Why AI isn’t scaling faster? Hear what Nicholas Borsotto of Lenovo has to say. While AI adoption is still quite low, Lenovo is seeing a “Significant growth in enterprise adoption over the past 18 months.”
Nicholas Borsotto:
Hi, everyone. My name is Nicholas Borsotto. I’m the global AI manager for AI Alliances at Lenovo. And I’m very happy to be here to give a talk on why isn’t AI scaling faster and what can you do about it? So a little bit about me. Before I joined Lenovo, I was working as a founder for Archgriffin Consulting, a consulting company focused on data science and machine learning startups. And even after I joined Lenovo, I kept up my work in Meetup.ai, which is one of Europe’s biggest online communities for discussing practical artificial intelligence. I also often give talks on data economics and venture startups here in Berlin on the topics of product market fit and artificial intelligence. So a little bit about Lenovo, the company that I work for. I think many of you are well aware of our computers and laptops, but actually Lenovo is one of the biggest producers of servers and especially AI ready or AI optimized servers as well.
Nicholas Borsotto:
And we bring that reliability on our servers business also to the work that I lead inside of Lenovo, which is the Lenovo AI Innovators, our ecosystem for AI partners, so basically bringing all the people together necessary to deploy practical artificial intelligence in real life use cases. So let’s start talking a little bit more generally about AI. I think most of you here wouldn’t be surprised when we see some of that data, some of the overall trends of models getting more and more complex while we see this huge explosion of data. So I think the question of “The data is just not there, it’s not even possible anymore to be made,” there’s so much data being created every day. We see big improvements in terms of computing and hardware, so new accelerators that can do much faster, orders of magnitude what we thought would be possible before, at the same time that you see hardware really pushing the boundaries in terms of thermals and other requirements to bring the art of the possible to AI.
Nicholas Borsotto:
On the other side, we see edge AI, which before was mostly focused on IoT really becoming a huge thing with a lot of inferencing and even some training happening on traditional edge servers and edge scenarios. And finally that all balanced together to make the adoption of AI enterprise skyrocket over the last few years, especially as the pandemic forces a lot of companies to rethink how they were doing business. But when we pop up the trunk and we look a bit more carefully at it, with the exceptions of a few leading lights like telecom and automotive, you see that adoption of AI currently worldwide is so very, very low across large companies. Most of them don’t hit even the 50%. And this data here is based on interviews with CTOs, which means if anything, there’s a overestimation here as a lot of companies use any data analytics to be considered AI.
Nicholas Borsotto:
So we have all those trends going on. We see that adoption is increasing, but it’s still not getting to the size that we expect. So it brings to the question of why isn’t AI scaling faster and what can we do about it? This talk is going to be mostly focused on startups and ISVs because that’s where I come from and where I have most of my experience. But I wanted to start off with some generalized view of some good ideas and a good framework for developing AI for corporates and startups, which basically starts with a use case, starts with an idea why, why we’re looking at this, why is this worth it? The data, so the raw resources here, you can also see, I think, the algorithms that would be necessary for you to develop this.
Nicholas Borsotto:
Most work in AI nowadays doesn’t come straight from scratch. There’s already a lot out there that can be repurposed. So the search for data nowadays is also the search for what’s already out there and what can I use? Then there’s a question of infrastructure that is very important for us here at Lenovo, which is a question of what hardware, what infrastructure, what software will be used to manage and create the right environment to deploy this. And then finally people. You could also argue that people are the very beginning. But here, we mean after we know all the tests and we understand the challenge deeply, and we know why we’re doing this, then let’s go at it. Let’s get more people to help us, support us into this new journey. So this is the, let’s say, easy, clean breakdown, and we’re going to go many steps deeper and focusing a lot on how does that look from the ISV perspective, the people actually often developing the AI.
Nicholas Borsotto:
So first of all, let’s remind ourselves that a startup is a very, very specific construct. Nowadays everybody’s called a startup, but a startup basically is the tip of the spear. So they only try to achieve a combination of something that brings customer benefit, something that is commercially viable, it’s technically feasible, but very, very importantly something that is scalable. If you have something new that fits all those three options, and it’s not scalable, that’s consultant. And if you have something that fits all of those three things, and it’s not new, then it’s basically a traditional product. So that is the best way of understanding that scale is hardwired inside of what a startup is supposed to be about. That is the reason why people invest so much at companies that haven’t raised or haven’t made any revenue.
Nicholas Borsotto:
Nowadays, we expect to be around 25,000 AI startups all over the world. A few of them, actually 1,000 of them, 2,000 of them actually raised more than $10 million. And you even have a few hundreds nowadays that have raised more than a $100 million being serious potential inside of the market. But today we’re going to speak about those companies who are still scaling their product. So here to define, we’re going to be talking about early scaling compared to international scaling, or basically bringing a product all over the world. And this is a question that I faced when working with startups from the moment that I was a manager inside of one, as a consultant in Lenovo, as a mentor. And they can roughly be brought to three specific groups. You have technical, you have behavior and you have economic challenges.
