AI Agents and the Future of Global Trade with Alibaba’s Kuo Zhang - Ep. 291

Foreign.

0:00

Hello and welcome to the Nvidia AI podcast. I'm your host, Noah Kravitz. Since 1999, Alibaba.com has served business to business, e commerce buyers and suppliers from over 200 countries and regions around the world. Kua zhang, President of Alibaba.com is here with us today to talk about how AI agents like Alibaba's recently launched Aktio will be reshaping our massive global trade ecosystem over the next decade. And while we've talked a lot about agents in recent episodes of the podcast, there are few people in the world who have Kuo's perspective on how technology shapes global commerce, which is why I'm so excited to welcome him onto the podcast. Kua, thanks so much for taking the time out. Thank you to join Nvidia's AI podcast.

0:10

Thank you for having me.

0:57

And I know you're joining us while you're traveling, so an extra special thanks. Appreciate you making it work. So let's get right into it. Maybe we can set the stage. Or you can set the stage. Tell the audience a bit about what Alibaba does and Alibaba.com in particular, and then you can talk a little about your role as president of Alibaba.com?

0:59

okay, so Alibaba.com is the first business of Alibaba Group. It is founded in 1999 by Jack Ma and the 18 founders. So it started as a yellow page and now it's kind of evolving to a leading B2B platform in the world. It connecting around 50 million buyers, B2B buyers in yearly basis and more than 200,000 suppliers globally and now enabling more than US$60 billion transaction in yearly basis. So this is about Alibaba.com today. And actually I joined Alibaba since 2011. So my first role is in Taobao and Timor. It's kind of in China, domestic B2C business. And I joined Alibaba.com since 2017 and becomes a president about five years ago. So yeah, this is quite a journey.

1:19

Yes, I can only imagine how. I mean, maybe I don't know if you can speak to this in a short amount of time, but I can only imagine how technology and from your perspective has changed so much on the inside on the buyer and supplier side and of course what Alibaba has done and is doing building platforms, because as a consumer, the way that I purchase things way down at the end of the chain has changed so much. And so I can only imagine from your perspective and the work that you've done and continue to do now, just really how things have transformed.

2:22

Right. So what you say is correct. I want to echo that is one of our dream is to make global trade as easy as online shopping.

2:57

Yeah.

3:10

Think about how we kind of reshape about the E commerce. So the people buying stuff is really easy. And now when the buyers want to source or want to buy something from a kind of supplier overseas especially is still they will meet a lot of challenges including the language barrier, time difference, the time zones and the culture difference, the trust issues. So how to kind of, hey, how to settle the payment, how to settle the logistics, how to settle the kind of after sale services, so on and so forth. And yeah, so there's a lot of things to be done by technology. So before AI, actually we already set up kind of the first the demand and supply system, the search and the product listings and the online communications, kind of the live shows. And then we build up the transaction systems including the payment and the payment networks and the logistics networks and then set up the entire trust systems, kind of the B2B Astral payment for the Binance suppliers in the B2B scenarios. So that's, I think we already set up a kind of standard E commerce platform for the B2B and now we know that the AI era is coming. So we see a lot of improvement space in here as well. That's why we introduced Axio.

3:10

So let's talk about Axio. From the materials I read and I looked at some videos online, it's described as an AI agent designed to help you do business, which sounds simple enough, but as you've alluded to, there's a lot that goes into doing business B2B, particularly on a global scale. So can you tell us what does Axio do and how does it fit in? Why is it so important to Alibaba.com's overall vision?

4:49

Sure. So I can share with you our kind of vision while we do it and I can share you some data to prove it.

5:15

Perfect.

5:24

So Axio actually is an AI native application. So when we build this application, we are building it on top of the kind of SOTA model we see when the people are now using Axial, it's kind of the behavior is different. Like they're using the traditional search engine or traditional platform before the first four, they are using the kind of natural language or long sentences to describe about their request. So previously they may use the language like I want to buy a kind of how to say the portable energy storage to buy something like that, like

5:25

a generator or a battery?

6:05

Yeah, it's like a battery. Battery.

