- hace 6 semanas
Aprende a transformar la incertidumbre en una ventaja competitiva. Esta sesión mostrará cómo la IA ayuda a líderes mexicanos a tomar decisiones más estratégicas y efectivas.
Conferencista:
Sam Ransbotham - Editor de AI Initiative de MIT Sloan Management Review y profesor de Analítica en Boston College
Conferencista:
Sam Ransbotham - Editor de AI Initiative de MIT Sloan Management Review y profesor de Analítica en Boston College
Categoría
📚
AprendizajeTranscripción
00:00Hi, thanks for joining this session. My name is Sam Ransbotham. I'm professor of analytics,
00:14and I'm the editor for the AI and Business Strategy Initiative at MIT Sloan Management
00:20Review. We've been studying analytics and data since, gosh, since 2014 or even earlier,
00:27and AI in particular since 2017. And what I'm going to talk about today is some of the recent
00:34findings we've had from our research and a couple of hints about the research that's coming out this
00:40year. And what I want us to think mostly about is thinking about artificial intelligence beyond
00:46productivity. Let's start with who's actually using artificial intelligence now and how. One
00:54of my favorite stories comes from a discussion that we had with Vandy Verma at NASA. There is one
01:01planet on our in our solar system that has purely automated driving, and that's not Earth, but it's
01:07Mars. Mars is interesting because it's so far away that any sort of telecommunication between Earth and
01:15Mars takes about half an hour round trip. And what that means is if you're driving a vehicle around
01:21Mars, the people at on Earth in JPL and in NASA can't actually steer the car, meaning that they
01:30can't just in real time steer. So instead, any of the driving, any of the helicopter flying has got to be
01:37automated. And what's fascinating about that is that they have a very difficult set of objectives. On the one
01:45hand, they want to use their machines to explore as much as possible of the Mars surface. And what
01:53that then means is driving fast. You drive fast, you can cover more surface. At the same time, they want
02:01to study in depth. They want to look deeply at certain things, which takes time and requires going slow. So
02:09you can see those objectives fundamentally compete with each other. So what they've done is they've used
02:14artificial intelligence to help them drive the car and the instructions. And I'm simplifying greatly,
02:19but the instructions are drive fast until you see something interesting. And when you see something
02:26interesting, slow down. And that's something interesting is really hard to quantify, but it's
02:33something that the Mars Rover is doing. It's sort of looking at, hey, these are the kinds of things I've
02:39seen before. So those are not interesting, but this is something unusual. So I might want to slow
02:44down. And this is sort of anomaly detection and real-time processing of data that is allowing
02:51them to actually just last week, make some really interesting discoveries of rocks. So there's one
02:56planet where we're doing autonomous driving, and that is Mars. You may say, well, at least they don't
03:02have other traffic and other cars to deal with. So perhaps they have an easier problem.
03:05But there are other organizations, not exploring the universe. I talked with Daniela Patecki. He's at
03:15Pirelli, and they make tires. They're a company that makes tires based in Milan. And you'd think, well, how
03:23could tires have anything to do with artificial intelligence? But they do. Because historically, Pirelli
03:29had to create new tires, they would go in and mix something like 200 different compounds and different
03:35ratios and different recipes to create tires. Now with artificial intelligence, they can create those
03:42tires virtually. They can simulate virtually how those tires will perform. That's not, of course,
03:49all they do. But when they do that, they don't have to create and test physical tires. They can take a
03:59small subset of the possible tires and only test them. And what they found is a lot of interesting
04:05combinations that they would not have expected and testing results. Many of us are likely familiar
04:13with Slack. Slack is a workplace productivity tool. And if you've used Slack, you can note that the
04:21number of messages can be overwhelming. It's just hard to keep up. It can become a full-time job just
04:26to keep up with all the messaging coming in. What Slack has done is instead of using Generative AI to
04:32generate content, they've used these tools to summarize content so that if you have missed maybe an hour or
04:40two's worth of messages, you can catch up quickly. You can get just the latest information. And I think
04:47that's a fascinating use of technology to help us save time and be more productive. Shopify is a company
04:56that runs a lot of the infrastructure of the internet, particularly retail. And if you wanted to
05:02list products on a site like Amazon or other places, you might fill out some information about your
05:09product. You might start with a picture and then add a description of it so that buyers would know
05:13what they're getting. Within three months of artificial intelligence coming out with a large language
05:19model improvements that like ChatGPT 3.5, Shopify had it built into their product so that people could,
05:28from a picture, get a quick set of four or five descriptions. And those four or five descriptions,
05:35they could then pick one and tailor it. This wasn't amazing. This wasn't driving vehicles
05:42around Mars. Instead, it was just saving individual people three or four, five minutes every day.
