Baidu CEO Robin Li on AGI, Scaling Law, Open vs. Closed-Source, China's Market, and AI Super Apps
Hi, this is Tony! Welcome to this issue of Recode China AI, a newsletter for China’s trending AI news and papers.
At WAIC 2024, China’s largest AI event held from July 4th-6th in Shanghai, Baidu CEO and Co-founder Robin Li publicly shared his views on the generative AI landscape in a fireside chat interview with Yicai, a top Chinese business media, and Silicon Valley 101, a prominent Chinese tech podcast based in Silicon Valley.
Baidu has been investing in AI for over a decade. Its chatbot, ERNIE Bot, has garnered a total of 300 million users. Baidu is also one of the few companies worldwide offering an AI full-stack to developers, including chips, AI frameworks, models, and apps.
Robin Li discussed some of the hottest topics in AI, including GPT-5, AGI, Scaling Law, the debate between closed-source and open-source models, AI super apps, as well as some China-specific things like LLM price war and local use cases.
I found this interview particularly informative, as many of Li’s insights are useful to readers outside of China as well. For example:
Li believes that smaller models are in real market demand because people require very fast response times and low costs for practical use. Before I published this piece, OpenAI had just released GPT-4o Mini, a significantly smaller model with comparable performance.
An interesting analogy Li proposed regarding AI agents is that they resemble websites in the early Internet era.
Li is also a strong advocate for closed-source models over open-source models, and he explained why.
Unlike Elon Musk, who believes AGI will be realized in about 2-3 years, Li stated that it will take more than 10 years, possibly 20 years, or even longer.
You can find the full interview video below, along with a direct English translation of the interview transcript. (Disclaimer: I work for Baidu as part of its global communications team.)
1. Inevitable Price War: Burning Money Is Not the Essence
Yang Yudong, Editor-in-Chief at Yicai:
We have observed that the costs for calling APIs of many closed-source LLMs are getting lower, and there are many open-source models available. In such a commercial environment, we are concerned about Baidu’s involvement. Does the business model of charging for inference based on LLMs have a future? What will be the key factors in the competition for LLMs in the market? You mentioned speed and cost earlier, what are your thoughts?
(For more information about China’s LLM price war, check out my previous issue.)
Robin Li, Baidu Co-founder and CEO:
Honestly, we have been thinking and discussing this issue internally. I think a price war is almost inevitable. Having been in the Chinese internet industry for so long, we are very familiar with price wars. It has come even earlier than I expected, driving prices down to an almost unbelievably low level.
However, I don’t think this is necessarily a bad thing. When prices are low enough, or even free, it incentivizes more people to try developing various applications based on LLMs. This means that the penetration of LLMs across various industries will be faster.
When you offer it for free or at a very low cost, how can LLM companies make money? I think the ceiling for LLM technology is still very high. Today, there are many aspects of LLM capabilities that we are not satisfied with. We still need many excellent technical personnel, a lot of computing power, and a vast amount of data to train the next generation of LLMs. We might even need the next next generation, or even the next next next generation of LLMs.
So ultimately, I believe it will come down to who has the better technology. If your technology is superior and provides better value to customers, you will still be able to charge in the future.
The reason prices are so low today is that the current model capabilities are not at their best. When everyone’s capabilities are similar, the competition is about price—whoever has the lower price wins.
Over time, the market will return to rationality. For the same results, if my cost is lower than yours, you can’t win the price war against me. Over time, you will exit the market.
We experienced this process with Baidu Drive. In those years when everyone was competing in drive storage, I offered 10GB of free space today, 100GB tomorrow, 1TB the day after, and then unlimited space for free. When someone offers unlimited space for free, people will question how sustainable it is. It’s definitely unsustainable.
But for a period, maybe one or two years, everyone irrationally engages in price wars, and gradually, one by one, they exit. Why didn’t Baidu exit? Because I dared to fight. My technology and storage costs were low, so in the end, it was about who had better technology and higher efficiency.
Chen Qian (Co-founder of Silicon Valley 101):
You mentioned that the price war came earlier than you expected, and it came earlier than I expected as well. It’s been only a year and a half, and the price war has already started. How long do you think the price war will last? Some people might think that if the technology is not good and there’s no profit, they will exit. Eventually, the market winners will emerge.