6:07

Okay, okay.

6:09

And now they can describe in a kind of full sentence, like what is the scenario? This battery is usage. So what the tabs look like is like a kind of suitcase. It's portable with what kind of protection and the dimensions of the size, the kind of weight you can put it together. And then the axial engine can understand what you are requiring and kind of break down into different elements and match the products and match the suppliers. So it's a kind of a completely different user scenarios. And we see that people are putting much longer sentences, more natural sentences into the axiosystem. I think that's one. The second is now we are introducing agentic model, actually the agent model to the axiom. So now it's not only doing the search functions, but also it's acting as an agent. So meaning that you can give them a very complex task, you can execute upon the task and give you and deliver the result to you. And what we see is that in axial, the user, the audience in axial and in Alibaba.com, so only 30% of coverage, meaning a lot of people actually is using Axio to do the first online sourcing for the global tree. And it's completely kind of lower the barriers for people to enter in this field.

6:10

So how does just you described it, but kind of to dig in for a second, how does the experience differ? Or maybe you can unpack a little more of all the different things that have to go into actually buying an item business to business globally. And you kind of alluded to, you know, building the network of trust, the technical infrastructure, the logistics, payment, other trust factors. Can you describe maybe some of the things that traditionally have been done manually that can take up quite a bit of time, particularly as you said for someone new to doing this, that now Axio can sort of take care of and automate.

7:36

Sure. I can give you two examples. Okay, so these two examples, actually my team tell me that just one week or two weeks ago, how they see the people are using the system. So one example is this kind of a supplier for Bolivian Games. It's a kind of mini size of Olympic Games helping in Latin America.

8:18

Okay.

8:48

About six countries is attending that games. And once the supplier is actually sourcing the items for these games, you can imagine they are sourcing all different products from metals to gears to clothes, protection stuff. And it need to kind of apply to the local compliance regulations. Previously they need a kind of team with expertise to sourcing from all kinds of suppliers, maybe hundreds of suppliers to support such kind of game. So now what do we see? That they upload a file like Excel. So telling about all the specifications they need. The items are in hundreds and thousands of them. And you can just upload this file to axial and axial can understand what your requirement understand this is kind of this product is going to be used in Latin America. Need to follow the local compliance and guidance and then it will simultaneously execute these tasks. And previously you may take weeks, even months to finish this sourcing list. Now using asylt, it can be finished in hours or in minutes. And then you can give you all the kind of the suppliers who can make this product and give you a suggestion. Then you can use that to send inquiries and even can take this step further to communicate with these suppliers. I think that is one of the examples. I can tell you how the agents can collaborate together to help you. The other I can give you a more. So this is a kind of requirement from expertise. So who actually have this sourcing experience just need this agent to help them to execute on the complicated task. The other examples I see is that people just have idea. For example, one of the ideas I want to design a clothes for the adhd. So what a type of kind of materials is and how to design these kind of charges, which is it can help the ADHD children. So then the Axio can help you to start from the marketing research to see what the existing product is and give you suggestions step by step and then give you the suppliers and the product recommendations and even can help you with the design prototype. So all this kind of stuff from the marketing research to the product design or product redesign, all the kind of find the suppliers who can make the product. All this stuff can be executed by this agent.

8:49

So it's really almost the full business life cycle.

11:40

Exactly.

11:42

And so you sort of read my mind. You almost immediately got to what I was thinking which was how far down the line can it execute? Like you kind of mentioned, you know, providing that Axio could provide the sources and then you know, I would send the inquiries, but actually the system could send the inquiries and agent to agent collaboration. Can you talk a little bit more about that? Either what it can do now or sort of what you're. You're working on or is possible if you can speak to that.

11:43

Okay, so first I can tell you the how we design the system and then I can tell you about the boundaries. Like I think you have a lot of questions that where we go, where we stop, right? So the system is working like this. So you first start from your questions. You send a request. Either it's a kind of unnatural language or it's a multimodal, right? You send out a drawing, a design or kind of file or PDF or Excel list and interpret your request and then orchestrate into a kind of set of tasks that you can execute upon

12:15

that can be at the level of a business already running, that has sophisticated sourcing needs, or like the example you mentioned with the games in Latin America. Or it could be something like an individual, as you said, for instance, who has an idea for clothing design, tactile clothing for ADHD wares. And so it could be somebody who doesn't know anything about sourcing, even as you said, it could help the system can help them design. So any level of expertise coming in.

