Whilst everyone is getting super excited about AI and is predicting a singularity. Let me state a few things I learnt from time spent running an industrial tech start up (all our clients were $100bn+ industrials).

AI does not trend to infinity.

Optimal AI capability is not infinity, it is maximum profitability.

ML is expensive, data access layers (at this scale) are expensive.

The bottleneck for us was not CPU’s it was getting petabytes of data into memory. We did it with lots of sockets and a big fabric.

You cannot move the necessary volumes of data via the internet. We had to move data by trucks, it was faster, a lot faster.

Back to economics… let’s say it takes you $10m to train a model that can perform a task with 98% accuracy, typically it would take $100m to reach 99% and $1bn to reach 99.5%. At some point you have already captured all the value in the addressable market and further training isn’t just uneconomic, it is wasteful.

Now lots of people argue that AI will create myriad new markets and added value, this is largely BS. Yes it will do these things. But nowhere near the levels people are taking about. AI will expand very quickly to displace people and companies from existing addressable markets but expanding beyond that is very different and faces a multitude of headwinds many of which are demand limited and beyond control.

It’s usually not worth spending billions chasing the last 1% of the error. But there are exceptions such as autonomous cars, etc.

Understanding the size of the addressable market and the competitive error rate for that market usually tells you how much investment is worthwhile to spend on AI.

I suspect the next economic cycle will see a huge AI bubble, where people think infinite intelligence means infinite ROI. This is wrong. I’ve lived it. It’s a very hard won lesson.