Future of Business: Baidu’s Robin Li on the Technology Trends that Will Transform Business

Southern New Hampshire University is an accredited institution with over eighty online graduate degree programs. That means you get the flexibility you need, the credibility you expect, and career-focused courses you can count on. Visit SNHU dot edu slash idea. If you listen to IdeaCast for conversations and strategy, you'll want to check out Founder's Mentality, the CEO sessions. In-depth conversations with top executives about the judgment calls that define their careers. Follow wherever you get your podcasts. Welcome to the HBR IdeaCast from Harvard Business Review.

I'm Alison Beard. Baidu launched in 2,000 as a search engine platform. Fast forward two decades, and it's now one of the few companies in the world that offers a full AI stack. Its core businesses span mobile, cloud, intelligent driving and other growth initiatives, and its products and services have attracted hundreds of millions of users and hundreds of thousands of enterprise customers. Today's guest is leading all of that. For the third episode in our special series on the future of business, we'll hear from Robin Li, the co-founder, chief executive and chairman of Baidu.

He explains how his company has built generative AI into its business, the technology trends he's keeping an eye on, and how he anticipates these tools will transform our lives. Robin spoke to HBR editor in chief Adi Ignatius and took questions from the audience during our recent virtual Future of Business conference. Here's their conversation.

So, Robin Li, I know it is very late your time in Beijing, so we appreciate your joining us live for this welcome.

Hi, Adi, thank you for having me. It's great to be here.

Well, it's great to have you. Before we start, let me just remind everyone in the audience to put any questions you have for Robin in the Ask the Speaker chat, and I will try to get to as many as I can later. But Robin, let's get to it. So Baidu, uh, your company introduced a ChatGPT-like product, Ernie Bot, last year that last I saw has more than 300 million users. I assume it's been a learning experience for you.

Can you talk a little bit about what you've learned since the first version came out and how it has evolved, and just tell us a little bit about all that?

Yeah, sure. We launched Ernie Bot, um, I think March 16th, uh, of last year. Uh, I think that that was the first ChatGPT-like, uh, chatbot for all the public companies around the world. Uh, because we we've been investing in AI, especially, uh, natural language related AI for quite a few years. We were able to quick, uh, launch, uh, a uh, a chatbot based on our our large language models. Over the past year and a half, uh, a lot has happened, uh, the the technology has evolved very quickly and and dramatically.

Um, things we we learned, uh, there there are a lot of things I I should mention. Uh, the first is that, uh, a lot of people, uh, users, developers, uh, customers, they not only care about the efficacy of of the model, they also care about, uh, the the response speed. They also care, uh, about the cost of the, uh, inference, uh, cost. So after March of last year, we have rolled out a series of language models or foundation models to satisfy all kinds of different, uh, needs in in different scenarios.

Uh, uh, meaning that the the model size could vary greatly, uh, and the inference cost could, uh, be very different too. And in certain cases, uh, users don't mind to wait, uh, you know, ten seconds to get the best answer, and, uh, in other scenarios, you have to do it very, very quickly, sub-second, uh, response time. And also in terms of cost, we've, uh, uh, being able to reduce the cost by about 99%, meaning the current inference cost is about 1% of the original cost when we first launched that.

Uh, having said all of that, I would say that, uh, the probably the most significant change, uh, we we're seeing, uh, over the past, uh, you know, eighteen to twenty months, is the accuracy of those, um, answers from from the, uh, large language models. And I I think over the past eighteen months, that problem has pretty much been solved. Meaning that when you talk to a chatbot, a a frontier model-based chatbot, you can basically trust the answer.

That's a huge difference.

Now, from from my perspective, maybe this is a a US perspective, you know, there was a huge wave of excitement about AI, particularly with the release of of, you know, generative AI products. You know, you've talked about search. Um, there don't seem to be a lot of or or as many interesting use cases as as maybe some of us had expected by now. So I I'm interested in your view. You know, are we are we in an AI bubble at this point?

You know, w- what's the trajectory of of the technology?

I think, like many other, uh, technology waves, um, bubble is kind of inevitable. When you pass the stage of initial excitement, people would, uh, be disappointed that the technology doesn't meet the, uh, high expectation generated, uh, through the the initial excitement. We've seen this, uh, many times, uh, when the internet took off in in the mid to late Ninety's, uh, and uh, there was a huge bubble. For mobile internet, similar things happened. And this time for generative AI, I I think we are also go through that kind of, um, uh, period too.

But uh, I I think it's it's also healthy. It will wash out a lot of those, um, you know, fake innovation or, or, uh, products that doesn't have a market fit. After that, probably 1% of the companies will stand out and become huge and will create, um, a lot of value, will create tremendous, um, uh, value for for the people, for the the society. And I think we are just going through this this kind of process.

This year, the sector is probably cooler than than last year, but, uh, I I think it's also healthier than than last year.

What's the and what's the right business model? I mean, there there's some large models, you know, Meta's Lama, for example, are open source, others are closed source, like OpenAI's, uh, uh, you know, GPT. Baidu, I think, is advocated for a closed source approach. What's the thinking behind that? And and and you know, how does that set Baidu up to capitalize on on AI technology?

