Transcript
Transcript: Trends in Technology-Driven Change: Barriers and Opportunities Related to 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 "Barriers and Opportunities Related to A.I.".]
So, what are the barriers to implementing? Accuracy, we talked about. That excludes bias. There's bias in these models that gets amplified, and that's the fact that there's bias in the training data, that a lot of training data masks out certain racial profiles, demographic profiles, economic profiles, age profiles. And so, because the foundation models were trained on the internet, it's as biased as the internet is, to that extent. And so, we think, well, let's fix that by putting a human in the loop. So, a human is checking that output and changing it and so on, but humans are biased too. And so, it's this delicate balance of accurate answers, filtering out the bias of that, and challenging them and training them. There are some techniques that are being used to do that, to challenge the output of these systems. So, a technique called constitutional A.I., what it does is you basically have two systems talking to each other. So, one generates an output. The other one checks it and then makes it re-generated until it conforms to a policy or a regulation or an ethos, right? And so, those are some of the things that's happening. Data, we've talked about already. Data is generally a mess everywhere. It's going to take investment to get it there, to get it into a place where it can play more effectively.
The market, so these are the vendors. Everybody's coming at you with an A.I. message right now, everybody, whether they have it or not, right? And the truth is that very few vendors have been able to solve these problems that are on the slide here, and the market is chaotic. The pricing is all over the place. The other thing is that it's coming with some of the large vendors that you may have been working with already, like SAP or Workday or Salesforce, like the big application vendors are coming with this capability baked in. And so, then the question becomes, well, what's left for me to build on my own? Should I just wait for the market? So, that's causing a hesitancy there. And also, it's moving fast for them too. So, one of the things that I do, I have relationships with all the major vendors, like Microsoft and IBM, at their executive team level, and my perception of that is they're about two weeks ahead of the rest of the market in terms of them figuring stuff out and implementing it, understanding how it works and scaling it, and so on. Everybody is kind of figuring this out at the same time.
And then, there's security and risk. Security and risk, some of it is just there are different techniques that you need to use to secure the A.I. environment. There are techniques, bad techniques like model poisoning, where a bad actor can inject bad data that then gets amplified in the model itself. They can reverse engineer it and figure out how you did it. It's just a different kind of attack surface. The risk has to do a lot with IP leakage but also derivative use of fair use principles and copyright material because there's a lot of that in the foundation material, copyrighted material, and there's a lot of pressure to get that removed. It's not a trivial task. And so, where in Europe, a company has the right to be forgotten in Google, that's not so trivial here because of the way that it's computed, and it costs hundreds of millions of dollars to compute these models, which is another thing I haven't put on here, but is really, really important, is that the environmental damage created by building these foundation models is really big.
I mean, it's extremely harmful to the environment, the amount of compute that's required to build these and run these things, the amount of power. If every… and so, the projection is that if everyone that says they're going to build one of these models does, we will run out of power by 2030 worldwide. We'll be out of power. In fact, the Penn State University system was building a giant model for health care purposes, and as they got closer to releasing it and started testing it, they realized that it was causing the Philadelphia power grid to drop, sag. They could see the power drain on the grid. If they actually turned it on, which they ultimately weren't allowed to, it would have taken the grid down, right? So, that has caused a set of innovations towards better compute, better silicon, alternative energy techniques, smaller models, and so on.
[00:04:30 A slide is shown that reads:
"Key Trends Affecting This Market
-Models Will Slim Down
-Mainstreaming of OSS GenAI Models
-Growth in Domain-Specific Models
-Model Hubs Enable Developer Collaboration
-Emergence of Multi-Modal Models
-Regulations Intensify
-Potential Model Commoditization
-Emergence of Autonomous Agents".]
Okay, just one thing to point out here, that what that's leading to is smaller models that will need to be integrated with one another that are becoming more specific to industries or specific domains or specific types of questions, and so getting more accurate as a result.