Robin Li:
It’s hard to say. Some players are startups, but there are also large internet platform companies involved. Theoretically, burning money can last for a long time, but I think burning money is not the essence of the matter. The essence is still about who has better technology and higher efficiency. When your technology is good and your efficiency is high, you are not afraid to engage in a price war. So, it can last as long as it takes, and ultimately, it will be a process of survival of the fittest.
Chen Qian:
Do you think the Chinese market will end up being dominated by one winner, or will there be a few major players left after the price war, with some smaller models forming an ecosystem?
Robin Li:
I think this generative AI represents a major transformation of the entire IT technology stack. We generally think of the past IT technology stack as having three layers: the chip layer, the operating system layer, and the application or software layer. After the advent of generative AI, we believe the IT technology stack has become four layers: the chip layer, the framework layer (deep learning framework), the model layer, and the application layer. I think there will be at least 2-3 major players in each layer.
At the application layer, there might be millions, or even tens of millions of various applications, and super applications will gradually emerge. Based on the definition of “super,” there won’t be many, maybe three to five.
For the model layer, I think two or three might be enough, because ultimately, it’s about efficiency. If your efficiency is not high enough, people will eventually think it’s better to use someone else’s model and develop more applications on top of it.
So, there are opportunities at every layer, but the development patterns and required expertise vary.
Chen Qian:
I completely agree with your view. I previously asked Dario Amodei from Anthropic, and he also said that in the international market, there might be only four major models left in the end because the others might not be able to keep up, which aligns with your thinking.
2. Commercial Implementation: Enterprises’ AI Demand Is Much Greater Than the Internet Era
Yang Yudong:
You mentioned that the impact of LLMs on the business-to-business (B2B) market will surpass that of the internet. Why do you think the transformation of the enterprise market by LLMs is greater than the impact of the internet on the B2B market?
Robin Li:
If you think about it, the transformation of the business-to-consumer (B2C) market by the Internet is something we all deeply feel—it is very thorough and disruptive.
However, the transformation of the B2B market by the internet, I think, is just okay. Although we can’t say there is no impact, the “internet+” in the end might just be a process of informatization or digitization. The technology used is relatively simple, and the gains are not that significant.
But LLMs are different. Various AI applications, such as novel creation, reading novels can be considered a B2C product, but novel creation can be seen as a B2B product. Legal assistance helps lawyers improve efficiency (another B2B).
We have encountered similar demands in industries such as energy, power, and manufacturing. For example, in China’s electric vehicle industry, the in-car conversational systems of many cars are using the ERNIE LLM, and the usage is quite high. For Baidu, this is a B2B application. We do not directly provide it to users; it is integrated by OEMs and car manufacturers before being provided to end consumers.
There are many such cases, and if we look at the call volume, it shows that our customers need these things. The B2B market is generating value through the LLM and AI-native applications.
Yang Yudong:
Recently, I’ve been visiting the manufacturing industry more often. Some car companies directly call the ERNIE LLM (API). In the manufacturing industry, many entrepreneurs say that they believe AI will have significant value and be a revolutionary breakthrough for advanced manufacturing and intelligent manufacturing. However, they still don’t understand the relationship between general LLMs and their industry-specific vertical models in the future. They also wonder about the business chain between OEMs and platform companies like yours. This is a particular concern for the manufacturing industry, especially advanced manufacturing.
Robin Li:
This is indeed a critical question. The use of LLMs in various vertical scenarios has also gone through an exploration process. Initially, we thought of making the foundation model increasingly powerful. People call it AGI (Artificial General Intelligence), capable of performing well in any scenario.
However, we found that this is not so easy. Each scenario has its own characteristics and different requirements. In some cases, it is acceptable for the LLM to take two minutes to provide a result as long as the result is accurate and comprehensive. In other scenarios, if the result is not given within a second, the user will give up. So, these two scenarios have different requirements for LLMs.
Even today, the most powerful models cannot react quickly and accurately at the same time. Therefore, when application scenarios require quick responses, we need smaller models. Since these smaller models do not have the general capabilities of LLMs, they need to be fine-tuned in vertical scenarios using SFT (Supervised Fine-Tuning) with industry-specific data to achieve results comparable to LLMs in those scenarios.
We have seen many such scenarios. After launching ERNIE 4.0 last October, our primary focus has been on pruning the largest models into various sizes, such as billion-scale, tens of billions, and hundreds of billions of parameters.
For example, one model might excel at role-playing, another at logical reasoning, and so on. We develop various models based on the different requirements of customers in different usage scenarios.