Nicholas Borsotto:
And overall in startups, those are seen as, okay, let’s just throw business people at it and we’re going to fix it, or let’s just throw engineers at it, we’re going to fix it, sometimes is let’s throw marketing and design at it, and it’s going to fix it. And all three of them come from a natural misunderstanding between the client and the startup plus how they understand the situation that they’re in, the market, the people, the businesses that they’re trying to break into. So let’s go straight into the three challenges. I’m also going to be using a little bit of information from companies that I worked before without quoting names and exemplifying the case over there on the top. So to talk about technical challenges, I think it’s very interesting to keep e-health care in mind. And the origin of this issue comes from the very different situation, which is to develop AI on a research or experimental fashion to a production level. And often you will see people saying, “Ah, because this is good for production, I use this notebook for this.”
Nicholas Borsotto:
And in truth, that is still not production because production naturally means moving from your clean and predictable environment to a messy and chaotic one. So it’s not only about the tools you’re using, but how you move from something that you can control all the parameters to something that you cannot control all the parameters. So how fragile is your solution? So some of the points here to keep in mind, and we’re going to be going through this like this, it’s fault and advice. And in the end, I’m going to roll back a few of the points that we saw. So plan for data diversity on day one. So understand that whatever data you collected to develop your algorithm in the beginning probably is going to bring you very little when you actually hit real life cases. You would need to always be adding real life data, more data to your product, to your development.
Nicholas Borsotto:
But while a lot of people just think that the more the merrier, what you should be actually looking for is data diversity, to understand data coming from different places and different styles. So in e-health, for example, it’s very traditional that companies working in radiology work specific data sets coming from one specific hospital using one specific type of machine. That leads to critical issues when you’re trying to do comparisons between hospitals or between radiology machines from, let’s say, General Electric, GE, or Siemens. There are even cases of people getting misdiagnosed because the assistant radiologist used to hold the thing in a way that his thumbprint was in the image. So it’s very, very important to work for that. And that thumbprint part actually brings to a second part of that question, which is edge cases. Really understand what are the things that almost never happened when you look at it from a statistic point of view, but it still happened often enough that your client will see it at some point?
Nicholas Borsotto:
So a traditional use case would be for somebody who basically in autonomous driving, that somebody crosses the street without being in a proper place. That shouldn’t happen in a perfect scenario, but happens often enough that you need to deal with it if you’re creating autonomous cars. The other point is about vertical and horizontal representation. A lot of companies feel pressure to say that they can do everything and that they will cover everything. So when you look, when you think about how you represent yourself, the vertical part is about how well you know a use case, how confident you feel that your solution, your algorithm covers one use case to a certain percentage of reliability, while horizontal means how many use cases can you actually cover in the solution? That means that there is traditional trade off between an expert solution and a flexible solution.
Nicholas Borsotto:
And you really need to understand who you are in the situation and where your weak points are. Often communicating those and understanding what doesn’t work for you, it’s a big marker of maturity for a startup. So don’t feel bad about bringing those things up and understanding your own limitations. A lot or actually most companies work in AI, they actually work on the software side, which leads them to not understand and not really put an effort in understanding infrastructure or hardware questions. And I get it. That’s where I lived before. And it’s very easy for companies to just focus on that, what do they do best? But a lot of ISVs go beyond that, and they basically believe that they shouldn’t care about this. “I’m a software company. Why do I care? I run on everything.” When that’s just not true. When facing real life scenarios, you won’t be working just with the top of line servers being used today.
Nicholas Borsotto:
You might have a lot of tech data to work through, servers from different styles. Understanding exactly what is necessary from a harder perspective to actually get you to top performance, even if you don’t know how to build it yourself, it’s crucial. And it’s important that your clients are going to be expecting that of you. Not knowing hardware, it means that whatever you create might still crash and burn when it’s actually being deployed. And a final point that I really like to stress is the idea of, well, to stress test or to have minimal failure tests. That comes from aeronautic engineering, and it goes under the idea of what is the smallest thing that needs to go wrong to break your product and understand that kind of thing.
Nicholas Borsotto:
So in mental health, when I was working before was understanding that a lot of kids who using that product were not neuro-typical. That means that despite being just a small minority of the population, they happen much often when you’re working in that specific industry. So to know how those types of children would interact with an app like that was crucial to see how the product could fail in certain situations that were unexpected. So let’s say that you cover most of your issues on the technical side, you get to a second part, which is basically behavior challenges. You can use different wording to reflect this here, but this comes from a very basic understanding as well. Just like the technical comes from the change between the research experimental setting to the production setting, the behavior challenges comes from the difficulty of ISVs and startups to truly understand the people that they’re talking with. So here, the real motto is falling in love with the problem, not the solution.