12:55

Exactly.

13:23

Okay.

13:23

And for that part, actually it can involve the human, actually the users to interpret. So like here are the tasks orchestrated or designed for you. What do you want to change? If not, then it will execute upon these tasks using all the sorta model to give you the best result. So that's the second step and the third step is to reevaluate. So in many of the cases in this kind of global sourcing or global trading scenarios, so the decisions is not always quantitative tasks, many of the time is qualitative tasks or decisions as well. Then you need to evaluate whether the decisions or the answers that we give is the right answer. It will be proved by the platform like Alibaba.com to see if this is a good answer or this is a good output. If not, then we will iterate the whole process and see how we can kind of evolve and help the people. And whenever actually there's a decision cannot be made by let's say the machines, like when you make a deal, so what precise, what kind of conditions that you can offer, then it will involve the human to make the final decisions. Or when the kind of the AI actually exceeds its boundary, it does not have this kind of knowledge, it will come back to the humans to make sure that they understand and can execute the path. So this is how we kind of execute this whole system.

13:24

You mentioned in your example, sourcing materials for an event in a particular part of the world. Some of the just situation specific, cultural specific, location specific things that have to be dealt with for different projects, even if they might look the same sort of on the surface. How do you build for a global audience? Alibaba Alibaba.com been serving a global audience for some time now. But when you're building AI systems, agentic systems, how do you ensure that the AI driven solutions work across these diverse markets and use cases?

14:58

Right. So the first four is a kind of combination of the word model and the domain specific models. The word model meaning that so when there's a kind of specification for a kind of country or region, this specific situation or the rules will be learned by the model. So when you kind of ship product to that area, so what kind of laws or regulations that apply to and the domain specific kind of information or the knowledge is maintained by audible.com so you know that we have more than managed 200 different suppliers and they upload all kind of certificates and all different kind of product details that we can understand. And together we know that how we can apply different products into different scenarios, into different countries. And also it's kind of a human and machine interaction system. So in some of the cases that we involve human to make the decision or make the kind of judgment as well, if it already exceeds the boundary of the AI system or the knowledge base. And the third but not the least is that we already actually Alibaba.com already managed a system which we comply with all the regulations, all the rules, and also we already managed to process more than 60 billion US dollar kind of transactions. So we know that how to build a systematic approach to protect. So these three layers actually together that we can deliver a better result for

15:32

the global treaty when it gets into more kind of nuanced or maybe sort of qualitative as opposed to quantitative information insights, cultural nuances, you know, just things that a person might sort of know and act on intuitively, but maybe have a hard time or just never think to express in words. Do you approach training the systems in the same way? Is it are those types of knowledges and capabilities kind of more dependent on learning by use? How do you infuse the system with the sort of local nuanced information that can be really important to closing deals?

17:15

Okay, this is a very good question and I think we are keeping working on that as well. So what we're doing now is we can say that it's three layers. The first layer of course is about the data set. So knowing that we have more than 260 million products and suppliers and transactions and buyers, that we through this data that we understand what the demand and supply look like and we can kind of abstract the domain knowledge from that as well. So I think that is the first layer. The second part is about the industry know how so that part actually is based on the not only the model but also the industry expertise. So I can give an example. So when we say we design a product for a specific scenario, so how we evaluate that design is a good design. It's like the questions of how we evaluate the answer is a good answer. So that need to be kind of rely on the expert systems to evaluate all these answers and to kind of keep improving the system. And the third is about the platform itself. So when we outcome kind of a result based on the requirement and then we will put that result into our platform like Alibaba.com and then we will see the conversion rate and we'll see how the kind of the buyers and the sellers are iterate with this output. So if this is not good and then we will iterate the system, the models, the data set as well. So that is how we kind of solve these problems. So you can imagine that we are leveraging our data, our kind of industry know hows but we need to iterate the system in real time based on the system's feedback.