Okay, I'm going to take you through a couple of examples and then we'll switch into questions. Most of the way that this works now is that you use a prompt to go against your own data. So, you have… so, I was working with Food Inspection Agency, and it's a policy machine, tons of policies about what can come in and out across the border and that kind of thing. And so, their policy data would be some of my data. And then, you run a prompt against that, it queries that, and then it brings it back and creates a human-like response, and it's using a foundation model to do that but the data is coming from CFIA. All right, what's happening though is that you're going to start to use multiple foundation models on the bottom. So, an increasing number of these models are on the market. Some are big, some are small. Some are better for some things than others. And then, instead of some of your data, you want to use all of your data. And then, there are these domain models starting to appear.
So, in this case, it's a health care example. So, you have molecular models that pharma companies are using to create new drugs as well as medical models. There was one released, I saw a note about it today. I think it's a Google model, but it's specific with medical language which is relatively arcane and isn't in the original. And then, you get systems coming like Epic which has these capabilities built into it. And so, you have all of these models, and then some of your data you want to actually create a model of on your own. This is where we're going really fast, which means that you're going to have to integrate across all of them. So, it becomes more of an integration question. What I want you to be thinking about is what's important enough to be that red data, or what are the models that are starting to emerge that you think could be useful for where you sit in the government? Okay, but it goes beyond text. As I said, I'm going to just give you a couple of examples here, things like 3D printed parts or factory layouts, even office layouts. You can use it for power grid design, which we talked about power grids. And so, think about the stuff that isn't text in your environments that could be used. I'll give you an example.
[00:06:36 A slide is shown that reads
"Bionic Partition! Airbus Uses GenAI to Improve the Prototype Design Process
-Goal: Airbus wanted to design components for future plans that were much lighter, resulting in using less fuel and a smaller carbon footprint
-Solution: The implementation team digitally mapped the thousands of options created in the generative design process against weight, stress, and strength parameters to decide which to prototype. The team decided on using this technology to design the partition that separated the passenger component from the galley in the Airbus A320
-Results: The redesigned prototype of the partition reduces the weight of the plane. It is still strong enough to anchor two jump seats for flight attendants, with an opening for wide items to move in and out of the cabin. Assuming that for each 1 kg (2.2 lb) reduction in weight, jet fuel use is cut by 106 kg (233.2 lb) a year, helping to decrease the carbon footprint of air travel. Each partition is approximately 30 kg (66 lb) lighter than the standard partition. For a typical A320 aircraft this results in:
-166 metric tons of CO2 emissions cut per year, per plane
-3,180 kg of fuel saved per partition, per year
-95% less raw material
-Scaling potential is huge as thousands of A320s on order, and airlines have the potential to reduce CO2 emissions by hundreds of thousands of metric tons per year".]
Airbus is a good example. I was on an A320 coming here yesterday, in fact. So, what they've been doing is they wanted to see if they could apply it to the design of the parts of the inside of an A320, so the kit. And specifically, they were focused on the wall that separates the passenger cabin from the galley in the front. And of course, in airplane design, everything needs to be strong and light. The stronger it is, the lighter it is, the better it is. And so, what they did is they actually used an optimization mechanism using generative A.I. They took all of their schematics and fed it into this model, and they were able to re-combine, kind of like the Lego thing a little bit, right? And then, as a result, they were able to reduce the weight by about, what does it say, 30 kilograms of this one wall, 30 kilograms reduction, and that saves them 166 metric tons of CO2, 3200k of fuel saved per partition per year. You scale that out across the fleet, it's enormous.
But what I find really interesting about this, aside from the results, is what the other stuff they used was. So, they not only used their own schematics. They used models of how slime mold grows, which is the single-celled organisms that's extremely effective at connecting points together with strength, and it's extremely optimized in that way, but they also mashed into that mammal bone schematics. So, mammal bones are really strong where they need to be but light otherwise. So, their schematics, the slime mold, and the mammal bones led them to a completely new design that they could feed to a 3D printer and then print this thing out that's 30 kilograms lighter than the original. So, generative A.I. gives you the ability to take these things which are sort of marginally related to one another and actually merge them, and give you what they call their bionic partition. That's what that is. You can use it to create office layouts. And so, these are impossible to read. I don't intend for you to read them, but you can go for different things. You want interconnectivity of workers, here's how you should re-design the space, or you need people heads down, like low-visual distraction, here's how you would do it in the same space. And so, it's using your blueprints plus your parameters. What is the space to be used for?