Most of these models have faster speeds and lower inference costs than ERNIE 4.0. Therefore, they are more preferred by users.
Even today, the most frequently called models, among the 500 million calls I mentioned earlier, are not the most powerful models. Users find the most powerful models too slow and costly.
However, over time, costs will decrease, and response speeds will increase. For example, we recently launched the Turbo version of ERNIE 4.0, which offers improved performance but, more importantly, much faster response times and lower costs. Why?
Because we found that the real market demand requires very fast response times and very low costs for practical use. As for how powerful the next generation of LLMs will be and in which scenarios they can be used, that remains to be explored. But if you look at the current market demand, smaller models have greater market demand.
Yang Yudong:
Could you share some successful commercial cases or business models and product forms of LLMs in the B2B market applications?
Robin Li:
In addition to novel creation, there are cases like digital human live streaming in e-commerce. LLMs generate live-streaming scripts. For instance, when selling a nutritional product, it’s challenging for a human to remember details like the year a certain university professor conducted experiments on it and the results. However, these data points are crucial to explain the product’s benefits. A digital human can express these details generated by an LLM, often achieving better results than real human live streams. Such examples can be found in almost any industry.
Chen Qian:
How do enterprises view the costs of calling AI services? When helping companies calculate costs, if they spend 100 yuan on various AI services, they need to make 500 yuan to be profitable. Are enterprises willing to pay for AI services now, and what is their attitude when you communicate with them?
Robin Li:
This is a great question. In a market economy, enterprises are very rational, especially small and medium-sized ones. They calculate costs very precisely. If something can help them reduce costs and increase efficiency, they will use it. If not, no matter how much you hype it, they won’t use it.
The market will tell you whether LLMs are useful. The rapid increase in call volume indicates that it does provide cost reduction and efficiency improvements for businesses.
For example, in the recruitment scenario, a LLM can understand the requirements for specific skills and match the right candidates much more efficiently than a human HR screening resumes and conducting interviews.
Previously, HR would screen hundreds of resumes and interview candidates to shortlist ten people, which is highly inefficient. With an LLM, this efficiency can be significantly improved. In such scenarios, the inference cost of the model is almost negligible.
Especially in China, where the price war for LLMs is very fierce, lightweight models from Baidu are free, including the computing power. Enterprises just need to use it without worrying about additional costs.
Chen Qian:
So, the inference is very cheap as well?
Robin Li:
The inference is free.
Chen Qian:
I have a curious question. During the SaaS era and the internet era, SaaS didn’t really take off in China. Why do you think AI as a service can succeed in China during the AI era in B2B?
Robin Li:
In the U.S., each business typically has its own e-commerce website where users can shop. In China, however, people usually shop on e-commerce platforms, which handle the B2B capabilities for the merchants, so they don’t need to develop these capabilities themselves. In the U.S., if you want to build your own website, you need to be able to sell products on it, have various features, and understand your users, which requires significant investment.
So, in a sense, it’s not that China doesn’t have this demand; it’s just that these demands are met by larger platforms.
However, in the AI era, I think the situation has changed. As I mentioned earlier, AI doesn’t represent a completely new business model that no one has seen before. It initially enhances existing business models. When existing business models are enhanced and become more efficient, they can better resist the encroachment of internet platform companies on their business. Therefore, I believe that the demand for AI in the B2B sector in China will be much greater than the demand for internet-based B2B services.
3. Open Source vs. Closed Source: Choosing the Commercial Path for LLMs
Chen Qian:
There is intense debate globally over open source versus closed source. For example, Elon Musk criticizes OpenAI for being closed and suggests renaming it to CloseAI. We see that Musk’s xAI and Meta’s Llama are open source. Why do you insist on closed source?
Robin Li:
I think open source is a kind of “IQ tax” (a Chinese Internet slang meaning someone who has bad judgment or makes bad decisions about their purchases, paying a high price for a poor product). Think about it carefully. Why do we develop LLMs? It’s because they can have applications that, in various scenarios, improve efficiency, reduce costs, and achieve what was previously impossible.
When you rationally consider what value a LLM can bring at what cost, you will realize that you should always choose a closed-source model.
Today, closed-source models like ChatGPT, ERNIE Bot, and various others are generally more powerful than open-source models. Their inference costs are also lower.