17:55

I'm speaking with Kuo Zhong. Kuo is president of Alibaba.com and we've been talking about Accio. They're relatively new. When did Achio launch?

19:54

Kuo, the first version launched last year, but the agentic model, the agent version launched just last month.

20:04

Last month, okay, thank you. I didn't want to be imprecise when I have the perfect source of knowledge right here to tell us. So last month Accio Agent came out building on not only Accio itself, but as you've been saying, Alibaba's incredible database and just knowledge repositories of years of serving the global business community. I want to ask about the user end of things. You mentioned before, some of the user behavior kind of comparing between search and using the AI powered interfaces and that kind of thing. But when you're talking about users whose businesses rely on your platform for what they're doing and there's a lot of sophisticated moving parts as you talked about, how have the users responded to to ATIO in particular and moving to more AI automated systems because automation's obviously been a thing, but now in the AI age it seems like there's an increased or even new layer of trust that would have to be built with users. So how do you build that trust? And on a technical, from a technical side of that, what Guardrails, transparency measures. You've talked a little bit about safety, but when you're looking at an agentix system. How do you go about building those guardrails into it?

20:13

So as I mentioned. So first of all I think it's always a human and machine interaction systems. So whenever that we think the decision that we made by the AI system or by the big large language model exceeds the boundary or without the knowledge that we have, we always involve humans to make decisions like when you make a deal, when you're negotiating a price or conditions and we always using our platform kind of to reiterate the model to see whether it gets a better result, there's a better kind of conversion rate, so on and so forth. So this is basically we are doing in daily basis.

21:30

Can you talk a little bit about what small and medium sized enterprises mean to the world, have meant to and kind of inspired your own work and then kind of talk about that or talk about it through the lens of a geo trade and these AI systems that you know, not only speed up the process incredibly as you were talking about, you know, taking this weeks long process, boiling it down to hours, minutes in some cases, but also as you said, make it available, lower that barrier of entry so that you know, the solo entrepreneur or the team that has an idea, but maybe not the knowledge and the resources and the technical skills can now access this global market. What excites you most about doing this whole thing and putting these tools into the hands of smaller businesses?

22:14

Sure. So you know we hold a co create just tomorrow in Vegas. So we are, the whole team is preparing for that.

23:02

Okay, Early September. For listeners listening down the line, we're talking.

23:10

That's right. And during co create event we have a co create pitch. Okay. So in this year actually we say there's more than 25,000 applications amazing. For the co create pitch just in 30 days. And among these applications I think more than 40% of them mark them as a solo entrepreneur. Okay, so it's meaning that a lot of people actually have their ideas about to build their product, build their process based on the global supply chain. But they need to do everything by themselves from product design to the handle the customer complaints to execute upon all those logistics financing systems. I think that agentic model can help you help this out. Solo entrepreneurs at least in different perspective. So we see the number one data scenarios using by Axio is to find suppliers. It's just like who can make this? I have idea who can make this find their business partners. The number one scenario that they use SEO is to help product design or following the winning products in the market or redesign a winning product on the market. So this is all the product redesign part. And third is about how to find the products. I think these are the three major scenarios in Axial. It's completely different from the other kind of platforms. The other platforms majorly they just looking for a product buy and sell, something like that. But the Axial actually can help them much more. And you know Alibaba.com, when we set up this business back in 1999. So Jack Ma's mission for Alibaba Group is to make it easy to do business anywhere. So I think what do we do today with Axio for solo entrepreneurs extend

23:14

this mission makes sense. What's been something that surprised you in building Axio and trying to envision and then bring to life an AI system for global trade?