[00:08:49 A slide is shown that reads:
"Game Changing Generative A.I. Use Cases: Reducing Operational Risk, Event Prediction
-Asset Management
-Cyber-physical Protection
-Environmental Sensing
-Mixed Reality + Composite A.I.
-Marine Telematics
-Employee Health/Safety
-Arrival Prediction".]
Okay, imagine it at a port like the Port of Rotterdam here. And so, this would be the ability to get data from everywhere, so from the ships themselves, from the port, from the cranes which are pumping out data. Interestingly, in the U.S., the goal is to remove the Chinese-made cranes because they're communicating back to the homeland apparently. So, they're replacing them with American-made cranes or non-Chinese-made cranes. It is an election year after all. But then, you have the workers that are on the ground using mixed reality and data coming from all of these systems that are talking to them. Essentially, what you've done here is built a digital twin of the port, but think of all of the players involved, and if you're a government agency, what pieces of data do you have that sit there that could actually be used to make that more effective?
[00:09:31 A slide is shown with an image of Swedbank Arena.]
Just another fun example. So, this is Swedbank Arena which has since been changed to Friends Arena. This is in Stockholm where I was recently. This is a rendering of that, it's a relatively new stadium, but what was done when they were building this is they partnered with game designers, game developers to synthesize an evacuation of the stadium. So, you're able to give these characters true human behaviours, like avoiding objects, interacting with one another, and so on. And so, they're able to actually simulate the way that they would evacuate a real stadium. Apparently, they also do this to simulate halftime, getting to the restrooms and the food stalls (laughs), but what I've learned as I've shown this example to others is in big factories where they have shift changes, this is something they actually need to model because of the complexity of the shift change of people coming in and out, but it's an interesting combination of companies. So, you have the construction company that's doing the design, working with game developers that are creating these synthetic humans. And in this case here, there's also an insurance company involved because if you can spot the risk, where that red circle is, you mitigate it before you ever build the thing. So, you're mitigating it there, and it turns out that game characters are a form of generative A.I. They're generated, they're given certain parameters to interact with. They can interact with one another in an environment.
One of the earliest examples of this that I remember is in Lord of the Rings. So, the CGI for Lord of the Rings used a very similar kind of things, and you see it in the battle sequences where they rush in towards one another and clash. The instructions given were, run into the empty space full force. That was the instruction given to the character. And if you watch really carefully, you'll see some run off the field because their original orientation wasn't towards the battle, it was in some other direction. So, you see them, they just take off (laughs), right? So, this is far more sophisticated than that. It turns out it's more accurate than fluid dynamics as well, because fluid dynamics is far too rational. It behaves exactly like you should. Humans do not, right? Which is why on the A320, they said the exit may be behind you, because the mind says it's the one I came through, right? So, this was able to test those kinds of things. It turns out you can train these avatars with behavioural economics as well. So, you could watch how they behave in transactional situations, not just physical ones. There's a whole lot of power there.
So, that's it, kind of a run-through, where we are, where we're going, and kind of where some of the more progressive organizations are going as they head down this path towards full on A.I. So, summary, spending is intense. Over the next few years, most enterprise CXOs or leaders intend to implement it. You'll implement a mix of everyday use cases and some game-changing ones. The immediate future is multi-modal, so images, text, other types of things, and multi-model which has multiple models working together, and it's causing acceleration in related fields that what it's done, another conversation for us at some point, is it's brought quantum computing probably six years closer to us than it was three years ago, which is… you think this was disruptive (laughs), that will be even more so.
Thanks for watching. And again, I hope you found this useful and interesting for the work that you're doing in Canada.
[00:12:47 The CSPS logo appears onscreen.]
[00:12:54 The Government of Canada logo appears onscreen.]
