Building an AI-driven company: An interview with Kai-Fu Lee, Chairman and CEO, Sinovation Ventures

Adoption of artificial intelligence (AI) applications across industries is rising fast. Globally, companies are now more likely than not to implement AI; 56 percent of the respondents in the 2021 McKinsey Global Survey on AI reported adoption in at least one function, up from 50 percent in 2020.

A McKinsey study shows that AI can potentially deliver additional economic output of around $13 trillion by 2030. In China, AI applications are expected to generate up to $650 billion in economic value by 2030, according to another recent study.

Despite these advances, and the enormous value creation opportunities they bring, world-leading AI expert, investor, and author Dr. Kai-Fu Lee believes we have barely scratched the surface when it comes to realizing the potential of AI.

Lee is well placed to offer such insights as his venture capital firm Sinovation Ventures backs hundreds of successful companies that either deliver AI solutions or are using AI applications.

Lee recently spoke with Kai Shen, a Partner with QuantumBlack China, McKinsey’s AI practice, about his expectations for the development of AI.

Lee sees self-supervised deep learning (SSL) driving a second golden era in AI investment, as natural language processing (NLP) applications generate breakthroughs similar to those seen in computer vision over the last 5-10 years.

Lee also offers thoughts on how business leaders can best introduce AI into their operations, what it means to be a truly AI-driven company, and why we are still in the early days when it comes to AI adoption.

McKinsey: How would you define AI?

Kai-Fu Lee: As the study of emulating human cognition and intelligence. The most important subdiscipline of AI is machine learning, within which deep-learning is the algorithm that is having a huge global impact. The two are often incorrectly used synonymously. AI most often means leveraging deep learning in a system that uses a large amount of data to optimize an objective function – to make better decisions, predictions, and classifications – that are aligned with a business goal. A company can use deep learning to predict future sales and stock prices, or to recognize and classify objects and speech.

McKinsey: How do you expect AI to develop in the future?

Lee: Deep learning is a platform upon which other things are built. The two biggest such advances in the last 5-6 years have been convolutional neural networks (CNNs), which allow the generic deep-learning algorithm to perform computer vision: to see and recognize objects and understand scenes at a supra-human level for specific tasks. The other is self-supervised learning (SSL) – for example, training a system on global natural language data to learn English or Chinese – and then fine-tuning rapidly for certain domains. These two examples show that deep learning is not just about matching, decision making and optimizing simple objective functions, but the ability to see, hear, and understand. As we progress with complex tasks such as developing autonomous vehicles and using AI in health care, deep learning will likely yield new enhancements. In the next 5-10 years, deep learning will still be AI’s largest underpinning platform, but new, brilliant ideas like CNNs and SSL will be built on top to solve previously insurmountable problems.

On the other hand, some people believe that because deep learning appears to work better with more data and less human programming, it is not suitable for emulating human abilities, such as making inferences and analogies, or acquiring common sense. The idea is that human cognition is not easily modeled by data and deep-learning inputs and outputs, so it may be necessary to enhance or even replace it with brand new algorithms. That’s possible in the future, but we have not yet exploited all the benefits of the technique.

I am also a bit skeptical about that approach in the near term, partly because a lot of smart people have tried and not really succeeded, beginning with expert systems some 40 years ago. Moreover, amid an abundance of data and computation, deep learning continues to break new ground and achieve things that were hitherto thought impossible, so there is still a lot of room to grow.

McKinsey: What are your thoughts on how AI will be applied to language? Do you think the impact will be even greater than what we have seen in computer vision?

Lee: Yes. While we absorb real-world information primarily through vision and secondarily through hearing, language will more deeply impact AI’s business and scientific growth, because language is how we communicate and capture our knowledge and thoughts. We are at a similar stage as we were in 2012, when Geoffrey Hinton and others showed how CNNs enabled computer vision. ImageNet performance surged and was projected to be on a par with humans in 3-4 years. Among Sinovation’s biggest successes was to recognize that computer vision would outperform humans and change the world, and to invest in CNN and deep learning based on the expectation that when that happened, the result would be applications that either worked symbiotically with people, or in many cases replace them entirely, saving costs. We then saw a proliferation of applications, some controversial like deepfakes and facial recognition, but others that are universally accepted as technological breakthroughs, such as autonomous vehicles, robot perception, recognition in radiology and pathology, digital, image, video, and 3D data, manufacturing inspection applications, and so on.

About two years ago, OpenAI’s GPT-3, or third generation Generative Pre-trained Transformer, launched a new language-learning paradigm based on the fact that though AI works better with more data, you cannot generically label trillion-piece data sets. It’s clearly insufficient to label language data as just nouns and verbs. You can label for specific tasks, like building an airline reservation system, but you cannot do canonical-universal labeling.

So GPT-3 abandoned labeling entirely in favor of training a new data brain, feeding it all the data in the world under the premise of using the past to predict the future, with the highest level of fidelity as the objective function. This system self-organized into a network that encapsulates and understands the essence of language, probably not in the way that humans do, but sufficient to build systems for reservations, chat bars, speech recognition, machine translation, new search engines, question answering, advertising targeting, and so on.

Our view is that a second golden period of AI investment will unfold as natural language processing (NLP) applications proliferate in the same way as occurred with computer vision. We have already made four investments in NLP companies, including one that has delivered the leading performance for Chinese NLP and built a GPT-3 transformer-like model, shrinking it by 1,000 times to make it practical. They also built an English-Arabic machine translation system with just one engineer and two interns in about three weeks. No one in the team even knows Arabic. That’s a good example of how building a large SSL-trained model on global data and then fine-tuning for specific applications and languages seems to work. Similarly, we have illustrated how rapid-customizing NLP applications founded on large models work in fields like targeted advertising, which is now very powerful because you can show a different ad copy based on your understanding of the individual. It’s also being applied to speech recognition. In the next five years, we’ll see greater proliferation, impact, and probably valuations of natural language companies compared with what we saw with computer vision 5-8 years ago.

