Value is larger for traditional companies. Once you start using agents for digital workers, digital CEO, digital twins, and for the entire company, you gain efficiency. The sooner they do that, the faster they build up a lead and a moat. We train a large foundation model to perform many tasks, and giving this access from just a conversational interface makes AI more accessible, even to my own team. By empowering other people to get to our level in AI drug discovery, we're making our job easier. The whole definition of AI needs to change. They think of AI as a tool. That's already wrong. AI is a mindset. It's a way to reinvent your company.

Artificial intelligence has become a core part in our mission. Today we're going to talk about how AI reshaped our industry and human longevities. It's been a great pleasure to have two trailblazers sitting here with us: Dr. Kai-Fu Lee, a global AI leader, the author of AI Superpowers and AI 2041, the founder of 01.AI, and the CEO of Sinovation Ventures. Great to have you.

Thank you.

And Dr. Alex Zhavoronkov, the founder and CEO of Insilico Medicine. He led the company to become the real leaders in AI drug discovery and helped get the company recently listed in the Hong Kong market. Great to have you, Alex, and thank you for having us in the Silicon Lab today in Shanghai.

Very happy to host you here. We just listed under the ticker symbol 3696.

Insilico has been at the forefront of AI drug discovery, and I know Kai-Fu being a longtime investor and partner in that journey. Could you share more about how you first noticed Alex and Insilico's work?

I went to a conference in Milan about 10 years ago or so. I noticed one thing: everyone said, "Have you met Alex?" And I said, "So who's Alex?" He said, "He's amazing young guy, still young today, but younger then, with long hair and bangs." They said, "You got to meet him. He's doing amazing company." So I figured anyone who can influence so many movers and shakers must be great. So I did meet with Alex and I was really impressed with his knowledge of both AI and biology. At the time there was really no such thing as AI drug discovery. So I thought that was a clear area that Sinovation Ventures was interested in. And also in follow-up rounds, we kept putting more and more money in until you're worth so much money, we couldn't afford it anymore.

Okay, very interesting story. Alex, what is your initial impression of Kai-Fu when we met?

He's the world celebrity in AI, and everybody knew about his work with Steve Jobs at Apple, and at Microsoft, and at Google. He's one of the most established AI scientists who actually turned AI into commercial success as well. We were first introduced by Peter Diamandis, one of our very early investors and supporters, and one of the people who I deeply admire. He actually said, "Look Alex, there is one person you need to meet in AI who you haven't met before, and that's Dr. Kai-Fu Lee." This sounds familiar. He has the golden finger. After Sinovation invested, we really realized the superpower that Dr. Lee wields. He invited us to present at one of his AGM meetings, and that presentation was viewed by more than 10 million people.

Wow. So when I presented at the event, that gave us the superpower to hire amazing people. We really enjoy the privilege of the spotlight in AI as well.

We would like to spend a little bit more time on how we see AI today, especially after decades, we see multiple waves of AI enthusiasm. Maybe start from Kai-Fu since you're carving out the entire new model for the China AI landscape. How do you feel about the journey? Do you enjoy it and how you compare with your private chapter?

At Sinovation, we invested in 13 AI unicorns. We invested very early stage, they became unicorns, and Insilico was perhaps one of the ones we're most proud of, and there are others. They're all doing okay to great somewhere in that range. But I kind of felt not completely satisfied. I was more on the sidelines, a bystander helping, which is satisfying in its own way, but I was always thinking, "Hey, when the next big wave comes, I want to jump in and build something myself." So, when large language models started taking off, I realized this was the opportunity I was waiting for. The idea is we saw agents and we saw the enterprise market, and we saw that convergence to cause companies to have not just workers but AI workers, and to have the CEO and the C-suite not just have themselves, but have a digital twin. And that is what's going to take all companies, including traditional ones, to the next level. That is so large and such a big opportunity. So we spent the last three years really moving in that direction.

The major difference is that many agent companies build B2B applications, but they're typically targeting peripheral tasks. If my life goal was to bring AI to the main stage faster than ever, I want to apply them to core needs of companies, not just to replace people, save money, but also to rapidly and dramatically increase the revenue, the profit, time to market, etc. So core functions that create different outcomes for companies, those are the kinds of agents we like to build.

Alex, I think maybe we delve into the drug discovery industry. Where do you see the most common gaps that outsiders from the healthcare industry will still have when we're talking about applying AI technology into this field?