Especially in China, where we even provide computing power for free, if you want to use an open-source model, you must do SFT (Supervised Fine-Tuning) based on your scenario, possibly perform safety alignment to avoid risks, prepare various data, iterate multiple times, and buy computing power to run it. Since it is your unique model, you cannot share computing power with other applications. You must prepare the maximum amount of computing power needed for peak usage, which is very, very uneconomical.
In contrast, with a closed-source model, you can share computing power with others. Under the same parameters, closed-source models are generally more powerful than open-source ones. Given the same capabilities, closed-source models might have smaller parameter scales, leading to lower inference costs. So, in the long run, the usage of closed-source models will far exceed that of open-source models.
Of course, I am not saying that open-source models have no value. From an academic research perspective, if you are a scientist wanting to study the principles of large models—since the industry has always complained about the lack of interpretability of LLMs, like how to explain why they have such capabilities—or if you are a student wanting to learn and practice with these models, then open-source models do have their value.
But if you are a business seeking value-added benefits from LLMs, once you do the math, you’ll find that open-source models have no chance.
Chen Qian:
Baidu uses a public cloud approach with closed-source LLMs, along with a combination of closed-source and open-source models on the public cloud for customers to call. This is similar to the approach taken by Amazon Web Services and Microsoft. What considerations does Baidu have in adopting this strategy for enterprise customers?
Robin Li:
For B2B customers, the most cost-effective choice is essential. On one hand, the model must provide value for their applications, and on the other hand, the cost must be low enough.
Often, if something is useful but not cost-effective, customers will abandon it. This is why I mentioned that open-source models cannot compete with closed-source ones. Rationally considering the benefits and costs, customers will always choose closed-source models.
Closed-source models are not just a single model; they consist of a series of models. Depending on your usage scenario, you can balance the required performance, inference speed, and cost. My models have many variants, allowing users to choose based on their needs.
These smaller models are trimmed from the largest and most powerful models. This is an advantage that open-source models do not have. When the most advanced model is closed-source, I can trim it down to create smaller models that still outperform similarly sized open-source models.
We also see that larger open-source models are even less reasonable. As the scale increases, inference speed slows down. If you don’t care about speed, you can use the most powerful closed-source model, which is certainly better.
Moreover, larger parameter scales mean greater computing power consumption, requiring more servers to support the LLM, which is highly uneconomical. The higher the level, the more shared computing power is needed to balance the costs.
So, I believe that unless you are purely discussing theoretical concepts, for practical applications, almost no one will choose open-source models over closed-source ones.
4. AI Super Apps: Should We Chase AI Apps with 1 Billion Daily Active Users?
Yang Yudong:
The hype initiated by ChatGPT has been going on for over a year. You have previously expressed the idea of waiting for the emergence of super applications. Domestically, LLM products targeting the B2C market seem to be mainly in the form of search box Q&A. What do you think? Is there a possibility for differentiated competition? What good products might emerge?
Robin Li:
To be precise, I am not necessarily waiting for a super application to appear. I believe that on the foundation of LLMs, millions of various applications should be able to emerge.
Some of these applications might be in small domains that are not easily thought of but solve problems in those areas much better than before. There might also be applications with a vast user base and long user engagement, similar to super applications in the mobile internet era.
To be honest, we haven’t yet seen AI-native applications that can rival super applications from the mobile internet era. However, we have seen more and more applications, especially in the B2B scenario, using LLMs to improve performance, efficiency, generate more revenue, or save costs significantly.
Especially this year, we have observed AI application scenarios emerging across various fields and industries. These applications have resulted in significant labor cost savings and efficiency improvements. This might not be as exciting for investors and entrepreneurs because they often look for the next big thing, an application that people have never seen before that becomes a DAU (Daily Active Users) of 1 billion super app. This is important, and it will appear in time, but the current significant aspect is the application of LLMs across different fields and scenarios.
This is evident from the daily call volume of Baidu’s ERNIE Bot. In April, we announced that ERNIE had 200 million daily calls. Recently, we announced that ERNIE’s call volume reached 500 million daily calls.
This doubling in two months indicates its value in applications because no one would invest in these API calls if they didn’t generate value. This growth rate is quite exciting. The impact of LLMs on improving efficiency and reducing costs across various industries is already very evident.
Yang Yudong:
You mostly talked about vertical industries. Baidu has strong B2C DNA. As a regular consumer and user, what good scenarios can we expect for B2C market users? Including on the edge, even on mobile apps, where the hardware might directly use AI capabilities without going through an app. What are your thoughts?