25:09

I think first part is about the technology. So the problem we solve today for global trading is not like a kind of a B2C E commerce world. So in B2C E commerce world when you buy something so the price probably within a couple of dollars to a couple of hundreds US dollars. But when it comes to a kind of a B2B sourcing. Especially in the global trading scenarios, the questions to become is becoming much more expensive. And many of the times it's not quantitative task. It's a kind of qualitative task that decisions that you need to make. It's not easy. So I think that is for the kind of technical challenge perspective. The second part is about the business model challenge. So you know when you building kind of AI search meaning that you are not letting your users key in the keywords. Giving them kind of millions of results and letting them to click through. Actually you are understanding interpreting their kind of requirements come with a few results. It gives them kind of the better choice that may impact on the business model as well. So it's like the advertisement business model that it need to evolve. So we know that this model can bring more customer value and then for bring more business value. But still the kind of the business model need to evolve to kind of match with this new technology.

25:22

If you are speaking to an executive who's whatever the product may be. But we can say within commerce looking to build an AI product deploy at massive scale scale approaching what you deal with on a daily basis. Do you have a piece of advice or a couple things that come to mind that you would give them before starting out? Right.

26:57

The first I think the most important one is about the questions you're going to solve. So despite all kind of fancy terms about the technology. But still, whether your question is a real question or your question is a big enough question, I think that is the first one. And second is I can share some of the best practices that we experienced for the last two or three years. So we have kind of three layers of approach to kind of practice AI. The first layer is about AI native applications, which is axial, we talk a lot today. And in that kind of AI native applications, you can try anything that you need with a very quick speed and you can reiterate this product very quickly. The second part is about AI+alibaba.com which is, as I mentioned to you, which is the first business of Alibaba Group with years of 26 years history. And we need to add AI model or the AI value to Alibaba.com, which can expand in a larger scale. You can get more people to benefit from this amo both the buyers and suppliers. This is the second layer. The third layer is about AI insight. So within Alibaba.com's organization, every role has an AI KPI, from the user growth to the product design to the sales team to the technology team. And as you can mention, so every role in Alibaba.com organization, they have a kind of AI API for themselves. So everybody has a sense of urgency to improve, either copilot or improved by the AI technology.

27:18

The KPI is measuring use of AI or productivity or effectiveness of the AI tool itself.

29:10

I think different team or different roles have different KPIs. Like in sales team, smart about efficiency. Like in technology team, it's more about kind of throughput. So how many features that you can deliver in product team and in the user growth team, it's like how they can leverage AI to redesign the model, redesign their kind of daily basis work. I think the whole team can benefit from AI a lot. It's not only a single team or kind of single person.

29:17

Right? Right, absolutely. So if I can ask you as we wrap up here to look ahead five years, 10 years, somewhere in that timeframe, if that works, how is AI going to change just on a fundamental level, the way that we do business around the world, global business, what's going to change? And in particular, if there's something you think people might find surprising, we always like to end on a provocative note like that.

29:49

Okay, so it means how we define the success for AI when we talk about the AGI, something like that. I think the success of Define AI is whether or not we can add at least 10% of growth on the current GDP. For example, for global trading. Global trading today is more than US$30 trillion business. So if we can add 10% more to this business, it's going to be US$3 trillion, kind of add on value to the whole system. And we believe that with the help of AI, more and more people can anticipate can embrace this kind of global supply chain and can compete globally. That will dramatically kind of increase the value and as we said in the beginning, to make it easy to do business anywhere for everybody.

30:14

Cool. For people who would like to know more about Alibaba.com, obviously the website right there, but more about any aspects of what we talked about beyond the website, social media, perhaps there's a research blog, other assets. Where should listeners go? To learn more about the work that you and your colleagues and team are

31:11

doing at Alibaba, Visit the website xu.com or alibaba.com I think is the first go to place. And also we have a kind of a podcast we call the B2B breakthrough in us.

31:33

Oh, fantastic.

31:46

Yeah, we have a lot of customer use cases, a lot of kind of best practices that you can learn. There's a lot of fun there.

31:47

Great. And the name again, sorry. B2B breakthrough.

31:54

B2B breakthrough podcast.

31:57

Perfect. Thank you so much. Again, thank you for taking the time while you're traveling and I know you're preparing for an event. Best of luck with that. And we look forward to really, you know, living in a world that's powered by global trade and so making good use of your technologies every day and define the work that you're doing.

31:59

Thank you very much. Sam.

32:18