05:49These things add up when you do lots and lots of products.
05:54I like the story of Duolingo. Duolingo is, gosh, I think they told me that there were 600 million people
06:01learning English every year. And that's probably changed and increased since I talked to them.
06:06And what's fascinating about that is that they have an ability to observe humans learning
06:12at a scale that's not been seen before. Historically, an English teacher might see a dozen students
06:20every classroom, two dozen, three dozen, maybe over a career, a thousand or two students.
06:25But with Duolingo able to observe 600 million people learning, I'm really optimistic that all
06:31that data can help improve the way we as humans learn. What interests me about that is what's
06:39interesting there is that it's not the machines learning when we think about machine learning,
06:44it's humans learning and how the machines can help us learn.
06:47I talked with Ronald Nelson at Heineken Brewery, and he described their company as a 160-year-old
06:56startup company. And I thought that was an interesting take because as they've seen the
07:01opportunities come with artificial intelligence, they're recognizing that many of the ways that
07:07they've done things in the past maybe need a fresh new look. And with that, he cast the company more
07:14as a startup than an established long-term company. I think a lot of people are feeling that.
07:22At Expedia, they're using these tools at scale. They have 600 billion decisions a year made with
07:30artificial intelligence and machine learning. These are not huge decisions. These are relatively small
07:36transactions, but happening at massive scale.
07:39Now, this was fascinating. I had no idea what the World Wildlife Fund would do with artificial
07:45intelligence. But I talked with Dave Tao, who's a fascinating person in general.
07:51He described all the many, many ways that they're using these tools. For example, they have cameras in
08:00the woods looking for endangered species. And so rather than having a human monitor those, they can have
08:05image recognition help with that task. But more than that, they're monitoring ports. And so when a boat goes out
08:14to sea, and it sits in the water at a certain level, and it comes back and is sinking lower, they can get an
08:23estimate from that difference in the ship in the water, how much weight of fishing they've done. And so they can get a
08:32a way to check for overfishing at scale. Or they can use heat, heat image and image recognition on crates at ports to try to look for
08:43smuggling of animals. They're looking at social media to see what's the hot new exotic pet. And so these are
08:53fascinating ways of using these technologies that are not just your classic company using AI.
09:02And finally, another great discussion about innovation and exploratory innovation is Moderna.
09:09When the COVID virus was first identified, back in 2020, within three days, Moderna's AI models had
09:18identified a bunch of potential vaccines that could address it. Not all those worked out, of course. But the
09:25one that we ended up using was one of the ones that AI identified. And that was using AI to explore a search
09:33space. And so the point of all this is that we may think about, you know, exotic things like maybe the
09:39the Mars Rover or humanoid robots from from movies. There are many, many companies using artificial intelligence in
09:47many varied and, you know, from from somewhat routine to somewhat exotic ways. And I think that's fascinating.
09:57You might think that I've picked nine stories, but I have not. Well, I did pick nine stories, but
10:02they aren't the only stories. We've been tracking organizations use of artificial intelligence since 2017.
10:10And you see on the x axis here, the the graph going from 2017 to 2024. And on the y axis, we've got the
10:19number of organizations, the percentage of organizations that we talk to that are either
10:23piloting or deploying artificial intelligence of some sort. We had relatively stable, some growth
10:31through sort of 2023. But then suddenly in 2024, we saw a huge spike.
10:37In talking to people, the reason for that spike is right about the time that the large language models
10:44happened. And so suddenly there were a lot more useful. But then I think there's more to that,
10:50because there's also increased awareness of what the potential for artificial intelligence does.
