Transcript
Transcript: Trends in Technology-Driven Change: The Evolution of AI
[00:00:00 Text appears onscreen that reads "Trends In Technology-Driven Change".]
[00:00:06 The screen fades to Chris Howard.]
Chris Howard: Hi, I'm Chris Howard. I'm the Global Chief of Research at Gartner. Thanks for taking some time to listen to the advice that we have around A.I. and related subjects. I hope you find it interesting.
[00:00:17 Text appears onscreen that reads "The Evolution of A.I.".]
This is an astonishing number. Our forecasts show that between now and '27, three trillion U.S. dollars will be spent on A.I. It's an enormous amount of money, and not just on generative A.I. to be clear. This is all of A.I., that I'm going to talk about here in a minute, but this level of investment my forecast team has never seen. The closest things were the adoption of the iPad interestingly, so not just mobile but the iPad specifically. It had a similar kind of flow, like senior executives were using iPads for their own productivity. The same is true with A.I. And so, there's that motion that's happening there. So, it was like… but like the internet, but the difference with the introduction of the browser is it took 30 years for this to play out. This is three years. So, it's an order of magnitude faster.
And again, a lot of that money will be wasted, and we're sort of a third of the way through this story right now, but that's just the nature of economics, right? When something new comes along that requires investment, you burn a lot of fuel in that beginning phase. Think of Carlota Perez's S-curves. You build… you spend a lot until it finally gains momentum of its own and then you don't have to spend as much fuel. And so, we're in that first third of that right now, and the complexity of this is that it keeps restarting. You hope that simply something would trigger and then it simply finds that spot, but it keeps re-triggering because innovation is happening so quickly. Every week, there's something new. It's really, really hard to keep track of. And so, it's, again, unlike anything we've seen.
All right. So, I told you that the A.I. story has evolved. Last year, the emphasis a lot was on ChatGPT. ChatGPT is essentially a vendor product now. It's a vendor product from OpenAI. Microsoft has a funding relationship with them. And so, as that started to happen, people's understanding of the space started to grow. So, ChatGPT is a class of generative A.I. GPT stands for… of course I say it and I'm going to forget it now (laughs), generative pre-formed transformer. So, basically what it does, all a GPT is is a predictor of something. It predicts the order of words, and it's trained on so much data that it's really, really, really good at it. And so, if you can give it a sentence with a missing word, it has a pretty good probability of understanding what that word is, and it's right so much of the time that it feels really human. But when it's wrong, it's really wrong (laughs). When we first started experimenting back in November of '22, my analysts, being curious people, said, well, write an obituary for me in the style of Wikipedia just to see how much they would get. No, not an obituary, excuse me, a bio.
(Laughter)
I just gave away the punchline, because what happened is it would write a pretty realistic bio of them with pretty good detail and a few mistakes. But in every case, they would also write an obituary for them, always. Why? Because they asked it to do it in the style of Wikipedia and most Wikipedia bio entries have an obituary. And so, it thought, well, this is a missing thing, I'm just going to make it up. I'm just going to put it in there because that's the next chunk that would come, I predict the chunks. And so, it's that prediction machine that is really the magic behind the thing, but it can predict other things. It can predict the order of Lego blocks and how they fit together, which is what I showed you in the beginning. It can predict how pieces of a policy fit together.
So, one of the pieces of work that I do is with the big insurance company AXA in Europe, and they are doing what's called computable contract studies. They're taking all of their insurance contract language and converting it into code, and so that they can re-combine that for highly personalized policies for which I think there's application in government as well in terms of policy language, but one of the interesting things that happened there is when they converted it, they found mistakes in the language. They were able to test the code in different ways and actually find holes in logic or nested things within the policy language that then they fixed. And so, they had this sort of interesting thing, but it was able to predict the re-organization and combinations of policy within contract language. Some things are allowed to be together and some things aren't, and it understands that. So, all of the magic that we think of here that makes it appear human is the fact that it predicts the order of stuff. That's accomplished through LLMs which are large language models and prompt engineering which is the way you ask the questions. So, people are thinking more about that.
[00:04:40 A slide is shown with a diagram that reads:
"Decide
-Optimize, Forecast"
"Discover
-Anomalies, Identity (e.g. faces)"
"Take Action
-Automate, Modify" in dark blue
and "Generate
-Text (+ other media)
-Code" in red.]
What's happening now is that generative A.I. is settling into the overall stack of A.I. The message here is this, that these A.I. functions that are the dark blue capabilities, and it's oversimplified I realize, have been around for 50 years in some form or another and have been maturing over the last three decades. Natural language processing is a very, very mature discipline, image recognition as well.
I think back to the time I was working on Stanford campus around 2003, '04, so 20 years ago. And at that time, we had all the math. What we didn't have was the compute. So, to actually ingest the trillions of parameters of data and all of the images of cats that you could possibly find took a lot more compute power than we had. And so, it took evolution within the computing space, things called graphic processing units or GPUs and other things that are coming that actually created the computing environment where you could make these things real, and that's a big part of why this feels new to a lot of people, because it's that that made it useful. It sort of brought it to the public view.
But you're building on top of things that have a long line of maturity that are part of them and probably are in use across multiple agencies and departments already, things like fraud detection, anomaly detection, those types of things. We have masses of data where we're using machine-learning to understand the patterns in them. That's not a new technique. What's new is the ability to ask it a question. So, in the case of the AXA policy stuff, not only is it creating new personalized policy but a policyholder can take a picture of something and say, this weird thing just happened, am I covered? And then, the prompt actually gives them the ability to talk to their policy in a natural kind of human way.
The mistake we sometimes make as we're building policies and services is that we expect the people that we're offering those to to thoroughly understand them as if they work with them every day like we do. The fact is, for like an insurance claim, they maybe do it once or twice, life insurance once, okay? And so, as we're designing better policies and processes and services for these people, you need to think about, what's the natural way that they want to communicate with this stuff, and this becomes a way to do it. So, sitting on top of capabilities is a more natural human way into it.
Thanks for watching. And again, I hope you found this useful and interesting for the work that you're doing in Canada.
[00:07:04 The CSPS logo appears onscreen.]
[00:07:10 The Government of Canada logo appears onscreen.]
