Innovation and challenges in AI: a CTO's perspective

I studied physics during my undergrad and then began a PhD before dropping out – sorry, I mean “taking a leave of absence.” I accepted a job in finance, which brought me to London, but after almost a decade, I wanted to try working on some new problems.

I joined an entrepreneur network and that’s where I met Ulrik, my co-founder. We went for a couple of pints down at the local pub and Ulrik told me about his seed idea for Encord. I thought it was the best startup idea I had heard, better than the ones I had thought of at least.

From the general public perspective, people think that AI is scientists’ attempt at creating consciousness. They conflate all AI with AGI, artificial general intelligence, which is not the focus for nearly all applications. As a result, people attribute some sort of magic to AI and if they thought of it as just an additional statistical technique, I think that would demystify a lot of the aura around it.

They don’t design things for scale. They think short-term about having success at the prototype or proof of concept stage. However, once they put the model into the development cycle, they realise a lot of those choices that they made to get the proof of concept up and running don’t work for a production system. They need systems that are robust and work for real-world use cases.

By the value that the technology creates. If it works and creates concrete, tangible real-world value, then that is innovation. If you see a lot of venture money chasing a thing that never actually gets put into the real world, then that’s hype.

Like a lot of challenges, climate change depends more on our own behaviours than on technology, so changing our behaviours is the main lever we have to combat that challenge. However, a lot of interesting technologies are emerging, AI being one of them, to help tackle this challenge. Those solutions are not viable yet, but the technology can move. For instance, solar panel prices have fallen faster than anyone anticipated 10 years ago.

A lot of people are starting to adopt the technology and put it into real world systems, which is great, but from an academic perspective, the most interesting developments come from this new self-supervised phase of machine learning, so having data learn from itself rather than having to label everything. The potential impact of those developments on the future of the field is very exciting.