McKinsey: To use a basketball analogy, what stage of the game are we at in terms of the use of AI in business?

Lee: We’re clearly in the first quarter. The score might be 7 plays 8 in a game where we have made a three-point shot in deep learning and a couple of two-point shots in CNNs and SSL. So perhaps the first 2 minutes of a whole game. We have much farther to go – as I mentioned in my book “AI Superpowers”, we’ve just scratched the surface. How many companies are really using AI? It’s a single-digit percentage, and even those companies are not using it as extensively as they could be. There’s so much more opportunity in implementation. For example, a McKinsey study shows that AI can potentially deliver additional economic output of around $13 trillion by 2030. We’re nowhere close to that, so we have much to look forward to.

McKinsey: What does it mean to become an AI-driven company?

Lee: Firstly, it means becoming data driven, because without data, you can’t do AI. The company needs to invest to digitize everything that can be digitized so you have the nutrients for AI to work. Think about data and storage not as a cost center, but in terms of creating your most valuable asset. You’ll never get there if you treat data collection and storage as a budget that increases 5-10% a year. A big mindset change is required. Then use big data to visualize business intelligence. Once that is done, more and more corporate decisions should be made based on data rather than experience – and certainly not on intuition. Then look for low-hanging fruit to automate, usually to drive cost savings – things that take longer for people to do than machines, including making decisions. Explore every possible way to improve margins or acquire customers. Link everything that can be defined as a quantifiable business metric to AI to optimize and propose solutions that work symbiotically with people. AI can also perform data-rich, relatively routine single-domain tasks entirely on our behalf.

McKinsey: If you were a CEO of a traditional company that wants to become AI-driven, what would be the first business problem that you would use AI to solve?

Lee: First consider that the company may misunderstand AI; that executives will naturally be skeptical or have unrealistic expectations about its efficacy. I would bring in experts to advise on finding a single, data-ready task that that can be linked to a business result, so that when AI is implemented people will say, “Oh, that’s what it is and it does work.” Then I’d think about other opportunities where we have a lot of data and a business metric that can be linked to an objective function, whether it’s cost savings, margin improvement, or marketing to customers in a more targeted way. If the company has no data, then I would add the hard problem of picking a domain in which data can be gathered with reasonable costs.

But my primary goal would be to inspire my executives and leadership team with a vivid working example of AI in action, as this will get people’s creative juices flowing and generate more ideas for applications. That first implementation is important because if it fails, either because it doesn’t illustrate business impact, or the data is sparse or faulty, or the implementation is sloppy, then the leadership team will lose confidence.

McKinsey: Can you give an example of how a company might start that process?

Lee: One way is for the company to describe, without thinking about AI, its business drivers and challenges, and allow experts to apply AI and other technologies to deliver a solution. For example, a steel producer’s biggest issue was that their liquified iron cooled too quickly during export. Smart AI in terms of logistics management, autonomous vehicles, and safety sensors solved the problem. Once we had gained the company’s trust, they raised more problems, which were easily solved because we’d already digitized and installed sensors to gather data.

If the choice about where to make that first impact is not obvious because you don’t have the data, it is important to understand how much it will cost to gather. Initial data collection is easy. It’s cleansing data that often requires more resources and time than executives expect. Once the data is in good shape, the AI part is not as time consuming and difficult as people think. Identify the problem, obtain the data, and understand how much cleansing it costs, and implementation will follow.

McKinsey: That’s a great way to think about getting started, but what does it ultimately take to become an AI-driven company?

Lee: The business process needs to be fully digitized, and every task where AI can do a better job than people should be AI-assisted or automated. Executives anxious to protect their turf who are hesitant to embrace AI need to be replaced or enlightened. That is, first and foremost to trust data, to make intelligent, data-driven decisions, and deploy AI. If you don’t do that well, someone else will eat your lunch, AI upskills the employee base because it is best at doing routine and quantitative things while employees do higher level creative things that improve competitiveness. AI should be used in every aspect of management – not just in R&D or technologies, but in the HR department to pre-empt key employees leaving and to filter incoming résumés, and in marketing to optimize and customize EDMs (Electronic Direct Mails) to increase the likelihood that the customer reads it. AI should be used in sales and IT operations management. There should not be a department in the company that isn’t using AI tools to improve its performance.

McKinsey: Some quickfire questions to finish: which is more important, data or algorithms?

Lee: Data. You need both, but without the data, you can’t do anything. If you already have a reasonable algorithm, I would get more data rather than tweak the algorithm.

McKinsey: Which is more important, domain knowledge or AI knowledge?

Lee: In some domains algorithms can be very important because the data is relatively simple, like for an internet company with user data. But in some areas domain knowledge is incredibly complex. It’s not just building the application, but knowing how to sell through the right channels. Health care is a good example. So you need both. First think about the domain that you want to go into and whether it really requires domain knowledge. If the answer is “yes”, I’d put domain knowledge first.

McKinsey: When a company is selling AI as a solution, is scalable product or faster customization most important?

Lee: It is customization, because we’re not quite there yet with an AI platform that can address massively different needs. Customization is a necessary evil without which there is no business. I hope if you ask me this question again in five years, I’ll say scalable products because we as an AI community will have that figured out by then.

McKinsey: Is building an MLOps platform (a means of automating machine learning algorithms) or driving cultural change most important for a company adopting AI to transform its business?

Lee: Driving change is the more immediate requirement because I’ve seen so many companies have a difficult time doing so. Once you’ve done that, then you can look at MLOps.

Explore a career with us