When it comes to AI for drug discovery, we need to look at productivity metrics. Think of it as a car business. You are Ferrari and you're competing with Lamborghini and Bugatti. When it's early, you are looking from zero to 100. How quickly it can go? How much does it cost? So just like with the car philosophy, we try to go from zero to PCC (preclinical candidate) one stage before human clinical trials. How efficacious and high quality is your molecule and how novel it is. There are very specific criteria after preclinical candidate. Everything is very regulated. You cannot move faster than the speed of traffic. So what we need to do is cut this, and we go from zero to developmental candidate, usually around 13 months. However, we try to increase the level of novelty.

If you were to ask SpaceX, "What's your secret sauce? Is it AI? Is it the engine?" The reality is the ability to launch. So we started building the capability of consistently delivering those PCCs that later do not fail. 28 PCCs and 12 reached IND, so reached clinic. That is a big deal.

I cannot wait to utilize Dr. Kai-Fu's 01 technology, the 'boss AI', to actually make myself even more efficient. I want to be the agent of AI where AI controls even me.

I know both of you must have already seen the news: weeks ago there was a guy in Sydney, an engineer working with university labs, just using ChatGPT and FORS, and then they developed a personalized mRNA vaccine on his dog. This attracted a lot of attention from the public.

It's a way to attract attention.

That really drew a lot of attention. The reason why we have clinical trials and why we have the FDA, for example, is a very good reason, because without clinical trials, we don't know if the story is true or not.

I'm extremely happy that this story attracted so much attention, because maybe some young kids, maybe they will feel empowered to use generative AI to identify treatments. However, very often those sensational news give you very false confidence that technology works and there is a miracle. The reality is that we need to have a lot of hard work, a lot of clinical trials, in order to convince the regulators, but also ourselves, that the drug or any technology works. That is why those stories are great. However, we need to be very cautious.

Coming back to being very cautious, Kai-Fu, you feel like it's not only about healthcare, it's about people may be over-optimistic on large language models in a certain way. What do you think is the critical missing link if we talk about allowing advanced AI to cross the chasm into mainstream industrial adoption? Where do you see the missing link there?

I think industrywide, there is a tremendous belief now that AI adds productivity. So what is really stopping that? Certain domains are going so incredibly fast, like AI for software engineering, for programming. That turns out to be the first wave in which in my company, 01.AI, AI engineers are delivering 10 times the productivity they were before. It's not leading to layoffs and job replacement. So that kind of acceleration sets an example for other people to be inspired and to believe.

Now, the first set of industries that will adopt it are the ones that will solve big problems and create tremendous financial value, and drug discovery is one of them. Although as Alex said, that is not for anybody to do it on a hobbyist basis, but really you need the discipline and the validation of results. But there are many other industries that are also very fast. Anything that's digital is very fast to do. Engineering is digital. What's magical about software engineering is that it uses a language that seems very difficult for people, programming that seems tough. But it's so easy when you have natural language that could be translated to something very structured. That's very easy.

There are other professions like that. If you were looking for investments, law is another one. Reading a legal document is really hard for me, but having AI translate it to legal is very easy. So going from natural language unstructured input and translating to structure, that's the easiest.

Then there are ones that are risk-free, that are also very easy. Risk-free means, you know, when you have AI spending money from your bank account, that's worrisome. If AI controlling your health or your risks, that's very dangerous. But having AI think about movies, science fiction, entertainment, if it's too imaginative, the story is more interesting. Then there are other digital domains: financial payment, investment, insurance, banking. Other than regulation, those are also the next wave.

The toughest ones are the brick-and-mortar companies that do manufacturing, retail, mining, agriculture. We at 01.AI feel they're the ones who need the most help, because if you go to a software house, internet company, or drug discovery company, they already love and embrace AI. But we don't want any industry to fall behind. In fact, the value is larger for traditional companies, because once you start using agents for digital workers, digital CEO, digital twins, and for the entire company, you gain efficiency. Your knowledge of the traditional company becomes a moat, because that knowledge means you have the data in the industry, know-how of the industry feeding into the agents you create. Your agents are becoming smarter all the time. But it takes a while to create that data flywheel. So as we are creating enterprise software tools, we want the people who need it the most. They may not know it, but once you get it going, the flywheel after a couple of months starts spinning. I think then it causes society to move forward, because I view my dream and vision is to move the world to recognize and use AI as quickly as possible. So I actually want to look at the harder problem, brick-and-mortar companies building special models for them, special agents for them, and they're now seeing tremendous return on investment.