Robin Li:
I think there are two categories. One is what people are more concerned about, applications that never existed before, like ChatGPT. We call it a chatbot.
Every domestic LLM company will launch an app or website for chatbots. Its role is quite clear. If you have a specific question, it can give you a good answer, and its accuracy is improving. Many people gradually rely on such chatbots.
For existing B2C market applications, the information gain is significant. In April, we announced that 11% of Baidu search results are now generated by AI, and this proportion is continuously increasing.
In other words, existing common applications are increasingly being transformed by LLMs and generative AI.
Another example is Baidu Wenku. It used to be a place to find ready-made documents. For instance, a middle school teacher preparing lessons might look for lesson plans from the best teachers. Today, Baidu Wenku, transformed by LLMs, has become more of a generative AI application. Whether you need a PowerPoint, a research paper format, or even a comic book, it can generate it based on your requirements.
Moreover, this product has a broad user base and is paid. This year, it has around 26 million paying users. By the standards of super applications, it may not qualify, but looking at its actual value, having so many people willing to pay for it is quite impressive.
These products existed before but have been completely transformed by LLMs. People’s understanding of them is also evolving and deepening.
Yang Yudong:
Because of the improved capabilities, people are more willing to pay for it, unlike before when it was just a free search.
Robin Li:
Yes.
Chen Qian:
I completely agree with your recent emphasis on developing AI-native applications. This gives LLMs real meaning. At the same time, I have a small confusion. For example, GPT-4 was launched eight months ago, and everyone thought that the AI application explosion was imminent. However, eight months later, we still haven’t seen that explosion. Some applications are also disappointing.
Recently, Perplexity, an AI search application, has gained popularity. Its annual recurring revenue (ARR) is $20 million, and it has 20 million users, but it hasn’t reached the level of a Super app or a killer app. So, my question or confusion is, if the capabilities of GPT-4-based models aren’t leading to an application explosion, does this mean it’s not yet time to roll out applications?
Robin Li:
You mentioned GPT and Perplexity, which fall into the category of applications that never existed before, from 0 to 1.
There are no super applications at this stage. Even ChatGPT hasn’t reached over 100 million DAUs, so it can’t be considered a true super application. However, the transformation of existing products is significant both in China and the US.
For example, in the U.S., Microsoft’s Copilot has gained many paying users. In the B2B sector, Palantir and Snowflake have seen real revenue growth from being enhanced by LLMs and generative AI.
From this perspective, the application of LLMs is gradually emerging. Its impact on existing industries is more apparent and creates more value than from 0-to-1 disruptions.
Satya (CEO of Microsoft) once said that the internet was like AutoPilot, where results were automatically generated. Now, generative AI is like Copilot, a process of human and machine co-creation.
This process might not seem as sexy at first, but its impact on improving work efficiency, reducing costs, and opening up new possibilities is greater than the impact of 0-to-1 applications.
If we only focus on from 0-to-1, we might see a few super apps benefiting specific companies. However, today, almost every company in every industry benefits from LLMs. This impact on society and humanity is undoubtedly greater.
However, people might feel that this is something they have seen before, so there’s no novelty. Or these applications are more likely to emerge in productivity scenarios with an audience, or single application users, that don’t reach hundreds of millions or billions. Especially on the consumer side, the impact isn’t as concentrated at the public level. This is why everyone is always looking for a Super app now.
Yang Yudong:
Your question challenges Robin’s point about developing applications. After listening to Robin, I understand that his concept of “super” applications might differ from the traditional understanding. He sees it as numerous applications emerging rather than a single dominant one. In vertical fields, it has significant acceleration and enhancement.
Chen Qian:
Or maybe the definition of “super” is different in the AI era compared to the internet era.
Robin Li:
Yes, it’s different. In the internet era, it was a single application from 0 to 1 or to 100. Today, super applications enhance existing scenarios. This enhancement is still in its early stages, but it will gradually change existing product forms.
For example, in novel creation and online literature, which is very popular in China, writers previously relied on their own abilities and imagination to update content continuously. This was very inefficient. With LLMs, writers can generate content based on their ideas. The more detailed the input, the richer the output. If the tone is too gentle, it can be adjusted to be more forceful.
From a user’s perspective, they are still reading an online novel, but the production cost, efficiency, content richness, readability, and quality are entirely different from before. Similar examples can be found in almost any industry.