Nicholas Borsotto:
A lot of companies come in of this idea that, “Well, I know what I’m doing. That is the solution. Just listen to me and everything will be great.” And that just doesn’t work. That works maybe in one customer, but that doesn’t scale. So my advice is to understand the status quo. The status quo exists for a reason. There are reasons keeping it in place. Know what those reasons are, why the reason that a competitor, that your client is still using your competitor from 10 years ago. Know what it brings to them. And know your audience. Know why they’re using it, what’s important, what’s the experience that they have training on it? Another part is that a lot, while there’s this model and startups of always be pitching, I disagree. I think that there’s a few times where you want to be pitching, but often it’s more about socializing, informing and educating.
Nicholas Borsotto:
By socializing your ideas, it means just adding information very lightly without necessary feedback from your audience. You’re just bringing the idea about. Often you’re not even talking about your solution. You’re talking about the idea. To inform means to add more information on top of this, but at the same point, not expect a large investment from your audience. You need to have those two well with multiple stakeholders inside of your company, to be able to get the few that you would then educate and then help them understand your solution and be your allies in the insight. Talking about allies, it’s very important to understand what are the benchmarks for success in your specific use case. Understand what is the KPIs that they’re looking and understand what would be a too good of a deal to refuse?
Nicholas Borsotto:
I know that a lot of ISVs and a lot of startups get caught on an unending loop of POCs. And that often comes because you don’t understand exactly what is being asked. What is the game changer for that specific team, not only company? Often what you think is great for the company is not necessarily what drives that individual or that team. And then finally is prepare for failure. Failure will always come, and using the retail use case here to help us out, it often means to understand that at some point, your AI algorithm won’t do what it’s supposed to do. And then what happens? If the answer to that is that you, as a startup, has to engage directly with it to fix it, that’s not very scalable. And also the fact that they don’t know what happens if it fails might increase the risk beyond what a lot of companies would feel comfortable with.
Nicholas Borsotto:
So a good example is autonomous shopping. If you only have AI running a store, what happens if something completely out of the ordinary comes to be? Somebody bumps into a shop and they drop 50 bags of Doritos, what happens? What do they do? How can a clerk get involved in this, or how would the person who is managing multiple stores engage? How does your product react to failure? So that could be a loss prevention product that when they don’t know if somebody’s actually stealing or not, that they can send a first message to gauge the intent, like a reminder, like, “Ah, you’re sure that you didn’t forget anything?” So understanding that there is always a chance for failure. Plan for failure. And make sure that you bring that to your customers as well, saying that, “We are ready in case it doesn’t work, that we are de-risking the situation.”
Nicholas Borsotto:
The final group of economic challenges is a little bit different in the sense that they deal with a much larger market. And this is the one that if you ask a lot of great ISVs out there, they think that they have it covered. They have the data, they have everything that they need. But the truth is it is not that simple. And as somebody who has studied economics for a good deal of my adult life, economic rarely is. Economic is a web of effect and counter-effect, and you cannot ever just pull one thread. One thread pulls, another thread, pulls another thread. So you really have to understand, if my solution is applied, what change and what stays the same? Know the chain of events surrounding your product. Understand that there is always incentives and prejudice inside of a company.
Nicholas Borsotto:
Maybe that company tried something similar to yours before, which you think is nothing alike, but they burned themselves and now they cannot do something else. And please stop believing that there is such a thing as free lunch. A lot of ISVs out there think it’s free to test their product. Just because I’m giving it for free doesn’t mean that it’s free for the client to test it. There will always be effort necessary and opportunity cost on applying any AI to any production setting, be that training, be that time, be that resources, even if you don’t see it. So understand that there’s no free lunch and don’t assume that everything’s going to be easy because you’re bringing this for free. It’s much more important to focus on improving the benefits and how you clarify the benefits that come from with than just reducing the cost. Actually, that’s a risk. Just keep reducing the cost is a risky way to get on POC how, where you never actually come to product.
Nicholas Borsotto:
Quantify, it’s important. You need to tell your client what you’re actually driving value. And here I would even add on top that focus on what brings value, not what’s cool. A lot of AI comes into this like the cool factor. So cool, the AI does this, so interesting, the user finds it so cool. But often that is a crutch. That’s a crutch on saying what the real value is about. So for example, I used to work a long time ago with access control systems, and the big issue there is it was extremely tough to say how the AI was better than traditional key cards, which is a technology from the ’80s. And that’s just a fact of life. But rather than trying to face this back in the day, the focus was like, “Ah, because it’s so interesting, it’s app, it’s mobile.”