Alex, I know Insilico is now partnering with Liquid AI, and you also launched a new product called Science MMAI Gym. How does the gym work?

MMAI Gym stands for multi-modal AI gym for science. We realized that at Insilico, we don't want to compete in the foundation model space, developing our own foundation models. It's pretty much like investing money into a microchip instead of making this microchip semiconductor work for you. So we realized that we want to focus on the very protected space where we are currently number one in the world, in our opinion, which is drug discovery and many other AI for science tasks, where we can take an open-source model or a closed-source model, if the company is willing to collaborate, to turn the model into a super-intelligent system for many AI for science tasks.

At Insilico, over the years, we created a set of benchmarks of different tests that are very proprietary, that only us know the answer to the question that we're asking the AI model to answer. Now we can utilize these smaller models and our tests in order to first test the foundation models that other people create to see how good they are at drug discovery tasks. Usually they're very bad. Then we utilize those smaller models as teachers to teach the large model very specific skills. It's kind of like if you want to create a super champion in MMA (mixed martial arts), you send them to Shaolin Monastery, and they train with the best Shaolin monks, kung fu. Then you send them to Brazil to train with Gracie family in Jiu-Jitsu. Then you send them to Japan to learn Judo. Then you go to UFC. So that's the MMAI Gym. Take the model into the gym, turn it into a very capable tool, and then distribute this intelligence to everybody who wants to use it.

One of the foundation model developers was Liquid AI, Liquid Networks. They specialize in non-transformer architectures that provide you with foundation models that have very dense intelligence, and small models can perform amazingly well in very complicated tasks. So they can be deployed on cell phones, on watches, everywhere. We showed that with a very small model, 2.6 billion parameters, we can perform extremely well at state-of-the-art or above state-of-the-art in several drug discovery tasks.

Now what we are doing is partnering with 01.AI, which is extremely competent in post-training and business workflows. They know how to make AI work for you. If we train a large foundation model to perform many tasks, put it into a prompt window, and give this access from just a conversational interface, it makes AI more accessible to more people, even to my own team. So by empowering other people to get to our level in AI drug discovery, we're making our job easier, because 01.AI is so good at making things useful and practical, reliable and trustworthy.

Please allow me to invest in you now. I understand the concept of the gym. And MMA as well.

We've been a follower from the very beginning about this technology, which I would call breakthrough. It's not widely known because it's a very deep subject, but users probably don't even know they can take benefit from it, because this is about general ways to make smart small to large to medium models for specific AI for science experiments, starting with biology, but going way beyond that. Drug discovery, material science, and others can benefit.

So 01.AI works with many industries to help them train really powerful models, sometimes quite small, but targeted for that industry. We've done it for supply chain, for agriculture, for retail. But usually we hold the hand of these traditional companies because they don't have the skill set, and it's been a challenging journey. Here we have Alex and Insilico which has already done most of the hard lifting. So we're very happy and honored to be partnering on this, because whatever expertise we can add, and industry and sales knowledge we can have, we want to leave Insilico to be primarily focused on curing diseases for the world. If we can add a little bit of help to extend the reach of what these great scientists have accomplished beyond drug discovery to other areas, we're more than happy to take what expertise we have and do this together.

A follow-up question: in the future, do you feel like enterprise will rely on ever stronger general large language models, or do you think general models will be trained more deeply for vertical intelligence via platforms like MMAI Gym?

I think both paths will proceed. For the short to medium term, they'll both have different successes. Obviously if you like a large closed model and you feel it solves most of your problem, then you can just be completely committed to that one single supplier and platform. But that has a number of challenges for some companies. One is it gives you a good but not perfect solution because it doesn't quite get retrained and refit to your industry. You may want a smaller model that knows your industry, your company. People generally think about these applications as chatbots, but they're not. Inventing a new drug is not a chatbot. Changing the supply chain for mining or agriculture is not a chatbot. So I think we have to be more imaginative about uses.

Once you get specific to an industry or a company, you'll have specific data, special applications, special ways to measure the success of your model. Then large isn't the main thing. The model you fine-tune and train and perfect to your application cannot be used for general applications. So, I think that's the likely outcome: for companies that want a really tailored solution to a non-chatbot real business problem or a science problem, they'll want to find a way to do their own. Using a very large language model hosted on the cloud is impractical. First, because it's expensive, and also you don't have the model, you can't see the model. It's all hidden from you. So, how do you tune it?