10:56Back before 2023, when I talked to some organization, when I talked to executives about
11:01artificial intelligence, they might not have ever done anything themselves. They might not have built
11:07models themselves. And so when we talked about artificial intelligence, it was somewhat abstract.
11:12When I talked about machine learning, it was abstract. Suddenly between 2023 and 24,
11:19people became aware because people could go to websites and use these tools themselves. And they
11:23became aware of both what it can do and what it can't do. And we've got numbers from 2025 coming out,
11:31on November 15, in a new report. So stay tuned for that.
11:39A point I want to make here though, with all these technologies, is that tools don't decide how they
11:44get used. People decide. And how we as humans use these tools matters. I have an example, and I'll start
11:53with a, you know, if I think about the very first caveman that picked up a rock. One of the options that a
11:59caveman could do was hit another caveman with that rock and use that tool as a weapon.
12:06Another option would be to take that same rock and use it to build things, to build a house, to build a
12:11fire, to build a, you know, whatever they're building, he's building out of rocks.
12:14The rock doesn't know any different. And the rock is a tool. So how we use these things,
12:22like artificial intelligence, how we use these technologies, makes a big difference. And we,
12:27as humans, get to choose how to use these things.
12:29And if you're paying attention, you can tell that I've, I've generated these images using AI,
12:36of course, because I'm not a good artist. But you can see, we've got three legs here on this one
12:41caveman. So I think it's amazing how well it does some things, and then how it makes this sort of
12:48crazy three-legged mistake. Anyway, there's a whole topic there about mistakes in AI. But let's come back
12:56to this caveman. So when you think about this caveman, and he's discovered fire, he discovered
13:03how to use and harness fire, I think he's probably pretty excited. He wasn't as cold anymore. He could
13:08use the fire to cook things, to cook whatever's on that stick. That looks like a marshmallow to me on
13:16that stick. But I'm hoping, I'm doubting that people are toasting marshmallows as first caveman.
13:22But my point is that I don't think that when the person in first saw fire and began to use fire,
13:29I don't think they had any idea of where the ability to harness fire would lead us.
13:36Pretty quickly, we started to use, as humans, we started to use fire as candles for light.
13:43And, you know, when you're thinking about heating food, maybe you're not thinking about it. But when
13:47you're thinking about what happens with the light, light let people stay up later. It let people go
13:53inside of shelters like caves. It let them draw on the walls of caves and preserve knowledge.
14:00So again, when someone is cooking and just getting warm, that's one thing. I don't think that they
14:05realized that that would lead to written language and written knowledge and preserving knowledge from
14:10one generation to the next. But it was part of it. And that one caveman probably never traveled
14:18more than three or four miles from the place that they were born. But if we put a fire inside of an
14:23engine, we can use it to travel. We can use it to go further than this caveman ever thought about
14:30traveling. And again, I don't think that that first caveman was thinking about that.
14:34And that caveman looked up at the moon. I don't think the caveman was thinking, hey, if I put this
14:41fire in the end of a tin can, that I could blast myself or someone up into space.
14:49Again, it's these are so far removed from our initial first uses of a technology. And I think that's
14:56that's a point here is that we have agency in how we use these technologies. But it's difficult to know
15:01where those technologies will lead. At the same time, that's what people are pretty excited about.
15:09If you think about what people are excited about is the potential for economic growth from AI. So I
15:15found this chart. It's from a Project Madison database. And on the bottom axis is the years from
15:21the year 1000 to the year 2000, 1000 years of human history. And on the y axis is a logarithmic scale.
15:30So it increases by the factor of 10 every time it goes up. Of how much the world produced. I'm going
15:38to be a little vague about that, but just how much in some sort of scale of how much the world was able
15:44to produce. And over time, from the year 1000 to the year 1500, the abilities of humans to produce
15:52only increased about 40% every century. We saw things like the printing press come along in the
16:00around this phase. And suddenly we could write things down and productivity was able to double every
16:08hundred years. But then things started happening. We got steam engines. Steam engines are responsible for a
16:18great deal of mechanical work. And so when people had steam engines and suddenly we could start to
16:24double, you know, we're not only doubling, we're quadrupling or tripling, I guess, how much we can
16:30produce every century. And then I've got some other technologies listed along here and we can see
16:40just what people are excited about is how much our overall productivity and how much we're able to
16:46produce has increased over time. You know, point out transistors here that I think that's a really
16:50big one that we're experiencing very recently. Now, I think we have to be careful here just because
16:57every, you know, everything on human, you know, human development is on this timeline. That doesn't
17:01mean it caused that growth in productivity. For example, I mean, there were world wars in here.