5. AI Agents: The Threshold Must Be Low Enough
Yang Yudong:
We have talked about developing applications. The next keyword is agents. You’ve mentioned several times that you are optimistic about agents in the AI era, but they haven’t had a significant explosion. Why do you think agents are the future trend?
Robin Li:
I think agents are currently exploding, but the base is still small, so the impact isn’t strongly felt. However, all major LLM companies are working on agents, and it is generally agreed upon by industry leaders.
Agents represent the future because foundation models need applications to show their value. An agent is a nearly universal AI-based application.
Based on your scenario, you can set a role without needing to program. You just need to explain what you want to do clearly. Sometimes, you need to connect your private knowledge base or workflow. It becomes a very useful and different thing from the foundation model.
Today, most AI-native applications can be created using agents, and the results are pretty good.
Because the threshold is low enough, you might not even need to program to create a decent agent. This means that more and more people can create agents they want.
It is somewhat like the mid-1990s when websites were emerging on the internet. You could create a very complex website, like Yahoo, which was very impressive at the time. However, even a college student could make their own homepage, linking their favorite websites, such as Java learning resources. It was simple.
Because making websites was easy, millions of websites emerged from the mid to late 1990s. This massive wave eventually led to some exceptional websites like Google and Facebook, which appeared years later. But in the early stages, you might have thought, “These websites are messy, and a college student can create one. What’s the value in that?”
However, when the threshold is low enough, it allows more people to participate and use their creativity. You never know which path might lead to a Super app. This is why I believe agents represent the future, and I am particularly optimistic about this emerging industry.
Chen Qian:
The concept of AI Agents is very important, including in Silicon Valley in the U.S. Within the industry, there are still different discussions on the definition of AI Agents. Some people say that GPTs are also a type of agent, while others argue that only more advanced agents that can call different tools and act as virtual world robots qualify as intelligent agents. What is your definition of an agent?
Robin Li:
First and foremost, I think the threshold should be low enough that even a beginner, like a college freshman, can easily create an agent.
Of course, there can be various interesting functionalities on top of this, such as using tools, reflection, long-term memory, and so on, which will gradually be added.
This is similar to the development of websites in the 1990s. Initially, websites were very simple. Later, I could use Java to add some dynamic features to a website. Then, I could add cookies to remember what you did on the previous page when moving to the next one. These features were not available in the early stages of website development. But as more people created websites, the technology advanced to meet their needs.
Agents are the same. It doesn’t have to use the most advanced capabilities to be called an AI Agent. I believe we should lower the threshold enough so that everyone feels they can create an AI Agent. Over time, as new problems arise and solutions are found, the most advanced technologies will be incorporated.
To be honest, the capabilities used by AI agents today are still very basic. In the future, there will be agent capabilities that we can’t even imagine today.
However, the emergence of these capabilities depends on millions of developers creating various applications. During their usage, new demands will arise. The process of solving these demands will lead to innovation and the evolution of AI Agents. This is something I am very much looking forward to.
Chen Qian:
Do you have any interesting cases of AI agents from Baidu to share?
Robin Li:
There are many. For instance, during the college entrance exam season in China, a significant event, both students and parents are very concerned. Traditionally, LLMs were used to write essays for the exam, but this wasn’t practical for students during the exam.
You can’t take a LLM to the college entrance exam, right? But after the exam, you need to estimate your score, choose your college, and select your major. Some people care about how many people share a dorm room, others care if the school has a swimming pool, or which major is better for their future development. Everyone’s situation and questions are different. In such cases, no all-knowing advisor can tell you which school or major is best for you, but an AI Agent can do this.
In the past, people would search on Baidu or elsewhere for existing information. There was no content specifically produced for that person, at that time, in that particular situation.
Today, AI, especially AI agents, can do this. You can tell me your situation, and you don’t have to explain it in one sentence; you can use ten sentences, and I will remember everything. There are many such examples.
Chen Qian:
And it will give you highly personalized plans, like I saw your collaboration with the Singapore Tourism Board.
Robin Li:
Yes. Your spending level, how much time you have, your preferences, what you like to eat, what you don’t like to eat—everyone is different. It can generate the answers you need completely based on your situation.
6. AGI and Scaling Law: A Belief Worth Pursuing Long-term
Yang Yudong:
There have been discussions about Scaling Law. Initially, everyone believed in it, but recently, different opinions have emerged. Some industries are focusing on faster and more efficient small models. Will Scaling Law be overturned soon?