Yeah, you mentioned closed source, but I would prefer call it a commercial grade, uh, model, foundation model. I I think, uh, when when you look at the most advanced language model or foundation model, most of them, uh, are closed. And when people talk about open source, it's kind of, uh, misleading to me. It's different from the open source of Linux or or or PyTorch, because for an open source model, you what you basically get is is a bunch of parameters.

You don't know, though, how those, uh, parameters was, um, uh, derived, and you you have no way of changing of those parameters. So it it doesn't have the effect of many, many people from different part of the world and contribute back to the main branch and make it better a and better. Language models are very different. And uh, you you can use an open the so-called open source model to do things, but, uh, it's very hard for you to contribute back.

And another perception people have on the open source, uh, stuff is that it's free or it it's actually cheaper, at least cheaper than the the, uh, commercial ones. But, uh, in in the, uh, foundation model area, that's also not true. For our bins, uh, model, I I think, uh, we try to support all kinds of applications, all all kinds of customers, both, uh, external are, uh, you know, cloud computing customers, and also our internal customers, Baidu search, Baidu map, and, uh, uh, Baidu, uh, Wenku.

There there are lots of, um, applications that leverage the power of, uh, Ernie Bot, and, uh, basically we we charge, uh, the the the inference cost for those, uh, API calls.

So I want to switch now to, um, robo taxis. Um, a few days ago, Tesla announced its robo taxi plan. Waymo has expanded its service. What do you think? You know, are are are we at a point where large scale development of robo taxis, including, you know, your own product, you know, has it arrived? Is this the moment?

Yeah, we've been investing in the self-driving technology for over a decade. It took us a very long time to get to this point. Uh, meaning that, uh, we now have more than 400 cars in the city of Wuhan that covers about 9 million people, and, uh, a lot of people in in that city already used to taking a robotaxi. Our brand is called Apollo Go. They pay a fare that's typically, you know, cheaper than, uh, uh, regular taxi.

And, uh, I I think the technology is ready in certain restricted areas. It it's not ready for, uh, anywhere, anytime. Uh, in, um, our, uh, professional terms, it's called the level level five is that you can drive anytime, anywhere. But, uh, for level four, I think we are at this stage, the the we we are at level four, which means that when you know where which area, uh, you are in, and, uh, you you you can get rid of the human driver and provide a right hailing service.

Uh, we we cannot do that in the most crowded, most complicated, uh, traffic area, uh, yet, but we can do that in most of the the, hm, areas in in in, uh, most the cities. I think US is probably at a similar, uh, stage. Right now, the the bottleneck is is more about regulation. In, I think, in most cities around the world, taxi service without a a driver is not allowed yet. So we basically try to go to those select few cities, uh, that regulations allow us to operate such a service.

And, uh, it's it's a gradual process. It's probably gonna take probably ano- another decade for the robotaxi service to be really everywhere to to become mainstream, but, uh, uh, it will gradually become a service that people, uh, prefer.

So you you've talked about the inevitability of robo taxis putting human drivers out of work. I I'm curious your perspective. You know, you can't talk about AI without at at least discussing the issue of job displacement more broadly. What do you think? You know, do you think AI generative AI will replace humans on a on a large scale? And and if so, how do we prepare for that?

Yeah, uh, a lot of people compare the generative AI revolution to, uh, the industrial revolution. If you look at look back the industrial revolution, similar things, uh, happened. Uh, a lot of old jobs were, uh, you know, get rid of, but, uh, uh, more new jobs were created. Every time when, uh, innovation, when when, uh, you know, technology revolution happens, the jobs that got lost are those hardest, toughest jobs, tho- those jobs that's not so pleasant to human being.

And the jobs, the new jobs that got created were the jobs that that's more comfortable, more decent and less, uh, stressful. Uh, I'm optimistic that this wave of, um, innovation or generative AI will do the same thing. Another point I would like to make is that, uh, this kind of process, it's not something that's gonna happen overnight. It will take, you know, ten years, maybe twenty years, some maybe thirty years to happen. So I think human being have time to prepare for that.

And, uh, uh, we need to be proactive. And, uh, companies, organizations, governments and and ordinary people all need to prepare for that kind of, um, paradigm shift.

So I'd love to talk about China more specifically, and, you know, China's approach to AI. Um, how how, you know, do you see a difference between China's path to to generative AI development and the approach taken by the rest of the world?

Yeah, I I do see some differences. The the most obvious difference is that China is more application driven. Uh, we hear more about, uh, what kind of applications can benefit from this kind of frontier models. And, uh, a lot of, uh, startups are are trying to find, uh, ways to leverage the the power of foundation model. Companies like Baidu, our strategy is to reconstruct and rebuild almost every one of our existing product based on the Ernie Bot, um, foundation model.