17:07They didn't, maybe they did, maybe they didn't, but not every invention is a, is one that we want.
17:13But the real question is, you know, and actually, I don't know, just as a joke, I put when I was born
17:18in here, I like to think that I increased worldwide productivity by 20 times, but I don't think that
17:24I did. But people are excited about artificial intelligence and how that could potentially help
17:29that curve keep going up. It's up to us to decide how it does though.
17:35What I want to focus on for the rest of the talk is the idea of using artificial intelligence for more
17:44than just efficiency and for more than just revenue. And one of our recent studies looked
17:49at using artificial intelligence to help manage uncertainty. We have a lot of sources of uncertainty
17:56in our world. I don't have to tell everybody here that the, the title of this, of this whole conference
18:03speaks to the, some of the geopolitical tensions that are, that exist in the world now. But we have
18:11more than that. We have fast moving customer preferences. We've got talent disruption, people
18:16leaving organizations, moving to organizations. We've got regulation changing all the time. We've got
18:21technology. So there's a lot of uncertainty in our world. And the less that an organization knows,
18:29the greater its uncertainty and the less it's able to manage its resources strategically.
18:35And what I'm going to do is try to, in our research, what we've done is
18:38connect managing uncertainty with learning and with artificial intelligence.
18:46So as a case study, let's think about Estee Lauder. So Estee Lauder is a
18:50a consumer goods company that produces beauty products. And what they found was they had a
18:58strategic need to try to anticipate consumer trends before their competitors. So the degree that they
19:06could get ahead of consumer trends put them in a better position. This became important because
19:12historically consumer preferences might have shifted once a season. But now they're seeing shifts happen
19:19really quickly. Social media, digital influencers, shifts are happening fast. And so what Estee Lauder
19:26does is they're using AI to detect these changes and try to have a market response ready to redeploy
19:33inventory, to redeploy supply chain resources when they need to. Now, I talked with Somya Gotapati, who's a
19:41very fun and fascinating person. She's the Vice President of Global Supply Chain Technology and
19:48describes how they're using artificial intelligence to discover these consumer trends. And then the
19:53key is matching up their existing products to those trends so that we can repackage them and position them
19:59for market for that trend. So they may not have time to produce something new, but they may be able to
20:04have get something they already have into the right position at the right time in the right place. I think
20:11that's fun. As part of our research, we looked at how companies were learning and how they were using
20:19artificial intelligence. And because I'm a business school professor, we tend to think of things in
20:24two by twos. On one axis, we have organizations that are involved in organizational learning. They feel like
20:31that they have capabilities in organizational learning from low to high. And then on the other axis, we
20:39looked at how they used artificial intelligence to support those. So for example, in this bottom corner,
20:45we found that the vast majority of organizations, almost 60% of organizations, felt like that they were
20:51low on organizational learning and low on using artificial intelligence to learn. Now, some organizations
20:58were very focused on AI learning, but not particularly strong in general organizational learning. And that's
21:03about 12% of those. Other organizations were high on organizational learning, but hadn't quite adopted
21:09artificial intelligence to help with that. And we found about 15% of the people who what we called
21:15augmented learners, people who were learning and were also using artificial intelligence as part of that
21:20organizational learning process.
21:22We tried to tie a link between their learning capabilities and their ability to manage uncertainty.
21:31And so at a very high level, we see a strong influence of learning and artificial intelligence
21:38supported learning on managing uncertainty. And I think, you know, obviously, it's really interesting that
21:44artificial intelligence plus organizational learning is the highest. 82% of the people agreed that that helped
21:50them. That's most people manage uncertainty. I thought it was also interesting that the sort of historical
21:57non AI based organizational learning was really faltering compared to AI specific learning. I think
22:04there's a lot more to explore there. We broke some of this down. But our key message is that
22:11organizations that are using AI to learn can manage uncertainty better. But how?
