I often tell my students not to be misled by the name ‘artificial intelligence’ – there is nothing artificial about it. AI is made by humans, intended to behave by humans, and, ultimately, to impact humans’ lives and human society. – Fei-Fei Li
A remarkable number of conversations that I have are now about AI. Whether it’s on my social media feed, or in a professional capacity, everyone is talking about AI and what it means for us.
I think AI has transformed our lives in two ways, one that we don’t notice and one that we do.
We rely on AI for our day to day activities far more than you might think. Google maps uses AI. So does Spotify. As does that scheduling thing that tries to work out a time when everyone is free to talk. There are lots of tasks that are being speeded up and because we’re using the tools without thinking about them we don’t realize that AI is embedded in more and more of what we do every day.
The second way is the visible one – the models you can talk to and which help you with intellectual work – as a sort of research assistant.
That’s again been immensely useful. Want to write a python script or do an analysis in R. Here you go. It gets you going, gives you a starting point, that often works. It sometimes needs fixing and won’t always follow what you want, but it’s certainly faster than reading the docs and starting from scratch.
It’s not human. Remember that. I once asked for a script that would go through some data and pick out numbers that looked odd. I meant odd in the sense of outliers or unexpected patterns. It thought I meant numbers that weren’t even.
Going from this step to engineering an AI workflow that does something more is a little harder. I have a decent workflow for document analysis, something that comes up fairly often.
But after that we’re in a bit of a gray area – between stuff that needs thinking to stuff that needs a lot of mathematics. New models like o1 are supposed to be better at thinking type jobs but we run into the issue of validation. When your maps app gives you a route you know it’s working if you end up where you wanted to go. When your AI tool sets out a strategy you have to follow it for it to work – and we run into the human problem. If you succeed, how do you know it was because of the strategy? And if you fail, how do you know if you did the strategy right?
When you’re managing people you try and train them as best you can. Then you let them get on with the work and you try and check in, make sure they’re on the right track. Some managers micromanage, look over their worker’s shoulders and tell them what to do, but that’s like a prison warden and prisoners. Both are in prison. A good manager should be able to go and read a book knowing that the team is doing the work and it will be done right. The point of checking in is to validate what’s going on.
I think that word – validate is an important one.
We need ways to validate what AI tells us, an ability to test its outputs and treat them with some scepticism until we see outcomes that suggest we’re doing the right thing. Validation is about having a mental model that tells us what to expect from the AI.
For a long time we’ve talked about digital twins – digital models of physical processes.
I wonder if it’s time for a new approach – an analog twin.
A brain tool that can help with talking about what we need from our AI assistants and validating if they’re doing the right thing.
Something that we can understand.
Such tools exist – they are mental models such as purposeful activity models from soft systems methodology (SSM).
They’re just not very well publicised.
Maybe their day will now come.
Cheers,
Karthik Suresh
Handcrafted insight 