That's why this open-source approach is a much better approach to fit, and also a lot of these businesses care about not leaking their data to the world. There's intellectual property, confidentiality, licensing rights. So having a local solution with on-premise compute, using an open-source model with your local data never leaving your corporate walls, becomes a very good option not only for the gym or for science, but also for banking, financial, and many financial.

Like we all like this for pharmaceutical.

Exactly. Because your data becomes your most.

Open claw.

It's open cloud people are so fascinating about it. Everybody's talking about that. It's already running half of my life.

In Insilico, you also have the Panda Cloud. Does that mean a scientist just sets a goal and the Panda Cloud can do the rest of the work? Is that how it works?

We have a system at Insilico called Panda Onyx. It can help you identify protein targets in a very wide range of diseases, including aging. This system is used by thousands of scientists around the world. It utilizes multiple AI platforms that we've developed over the years. So Panda Cloud applies the open-cloud concept to target discovery. You basically give it a task, like "go discover targets with these properties." And forget about it. It will go and create workflows automatically, try to identify new ways to discover targets, utilize all of the resources available to it, and all the skills in order to try to achieve that goal. It prioritizes them, creates reports for you, and basically anything that deals with deciphering the mechanism of disease, to understand why the disease happens and what's driving the disease in specific context, this tool can do for you.

Who do you sell it to? Is it for academic scientists or pharmaceutical companies? Mostly it's pharmaceutical companies, and some academics. Pharmaceutical companies usually rely on massive numbers of vendors who provide them with target hypothesis, target validation. Here they can deploy Panda Cloud, and the Panda Cloud system will just go around and do the work for them over time.

Is that replacing some of the vendors or managing vendors?

Let's call it complimentary, because we don't want to completely eliminate humans from the decision workflow. But I think in the future, it will eliminate much of the human decision-making, because human decision-making is often biased. If you want to come up with a framework where you can take more risk, where you can accept higher novelty, you need to have AI that increases your confidence consistently. And I think that Panda Cloud is one of those tools that will significantly increase your confidence by providing an extreme overview of everything available about the target, by putting it in front of you and convincing you that some novel target works better than something old and proven to work.

Speaking of that, I think it will be interesting talking about the mindset change, especially in 2026. Kai-Fu mentioned it's a year of enterprise multi-agent deployment. What mindset shift should enterprise or leaders have to have this intelligence deployed as a skill?

What's already happened is the enterprise leader, the CEO, is aware AI is here to stay. AI is a big deal. "I must embrace AI." I think that has been achieved. The second stage now, I think, is very important: for CEOs to start depending on AI for critical core tasks of the company. Core means what's the most important thing for the company, right? For a mining company, it's getting more tons out because they equate to revenue. For an agriculture company, it's revenue, but also margin and profit: how to make more money. For a product company, it's time to market, how to get more productivity out of the producers or software engineers in my case, in my company's case.

So when more and more CEOs shift their mindset from the chatbot, from the customer service, to this core function, that will make a huge amount of difference. I think the challenge is the traditional company. Typically the CEO doesn't understand AI. The CIO understands a bit about AI applications, but they're not the ones tasked with changing the company. So I think either the CEO just needs to take ownership and say, "I drive this myself," or when they need help, they might need a Chief AI Officer, and they could use 01.AI for that purpose or another consulting company.

The sooner they do that, the faster they build up a lead and a moat for their company within their industry. The third thing I think is very important is the whole definition of AI needs to change. If they think of AI as a tool, that's already wrong. AI is a mindset. It's a way to reinvent your company.

So the first step is not only a tool, but you're thinking from the fundamental what impacts your company. Every employee needs to rethink the way they work. I'll give you an example for myself. I no longer try to think original thoughts from zero. I pride myself that I am a relatively original thinker. But I would always delegate anytime I have an idea that's at ground zero. I always do deep research with my favorite AI tool and have it produce everything it can by covering every data it has, then have it summarize it for me. Then I start thinking. So I'm still an original thinker, but why should I be handicapped by doing my own research when AI can do it for me? Later I can do my own research. That's changed the way I think as a person. I give that not as an example everyone should do that, but as an example that if a person has changed the way they think out the problem, isn't that the same as an organization that needs to rethink how it should organize and tackle the market opportunity?