22:16We thought about three different types of uncertainty, we explored those deeper,
22:25talent, technology and legal disruptions. So a lot of the world is dealing with a lot of
22:32uncertainty due to talent. So people leaving companies taking what they know about the organization with
22:38them, people retiring. These are what we've called talent disruptions. We found in general that
22:46augmented learners, people learning with artificial intelligence are much more likely to be able to
22:51handle uncertainty from talent disruptions. And this makes sense. And I'll get into a little bit more
22:56why I think that can happen in just a minute.
23:01There's also a lot of technology disruption. And again, we find that that augmented learners are much
23:07felt like they're much better able to handle these technology interruptions than the
23:12limited learners. And with technology disruptions, we think about the fact that there's a new version
23:19of every tool coming out every week. And finally, we also thought about legal disruptions. There's a
23:25lot changing in the legal landscape. And augmented learners were better able to handle the uncertainties
23:31from that, whether it's the effects of changing legislation, changing rules, changing regulations.
23:39These organizations that are both using artificial intelligence and using artificial intelligence to
23:45learn are able to handle those uncertainties.
23:47So compared with limited learning, augmented learners are better handled talent disruptions,
23:552.2 times better talent, technology disruptions, 1.8 times better, and even legal disruptions,
24:001.6 times better.
24:02So how do we draw a link between organizational learning and artificial intelligence?
24:08We think of three things. And the first is knowledge capture. So an example is taking
24:16information that's coming in, there's a fire hose of information in the world, just vast quantities of
24:21information coming in. And people are using artificial intelligence to try to extract those and just
24:27and to capture all that information. Let's go back to my NASA example. When a rover is moving across
24:35Mars, the amount of sensory input that they have on those devices is amazing. And it's really difficult
24:43to capture it all. And so our AI is helping capture that sort of knowledge that's resistant to legacy
24:52techniques. And when I think about legacy techniques, I think about organizations that
24:55may have had a wiki in the past, but no one wants to type in things into the wiki.
25:01With AI monitoring and filling out and capturing that information, organizations are able to capture
25:07more of that tacit knowledge. But capturing alone isn't enough. Second thing we thought about was
25:13knowledge synthesis, taking that information and distilling and finding patterns. So an example here is
25:20Stitch Fix. Stitch Fix is a fashion company in the US that will try to understand both what
25:25you as a person want and what's available in the market and match those together.
25:31And so at the time, what it's doing is organizing data, pulling it together from lots of different
25:37sources and trying to make that understandable enough for people to pick up on these patterns.
25:44So not only a huge amount of information captured, but synthesized and distilled into something that
25:50people can understand. And the third part of that is just getting that information out.
25:57So Slack, which I mentioned earlier, will take a whole bunch of information, will distill it to a
26:02small thing, but then it puts it at the place of the person who is needing it. So they gave an example
26:11of a salesperson who was about to go visit a company and they were able to go into the tool and say,
26:16what's the latest happening with this company? And so that they, when they visited their customer,
26:22were aware of everything happening within their entire organization so that they were more
26:28knowledgeable walking in. So we felt like it takes all three. It takes capture, it takes synthesis,
26:34and it takes an ability to disseminate that information.
26:39All right. So great. You know, if I've convinced you, just go and do some AI stuff.
26:43And it's true that, you know, lots of organizations and boardrooms are more particular and more receptive
26:48to AI initiatives than they've ever been before. But there's several challenges. And, you know,
26:53I don't have to tell you these challenges. You're very aware of them. We think about process redesign.
26:57We think about the pervasive potential and how widespread it can affect. We think about the
27:02competing incentives for organizations and individuals. And we think about, you know,
27:07whether you try to improve existing things or try to explore huge measurement issues and scarce resources.
27:15And I don't have time to go through all those in a short presentation. So I'm just going to focus on
27:19this last one, these scarce resources, because I thought that might be something that people are feeling.