Nicholas Borsotto:
And in the end that company failed because talking about how cool it is doesn’t cover the huge hole, which was what are the economic metrics they were bringing to the customer? Sometimes AI just doesn’t solve the issue and that’s the clear truth. And that’s actually to bring to a point that is not every customer is your customer. Actually in the beginning, as you’re scaling up at the start of your company, I know that scaling up means selling to a much broader group of people, but that doesn’t mean selling to everybody. Especially in the beginning, you still have to focus on what’s the perfect customer or what are customers that have the right mindset and the right points that will make sense to you. Finally, something that often doesn’t come too often is understanding trade offs, understanding that by choosing one thing you’re saying no to another thing.
Nicholas Borsotto:
A lot of companies go on this, “We can do everything.” And a jack of all trades is a master of none. At the same time that if you’re always willing to change your code or change your product to fit individual customers, you would never achieve scale because if you’re always bespoke, it means that you’re a consulting company, you’re not a solutions company. And I understand the pressure, and it’s important that sometimes in the beginning, as you’re still developing your product, you do that. But you have to keep in mind that with time, you have to do less and less of that to actually achieve scale. So this is roughly the breakdown that I wanted to give to all of you, so the technical issues, behavior issues and economics issues.
Nicholas Borsotto:
Keeping in mind that often the world is much larger than what we get in our tech bubbles. And if we want to sell AI, we are selling value. We are selling solutions to real world problems. In the beginning at least, nobody’s going to ask you about the algorithms that you use and how you get there. You need to first defeat the adoption problem. And that comes from understanding the people that you’re talking with, understanding the chaos naturally implied in the production environment, and understanding that the market is much more complex than what you’re looking because the startup again is made for a very specific purpose. It’s not its objective to understand everything at the start. So from the technical side, keeping in mind that often accuracy is not the most important thing, and that integration and dealing with tech, that is as important or even more important. Use POCs to understand infrastructure, understand why that infrastructure is in place, even if you think it’s not a logical thing in the first whirl.
Nicholas Borsotto:
Understand what failure means and what it means for your product and your client. So when you’re looking at the behavior, fall in love with the problem, not with the solution. Always be asking about the problem from different points of view, from different people, from competitors, and try to understand how that problem actually looks and how it affects other people and other roles. Socialize your idea. So basically bring your idea to as many people as possible, even if you have a solution. Let people digest, let people tell you what they think. You are going to figure out much more by talking about ideas than by talking about solutions. Especially if you have opportunity to talk with possible clients or past clients, make it easy for your internal champions to defend you and show your clients that you know to do shop talk and not only tech talk.
Nicholas Borsotto:
And finally, when it comes to economics, most clients won’t be your clients. That’s a harsh reality. And I’m not saying to only focus on the 1% that actually is the perfect client, but to understand what is actually necessary, what needs to be true for your solution to actually add value for that customer? And what is the opposite? Why doesn’t it add value at all? I really like this expression from a colleague of mine, which is the classic, “Where do we play and how do we win?” What is actually the situation that creates value from your solution, not the general perspective, but the one that you got from interviewing people, to talking, to reading magazines from the industry, to really where do you add value and where don’t you? And focus on that, focus on quantifying the value that you add for that customer and driving those wins. Wins bring wins.
Nicholas Borsotto:
So it’s more important to be in the industry that you’re … the niche that you think it’s right, and to keep expanding than trying to swallow the world and not go forward. And finally understand the trade offs. There’s no problem having a flexible product, there is no problem having a partially consulting business, but understand that that comes with trade offs, understand that you’re giving up another option and that it’s tough to be both. So pick your strengths, pick your weaknesses, and maybe even more important, pick your market, understand it, and then repeat. And maybe the biggest part of it all, if you can, don’t do it alone. Nowadays there’s a lot of great people working with AI in different sections, from integrators to our work that we do in Lenovo with the Lenovo AI Innovators, big companies, incubators. There’s a huge ecosystem out there of people trying to get practical AI to really happen.
Nicholas Borsotto:
So if you understand those issues, but you cannot really cover all of them by your own, look for allies inside your clients, as partners, as integrators, everything else. There’s a lot of people out there trying to make this happen. So pick your allies and go fight together. And if you’re interested in the Lenovo AI Innovators, I’m always happy to talk with you. We work with ISVs from all sizes. Chooch, who is hosting is one big partner of ours. And I’m pretty sure they can talk about nice things. And I’m really looking forward to all the great AI that all of you in the audience will be bringing to the market soon. It was my pleasure to talk with all of you, and I represent here the Lenovo AI team. Thank you very much, Chooch for hosting us and have a good night.
