace and technological progress, albeit not always positive when looking through a societal lens. Advances in the mainstream have largely been driven by a proliferation of deep learning techniques aided by continual hardware performance gains.
However, throwing ever increasing volumes of data at power hungry black boxes is not an appropriate solution in many domains, including defence, security and regulated industries. What if data is hard to come by and the environment changes rapidly? How do we satisfy needs for transparency, assurance and trust?
I’m not suggesting we abandon deep learning but I feel its real world utility will peak below the claims of the hype. What then? How can we augment the undoubted power of deep learning while mitigating weaknesses? Maybe the way forward lies in combination with symbolic AI or other branches of a field that’s much more diverse than just machine learning. One of my goals for Origami Labs is to explore these possibilities and find a path to real world utility in cyber-physical systems. Perhaps combine ML with representations of human knowledge, after all we’re pretty good at what we do as a species (for good and ill). A kind of development-time human-machine teaming to complement a powerful run-time partnership?
Thom: Twenty years ago I was writing chatbots for IRC, mostly to annoy my friends on our 56k modems. They were simplistic and extremely limited, but the idea of a computer reading and responding to what I wrote was exciting and wedged its way into my brain.
A decade ago I was working in a research lab using high power desktops to detect faces in video. It was slow, inaccurate, and magical - a handful of simple binary wavelets could respond in such a way that we can locate objects (if unrotated and well lit) in images!
Today, I used my phone to ask a computer on the internet to generate a picture of the optimisation process in origami form. I got slightly annoyed that it couldn't capture exactly what I meant about the abstract concept of 'optimisation'.
This rate of progress is absolutely astounding, and is why I want to focus my time and effort into understanding what it means and what happens next. The impact of AI & Autonomy in the coming decades is likely to be far greater than we expect or understand. I want to create both value and joy from these systems, and to anticipate what problems we will face as they develop.