Any career gives you a particular insight into the world. Accountants seem to see the world in terms of money, doctors in terms of health and mortality, and so on.

I've spent a career creating AI software. This has meant also applying my software, and sometimes others', to client's problems. In the process I've analysed an enormous amount of data in wildly different fields. I've also thought very hard about the methods one might create to enable machines to understand the world, and thus for us to understand it better too.

In the process I've discovered commonalities in the way machines analyse data and the way that we do the same thing. This has made me think about the limits of what can be inferred, by human or machine.

Back in the 80s, when I first started doing all this, I particularly focussed on time series prediction. I discovered a class of problems that were effectively unpredictable. Sometimes things are unpredictable because you just don't have all the data. But in these cases they were unpredictable no matter how much data you had.

The first users of my software were mostly academics, and I was frequently invited to conferences where I was the only attendee without a PhD. I decided I needed one too, and found a new local university that wasn't going to be too demanding in attendance requirements, that was happy for me to do one.

I chose these difficult time series for my thesis, mostly because my investors saw no commercial application for that research.

Financial time series were obvious examples of these "difficult" series, but one of my software users, now a very eminent scientist, also had a very interesting example arising in biotech.

What was interesting about these series, was the prospect, not that they could ever be predicted, but that you might at least be able to establish bounds on their predictability. This means establishing how the uncertainty grew as you predicted further and further into the future.

So the thesis became about how you could identify such time series, and how you could estimate these limits on predictability.

Having completed the thesis, which gained a little notoriety, academics often sent me data to analyse, and I discovered how widespread these unpredictable series are. Time series prediction using supervised learning and tools from chaos theory | Andrew Edmonds - Academia.edu

I thus realised that the zeitgeist view that in the fullness of time all aspects of the world would be rendered capable of analysis and prediction was fundamentally wrong.

I've been trying to point this out to all and sundry ever since, especially in the field of climate change, but like Cassandra I'm doomed to be believed only infrequently. But what kind of person would I be if I didn't try?