Whilst I was studying at university, neural networks were a curiosity item to which we dedicated no serious time. It was still normal to hand-differentiate and code the entire network, a laborious process which greatly limited experimentation and scale.
I’ve always been mesmerized by the greater field of Artificial Intelligence. It carries with it such grand aims, of creating vast powerful intelligences, and has a rich history of interesting, awkward attempts to achieve them.
Over the years I’ve followed the field’s progress, devouring the latest papers with the curiosity of a child. Each successive achievement seemed like a little bit of magic.
With the dawn of our current era of neural networks, the pace and breadth of new achievements has become astounding.
The tools we’ve discovered are now remaking most of the fields that they touch: Voice recognition, language translation, face recognition, scene understanding, image subtitling, to name a few. Many previously impossible tasks are becoming possible, spawning new fields and businesses.
We’re amidst an international land-grab, as scientists and businesses race to apply these techniques to new problems. Rarely is there so much fertile ground to explore; we may look back on now as a golden period of exploration.
As new discoveries are re-combined with each-other, a vibrant kaleidoscope of new technologies has appeared before us, its depths unlikely to be reached for quite some time. It is hard to imagine the magnitude of disruption this will cause, as new technologies ripple and reflect across industries.
With this backdrop, I feel it is a fertile time to make real progress towards general intelligence. More specifically, towards reasoning and understanding:
Octavian’s mission is to create software that can reason over and understand information as well as a human can.
To make this more concrete, some milestones along the way are to:
- Answer questions requiring recall and reasoning over knowledge (“What color is France’s flag?”, “How many countries have blue in their flags?”, “Would reducing the speed limit save lives?”)
- Answer ambiguous questions (“When’s the best time to visit London?”, “How did the UK economy perform last year?”)
- Capture knowledge from writing and speech (e.g. turn television, textbooks and blogs into machine-knowledge)
Creating such a system would be transformational for humanity; it removes key barriers to information and enables us to think bigger and faster. Just as machines can free our time from manual labour, software can free us from analytical labour.
In a later post I’ll share a more detailed roadmap of the immediate goals we are working towards.
We have a set of beliefs and approaches to help us achieve this mission:
- We believe that our systems should store and process information as a graph
- We believe that architectures quite different from today’s (possibly ones that do not easily fit current hardware) need to be explored
- We believe that hand-crafting models is unlikely to be adequate to reach the level of complexity and sophistication required
- We believe the best way to direct research is to try to solve realistic problems
- We love synthetic datasets as stepping stones to solving real world problems
- We are happy to commercialize discoveries to help fund further research
- We freely share our code and discoveries
- We build things that are helpful, not harmful
Achieving this mission will take a lot of time and effort; there are many uncertainties along the way. We’re excited to both work towards it and the interim goals, as they themselves are valuable and worthwhile.
If you are interested in learning more, funding or working with us, get in touch.
Thanks to Rebecca Ross, Xander Dunn, Andrew Jefferson for providing feedback during drafting this article.