We have already seen, uh, very big, uh, changes, um, in our existing products, search being the first and foremost thing. Uh, right now, over 18% of the Baidu search result, uh, are they generated by, uh, Ernie Bot. Uh, we also have this kind of, uh, phenomenon in China, uh, live streaming shopping. I know it's not that popular in the US, but it it's a big business in China. Uh, but live streaming requires a a real human to devote to that kind of, um, uh, service, uh, full time.

Uh, but uh, uh, we now can create digital digital human to do live streaming shopping. Uh, the the scripts, uh, can be generated by Ernie Bot. It look like a real person. It's it's it's not a avatar per se, uh, can be can look very real. Sometimes the consumer, the the shoppers cannot just tell whether it's a digital human or or it's a real human. And during the live live stream, the digital human can also interact with with the audience and answer questions and react to, uh, their shopping activities, things like that.

Um, I I think we just got started. There are there are lots of lots of, uh, use cases. We we're saying that, uh, by leveraging the the power of generative AI, businesses, uh, can, hm, uh, get much better ROI. They they can get more revenue, and they can save, uh, more, uh, cost.

So just to follow up on that, uh, I I don't want to get into politics at all, but you know, there's there's a concern, uh, certainly in the US, about, um, you know, in in dealing with with with Chinese tech companies, you know, TikTok, uh, is facing a situation now that the the, um, uh, that the data will will, you know, could be misused, that our our data could be, you know, just given the difference between the Chinese system, the American system, the, you know, Western system, that that data is not safe, that the government could have access to it, et cetera, et cetera.

How how do you how do you respond to that in terms of your, you know, your company, your products, and about those concerns?

Uh, well, first of all, Baidu is a NASDAQ-listed, uh, public company, and, uh, we would follow, we will we we would comply with any applicable law. So for whatever, uh, practice we have, we would fully disclose that. We do respect, uh, uh, user, uh, privacy. We have, uh, uh, uh, data, uh, compliance committee within, uh, our company with, uh, very, uh, high-ranking executives, um, uh, being the the chair, and, uh, we we, uh, we take this very seriously.

And, uh, Chinese consumers are not different from any consumers for the rest of the world. They they care about their data. They don't want other people to to look look at their data. And, uh, in order to, uh, gain the trust of our users, uh, we need to do the right thing. And in order to, uh, stay, uh, as a public listed company in in the US, and, w, so we we also need to comply with all the applicable laws in in the US, too.

Okay, uh, I wanna get to a couple of audience questions, and there are a couple, uh, one from CLJ, one from Vicky. Not sure where they're from, but they're both asking about, uh, sustainability and environmental costs. So here's one, you know, how do you think about the environmental costs of AI? When you consider a return on investment, when is the use case, you know, too taxing on the environment, too much for the environment to be valuable as a product or to remain valuable to society?

Yeah, that's a very good question. Uh, but uh, I I would argue that, uh, the efficiency gain, the value created through this kind of innovation will more than, uh, you know, offset the the power consumption. We will find ways, uh, to generate green energy faster than without generative AI. Uh, we will be able to complete other kinds of tasks better and faster and, uh, at at a lower, uh, cost. So net net, I I I think, uh, generative AI will have a positive impact on the environment.

All right, so we probably have time for just one more question. This is an audience question is from Brian, and do what you want with this one. Brian asks, What do you think the world will look like in ten to fifteen years? How, you know, what will our interactions with technology be like then?

Yeah, it's very exciting to think about that. I I think, uh, uh, generative A AI is, uh, is really, uh, disruptive, uh. I I think it will give people, uh, give a lot of the ordinary people the power to be a programmer. What what does that mean? Nowadays, the the the engineers, the programs get paid like, let's say, you know, 200,000 per year, and that's because they they are quite powerful. They they can program things, they can create software that's very valuable.

And I I think ten probably doesn't take more the doesn't take ten years. It we should see that prob around five to ten years. And, uh, uh, and a person who can speak natural language, be it English or or Chinese, can have the power of a software engineer. You can imagine that how, uh, how much, uh, you know, productivity we we can have just by having that kind of, uh, capability. And when when when I was in college, we we learn programming language, we we then we learn assembly language.

Nobody use assembly language to do programming anymore. Nowadays, people use Python or C plus plus. Uh, five, ten years from now, nobody use Python or or or C plus plus anymore. They are just use English or or Chinese to do whatever they want, and everyone can do that. So think about that. The the world will be completely different ten years from now.

That was Robin Li, the co-founder, chief executive, and chairman of Baidu, speaking to HBR editor in chief Adi Ignatius at our recent virtual Future of Business Conference. I hope you listen to all of our Future of Business series and all of the episodes we have on the HBR IdeaCast about leadership, strategy, and the future of work. Find us at hbr dot org slash podcasts or search HBR on Apple Podcasts, Spotify, or wherever you listen. And if you don't already subscribe to HBR, please do.

It's the best way to support our show. Go to hbr dot org slash subscribe to learn more. Thanks to our team, senior producers Anne Sanni and Mary Do, associate producer Hannah Bates, audio product manager Ian Fox, and senior production specialist Rob Eckert. And thanks to you for listening to the HBR IdeaCast. I'm Alison Beard.