Professor Gans is one of the world’s leading economists on AI and his most recent book, co-authored with his colleagues at the University of Toronto, is the acclaimed Power and Prediction: The Disruptive Economics of Artificial Intelligence published by Harvard Business Review Press.
AI is really at its core an advance in statistics, in particular the statistics of prediction. What these new techniques have done by virtue of clever algorithmic design and also a massive amount of computing power is to start to dramatically drop the cost of prediction. And every other time we’ve had that sort of transformation, when you’ve taken something that’s quite costly and made it really cheap, there’s been a whole lot of flow on applications and progress.
For things we’re already predicting, like the weather or traffic, then we can improve our predictions and do it better. But the more interesting applications are where we didn’t previously realise that the problem we were facing was a prediction problem, such as self-driving vehicles.
The biggest gains of all, however, come not from mere applications but when you get true transformation, when an entire industry is essentially reorganised by an input becoming cheap. Historically, think of how the taxi industry was effectively completely upended by the mobile phone. No one saw Steve Jobs introduce the iPhone and said “well, that’s it for the taxi industry”.
But these transformational impacts will take time because human reorganisation always takes time. AI will improve prediction, but it won’t necessarily improve all the stodgy human stuff that slows down transformation. The real exciting developments for AI, which we’re starting to get a glimpse of, is still to come in ways that we can’t quite anticipate at the moment.
There will be environments in which it’s a good way to go to completely reorganise and design a technically superior system. But, more typically, it will be about tacking on automation and that will be quite difficult. It will be case by case based on the benefits of AI prediction versus all the costs that come from integrating this prediction. You will have to decide whether the juice is worth the squeeze.
For example, I recently visited the factory of Lego in Billund in Denmark. There is lots of automation in producing these billions of different types of bricks, 24/7. We could use AI to do better demand prediction for different models of Lego and send that back to the factory floor in an automated way so it can reconfigure and churn out the right bricks accordingly. But think about all the other moving parts, getting it onto the shelves etc. Lego has a very fine-tuned process already and you need to make sure your new overall system operates better.
Furthermore, even where it is appropriate to reorganise the whole company for AI, you’ll still need some way during the transition to take care of all the exceptions through human intervention when AI makes mistakes. AI can perform well, but only within the bounds of the laws of statistics. So if you’ve got model instability or bad data, it’s going to do bad things as well.
Well it’s been hard to sit there during the last year or so and not just see that there’s all manner of exciting things going on. Firstly, AI, and generative AI especially, has the biggest benefit through allowing someone who doesn’t have skills or experience to mimic the productivity of somebody who does.
Sometimes it’s given the term “upskilling”, but it’s not really upskilling because no one’s learning any new skills. It’s just that the skills you previously had to acquire through years of training have now been handed to you. Much like when you’ve got a navigation app on your phone, you can now perform at the level of a London taxi driver who has acquired the knowledge. That’s an interesting set of effects occurring all over the place now and it’s quite incredible.
For businesses, they’ll have an easier time hiring and expanding and not be limited by the amount of training or experience people have. We’re already seeing that in areas like coding.
Secondly, it’s the extra stuff that AI is going to allow us to do. People worry all the time about AI taking jobs. But as far as I can tell, especially in the service industry, there’s almost no one who hasn’t got other things to do if they could just be relieved of some monotonous task or can get it done a lot quicker. So I suspect that’s where we’re going to end up. People are going to be able to do more, they’re going to move to where the AI isn’t and we will evolve our work much the same way as we did when we first got a computer.
It’s very tricky. I must admit, I’m a long-standing regulatory economist and I’ve never seen such a negative reaction and call to regulate a technology than I have for AI. And for the most part, it’s purely on the basis of speculation. We haven’t had AI do significant damage to something yet. The danger is that you might regulate something and squash an opportunity to apply AI that would have been quite innocuous, and you slow down progress.
A reasonable approach to regulation for now is trying to improve some of the practices by which we develop AI models and understand what they are. I think companies are doing that anyway. If you put out an AI-based product, you are still subject to product liability laws so you can’t just throw it out there without thinking about what damage it could do, particularly when it comes to medical applications and anything involving human safety and human interactions that might be harmful.
The opportunities for AI in healthcare are enormous and there is huge potential to collect data, train models and do randomised control trials. But the challenge is it’s potentially held up by all manner of laws that limit our ability to do these things. We’ve got existing regulations in health that are very damaging to every single stage in the innovation process. With AI, there’s also the potential problem that even when you finally get FDA approval, you are back in the regulatory process every time you do a software update.
I can see a huge opportunity if there were one country that cracked open the book and said: “We can grab this entire industry if we just get rid of these barriers.” Because many of the barriers are for the most part harmful.
In terms of using AI within blockchains, blockchains use lots of data and some of the security problems there come from pushing forward predictions that have been hijacked in some way. So if you marry the two technologies together, you could have more reliable predictions in blockchain and you could trust a lot more. That said, this hasn’t turned out to be such a big problem yet, so we haven’t quite seen the benefit of AI in blockchain yet.
When it comes to whether blockchain can help AI – well you know, I’ve been fascinated by blockchain as a technology for a long time and I think there remain lots of interesting aspects to it. But after a period of time, you have to sit back and ask ‘are the transformative applications really there?’ ‘Is it really able to do something that you can’t do elsewhere?’ And for blockchain, for the most part, the answer is no.
There are lots of constraints on businesses doing AI in-house; most notably at the moment there is a lack of talent you can secure for that purpose. The talent has gone where the scale is and the scale is in these companies providing the more general purpose solutions. The good news is those seem to be progressing quite rapidly and we’re all learning a lot as we go.
On the second part of your question, every other time anyone has had that sort of vision, they’ve really had to do it themselves. They’ve had to control the artchitecture. If the architecture is running off an AI tool, this has to be superior, which means the chances are you will have to do this yourself. That said, Uber rearchitected its industry, and it didn’t have to control the phones or even the navigation software to do so. But I feel such cases are rare beasts.
Ultimately, time will tell. Certainly, for most businesses, off the shelf is really the only option. But there is so much opportunity to use off-the-shelf AI, that it is not too bad.
I think that it’s got the makings of a truly transformational technology. What I like is just how widespread and diverse the potential applications are. It is very wide and the opportunities are all over the place.
Is there some limit technically that we might hit before we can do all these marvellous things? Of course, that’s something we don’t know. There have been plenty of other technologies like blockchain, drones and 3D printing that just didn’t technically perform well enough to be transformational. But we are in a great upswing with AI and I’m pretty optimistic about it.