27:28When we think about these scarce resources, resources aren't infinite. And so we looked at how
27:33people who are using AI to learn well, these augmented learners, they're, how are they investing
27:39in projects differently than others? And we found that we ask many different things, but a couple I'll
27:47highlight is, one, these augmented learners are much more likely to focus on long term impact of projects
27:53than short term. And I think that's consistent with the idea of using these tools to learn.
27:58They also were much more open to risk. You know, they're still not, you know,
28:06they're still not investing only in risky projects, but they're more than twice as likely to invest in
28:11risky projects. And I think that's consistent with the idea of thinking not only about the risk of the
28:17project in terms of, it may not be financially beneficial, but augmented learners are also thinking,
28:24what could we learn from that project? And I think that's an important distinction.
28:31Some of these require sort of longer term investments. I spoke with SAP and Walter's son,
28:37actually he's a pretty fascinating person, turned out we had a history in common that I didn't know about
28:43before I talked to him. He's the vice president of the global head of AI there. And one of the things
28:48that they did was they thought about building technologies which are reusable for all the lines of business.
28:54And so this required them to create a platform where at the time they had more than 30 different
29:00large language models available. And the idea was that if they built this infrastructure for people,
29:06then it would help the organization as a whole. So that's different than thinking about one-off
29:14ways of addressing individual projects, but trying to address them holistically. But that did require
29:20an upfront investment and they're starting to see that pay off now.
29:27The big question is where to invest your efforts. So I wrote an article recently with George Westerman
29:33and Chiara Farnato and HBR that was talking about finding the approach that fits the problem you're
29:38trying to solve. And it's true that all these agentic tools are impressive, but I think it's also
29:43important to remember there's lots of other data science techniques out there. And I've got the link up
29:48there if you're curious about reading that article. And the crux of it is that we looked at the potential
29:54for automation, so just sort of normal coding automation, econometric approaches to learning,
30:01machine learning, and then generative AI. And we thought about it from five different dimensions,
30:06how much data each of those required, how much domain expertise they required, how repeatable they are,
30:13how explainable they are, and how rigid they are, how much they adapt to change. And what we did is,
30:19if you're curious, put out some classifications of how each of these approaches handled each one of
30:26these five dimensions. So for example, and I don't have time to go through them all, but if you think
30:31about the domain expertise required to code an automation, it's high. It's great if you have it,
30:38but if you don't have it, then what do you do? Something like a generative AI approach can require
30:44much less domain expertise. On the other hand, a generative AI approach and a machine learning
30:50approach requires a much greater amount of data required, much, much more than the econometric or
30:58automation approaches. So at the same time, automation can be explainable in a way that generative AI or
31:07machine learning approaches are not. As you're thinking about the different possibilities,
31:13there are different choices that every context will decide and dictate how you can approach them.
31:22I hope this was interesting. I think it's a fascinating topic. If you're interested in these,
31:26we've done a series of reports through the years, and these are all available on our MIT Sloan
31:30Management Review website in our big ideas section. And this is the last seven reports that we've
31:39produced every year on artificial intelligence, starting with how things are changing, how it's
31:46becoming more normal. And then more recently, we focused on things like cultural benefits and
31:52individual versus organizational benefits, and how companies are managing uncertainty.
31:58A lot of the stories that I've mentioned today come out of a podcast, and this podcast is called Me,
32:04Myself, and AI. We're in our 12th season now. We just released an episode with Hugging Face
32:11just yesterday. We talk with companies. We talk with normal companies, Walmart, Humana, DHL,
32:19Starbucks, all kinds of companies about exactly how they're using artificial intelligence,
32:24intelligence, and trying to connect how they're using artificial intelligence with the people behind
32:29it. I started this discussion talking about how we as humans have agency, how we as humans have
32:37choice in how we use these tools. What this podcast is, is about the people behind making those choices.
32:44I find the whole development of artificial intelligence fascinating. I hope you do too.
32:48If you have any questions or disagree with anything, I always like disagreement,
32:54feel free to contact me. I've got my information there at the bottom. Thanks for listening.
Recomendada
1:01
Sé la primera persona en añadir un comentario