Robin Li:
Scaling Law may still have several years of lifecycle, but various innovations will be added on top of it.
The reflective and evolutionary capabilities of agents, as mentioned earlier, are somewhat separate from the Scaling Law. They are developing along two different routes, but they are still based on Transformer LLMs.
In the next one or two years, we may see new technological innovations layered on top of this foundation, but we don’t know exactly what they will be. Everyone is exploring.
In other words, I don’t think the Scaling Law will be overturned in the short term, but many unimaginable innovations will be built upon it.
Yang Yudong:
In your opinion, what is the standard for achieving AGI (Artificial General Intelligence)? What paths can help us reach AGI faster?
Robin Li:
I think this question doesn’t have a standard answer in the industry. Previously, people thought that passing the Turing test would mean achieving AGI. In reality, LLMs have already passed the Turing test, but what people refer to as AGI today is not just about passing the Turing test.
So, what is AGI? In my opinion, AGI means that machines or AI can have the capabilities of humans in any scenario—general intelligence that is universally applicable. In any scenario, its capabilities should be on par with humans, which is a very high standard.
For instance, we have been working on autonomous driving for 11 years, and we still can’t say the technology is mature. It is still limited to specific scenarios, and if AI can’t handle errors, it won’t work.
Therefore, achieving true AGI will require many more years. Some in the industry say AGI could be realized in 2 or 5 years, but my judgment is that it will take more than 10 years, maybe 10, 20 years, or even longer.
We often hear people say that AGI is a belief. When you consider it a belief, it’s contradictory to think it can be achieved next year. If it is a belief, it is a goal worth pursuing in the long term. What does “long term” mean? If it can’t be achieved within 10 years, it can’t be considered a belief.
Chen Qian:
Too easy to achieve can’t be called a belief.
Robin Li:
Exactly.
Chen Qian:
Everyone is currently waiting for GPT-5, but its release has been delayed. I’ve been hearing increasing concerns that even with 50-100 trillion parameters, the improvement might not be as significant as expected. This could undermine people’s faith in the Scaling Law, suggesting that we may not achieve AGI using this approach. Do you have any concerns about this?
Robin Li:
I’m not too worried about this. I think people should focus more on applications rather than on the foundation models. In a way, slowing down the iteration of foundation models isn’t necessarily a bad thing. It provides a relatively stable base for application developers to work with, making them more efficient. If the model is constantly being updated, developers would have to rewrite their code repeatedly, which is exhausting.
However, continuously fine-tuning and incrementally iterating on the existing foundation models based on market demand is ongoing. Whether it’s OpenAI’s continuous updates or our turbo models and smaller models, they are all evolving according to market needs.
In the long run, I firmly believe that the next generation of LLMs will be much more powerful than the current generation. When will they be released? I’m not in a hurry. I think we should closely observe the real market demand and iterate the next generation of models accordingly.
If we believe that AGI can’t be achieved within 10 years, the next generation of models is still far from AGI. They will have strengths and weaknesses. If their strengths do not align with market demand, they are not meaningful. If their weaknesses align with market demand, we will waste resources on unnecessary tasks.
That’s why I focus more on applications. I want to know what the market needs. For example, if I’m a car salesman, I want to know what sales techniques can persuade a customer to buy a car, not how to pass a high school math exam. When market demand is unclear, rushing to create a seemingly more powerful model might lead to detours and wasted resources.
Chen Qian:
I completely agree. Do you think that in addition to rolling out applications, there will be more focus on small and medium models, like the international Mistral models that are popular among developers? What plans does Baidu have for small and medium models and model distillation?
Robin Li:
We have observed that the real demand is mostly not for the largest models. Instead, there is a demand for smaller models, which means faster speeds and lower costs. For instance, if a project brings me an annual gain of 1 million yuan, but the largest model costs 1.2 million yuan, I won’t proceed with it.
So, our requirement for LLM companies is to reduce the cost to 800,000 yuan or even 80,000 yuan. We need to find a way to distill the most powerful models into smaller, more cost-effective ones that still meet the needs of specific scenarios.
In this regard, closed-source models still have an advantage because they have the most powerful foundation models. The distilled or trimmed smaller models will be more competitive than those from open-source models. Your foundation is stronger, so the resulting models are more competitive.
We have seen an overwhelming demand in this area, and we believe the opportunities still lie more with closed-source models than with open-source ones.