At the very beginning we touched a little bit on longevity, and it's definitely a very interesting, attractive topic. Insilico also focuses on applying AI where it can matters to patient lives, shortening development cycles, improving outcomes, and expanding access. Alex, I know you have a bold, long vision theory saying with the help of AI, people can live up to 200. Tell us where we are today. How does it say against your roadmap?

When I started in longevity 26 years ago, I was actually more optimistic. At that time we also saw massive advances in technological progress because we've got mobile, internet, personal computer propagating, social networks. But now we are in 2026, and we still don't have a single drug that has demonstrated clinically, that's clinically proven, to reverse aging. There are multiple drugs that help you live a little bit longer by inducing certain behavioral changes and certain biological changes that could be achieved through lifestyle modifications, better diagnostics, etc. But in terms of clinical proof that something works in aging, we're actually lacking that at this point in time.

However, the past five years made me much more optimistic. At Insilico, we've managed to streamline drug discovery to the level where we can consistently launch from zero to developmental candidate drugs that target aging and age-related diseases from the angle of metabolic diseases, muscle wasting, inflammation, and fibrosis. We hope that once those drugs reach the clinic, we would be able to repurpose them into aging to properly test whether they reverse aging to some extent. Now we see some early evidence, very preliminary signs that some of those drugs have very promising efficacy in age-related diseases and potentially aging itself, using biomarker studies. Those are early days. So I would say five years ago I kind of lost optimism, but now I'm back on track. I am again very optimistic. I think that 200 may not be the limit. The results of our work are going to lay the foundation for many generations of AI drug discovery scientists to come, who are going to try riskier, financially riskier bets, where people would be able to bet on more novel mechanisms, try more approaches in the clinic, where we take a drug for a disease and then repurpose it into more diseases, more age-related diseases and biological processes, and then potentially aging itself.

Five years ago, Kai-Fu, you have the book AI 2041, right? In that book you pictured a very visionary picture, a world where AI can decipher the very codes of life. When you look back at your prediction five years ago, what surprised you most? The speed of the AI integration into biology.

I think the book covered things like AI drug discovery, longevity, large language models, education impact for businesses, AI externalities, safety. Those were all correctly predicted. Probably the only questionable area is the speed, because it turned out AI progressed faster than my prediction. It's a factor. You just have to change one number in the whole book to make it correct. AI 2041 should have been named AI 2031. I missed by about 10 years.

Some suggestion or advice, maybe a book, maybe a mindset, maybe a tool, they probably should think as the priority in 2026.

Don't be afraid of AI. It's here to stay. It's a tool that you have to use, have to learn. It will make you much, much better. Amplify yourself. The earlier you learn, the more contributions you can make.

Secondly, for every domain, there will be ways to use AI that you should embrace, because it's not just using AI to do the old job better. It's really redoing, rethinking, doing things differently. Before, a software engineer is given a set of input/output, and they have to write code. That work is now being done by AI. But the great software engineer who thinks in terms of modules and integration and product and users, they can now command dozens of armies of AI software engineers to build this. So code is always needed. It will be built differently with human-AI symbiosis. The earlier you understand what parts of your profession will be done better by AI, the faster it is you will move to the areas where humans add value. At the end of the day, it's always going to be human-AI working together. It's not the case that there's one pie where if it's half done by AI, human gets half less pie, because the pie will get bigger. What's limiting the pie from getting bigger is human limitations. We think too slowly. Our input is way too slow. Our output is way too slow. So now with AI and humans working together, the pie will easily be much, much larger. In that growth of the pie, there will be a chance for everyone. So be optimistic, learn AI, think about how the symbiosis will happen in your domain, and go after being the best that can partner with AI.

Alex, some suggestions for biologists, chemists embracing the AI workflow. Any suggestion for them?

My suggestion is don't be afraid of it. Embrace it. Drug discovery will be one of the last professions to be eliminated. It just takes so much effort to validate the drug to a specific point. I actually think that brain-to-computer interfaces are more likely to disrupt us as humans and improve our human life beyond what's humanly imaginable today. But drug discovery will still be done a little bit slower than people think.

I personally fully agree with you. I think that might be, I won't say the last, but one of the last parts that can be totally replaced by AI. I don't think so.

I hope so. And I hope that I'm wrong. I really hope that I'm wrong.

That means that we definitely shorten the whole process, the development cycle, and it benefits the humans. For sure.

So when I become 200 years old, you're still going to be working in your field, right?