T-EDGE Conference Panel Calls for Use of AI to Address Issues in Traditional...
TMTPOST -- While AGI (Artificial General Intelligence) may not be realized in the short term, large models will become more powerful by 2025, generating more products based on specific use cases, Zhou Hongyi, the founder of 360 Group, said during a panel discussion at the T-EDGE Conference.
In the coming year, the importance of six key scenarios, including universal AI access, smart everything, and the digital transformation of traditional industries, he added.
Zhou made the remarks at a cross-industry conversation "What for AI VS AI for What" at the conference last Saturday. He was joined by Kevin Baragona, founder and CEO of DeepAI; Zhou Yuan, the founder and CEO of Zhihu; Leo Zheng, the founder and CEO of Lighthouse Capital; and Liu Jianhong, a renowned host and soccer commentator, the chairman and CEO of Kailishi. They provided insightful examples of AI in different industries and discussed the technology's development, application scenarios, business model innovations, and future outlook.
The dialogue was moderated by Zhao Hejuan, the founder, chairperson, and CEO of TMTPost Group, and chairperson of the T-EDGE Global Committee.
Baragona emphasized AI's potential in artistic graphic generation and question-answering software, particularly in legal and financial matters. He highlighted the success of DeepAI's image design tool, which showcases AI's creative potential in the arts.
Zhou spoke about the ability of generative AI and large models to solve complex problems in traditional industries. He cited the example of Stanford Medical Center, demonstrating AI's practical value in healthcare.
Zhou also expressed optimism about AI's role in the movie industry, particularly in language alignment and content generation, noting that AI helps filmmakers create high-quality films more efficiently. He stressed that community interaction is about the stories and inspiration generated through human interaction, and AI can enhance the quality of content and services, expanding the overall market. He also highlighted the importance of building new closed-loop ecosystems as AI search technology improves.
Zheng examined the matter from an investment and banking perspective, saying that AI’s revolution lies in its ability to directly generate productivity. He foresaw AI moving from providing production tools to directly participating in the production process through an "end-to-end" model. This deep integration with different fields' knowledge and technology will reshape the entire production chain, offering one of the most transformative opportunities in AI development.
Liu discussed the application of AI in the sports industry, particularly through his company, which focuses on "AI + sports." His firm has leveraged intelligent recognition and production technology to enhance the efficiency of broadcasting sports events such as marathons and billiards. He expressed hopes for breakthroughs in broadcasting by 2025 and aims to improve the overall quality of sports content delivery using AI.
The discussion also touched on the responsibilities and challenges AI faces as large models evolve. Issues like the authenticity and professionalism of content, errors in generated content due to AI hallucinations, and the ethical impacts of AI on society were raised.
The following is a curated transcript from this summit dialogue, translated and edited by AsianFin for clarity and brevity
On "The Best Scenario": Focus Less on Grand AGI Narratives, More on Traditional Industry Pain Points
Zhao: Thank you for joining us at T-EDGE for an interdisciplinary conversation. The five guests on stage are entrepreneurs who are also pioneers in the application of AI in their respective fields. What inspirations will fly this discussion create? I’m so excited.
The first question is directed to all five guests. In your respective fields, what do you see as the most representative case of AI application? This case could be from your own company or not.
Balagona: I think it’s the Q&A software used for retrieval and search. I use it to answer legal or financial questions.
Zhao: You also create AI tools, many of them. Which one do you think has been the most successful?
Balagona: I think the image design generation tool has been the most successful.
Zhou: Image generation and text generation don’t have much persuasive power; they are just inherent capabilities of generative AI large models. The biggest challenge now is how to use AI to solve problems that were once impossible to tackle in traditional fields. I was planning to advertise at first.
Zhou: I was going to talk about how we use AI to solve cybersecurity problems. In fact, there are many startups in the U.S. solving problems in traditional industries. They aren’t working on grand models for industries like steel or pig farming, which are very hard to deliver because you can’t easily define the problem you’re solving for your customers.
I’ll give an example from Stanford Medical Center. The Stanford Medical School has broken down the medical service scenario and used three vertical large models to address three issues: First, when a patient needs a referral from a family or community doctor, it used to be faxed by hand, but now with multimodal capabilities, OCR technology solves the issue that once required a lot of human labor. Second, a digital human model is used to return calls to patients to discuss and schedule appointments using speech synthesis and understanding. Third, after a visit, insurance claims require highly specialized individuals to write reimbursement reports. Stanford Medical Center accumulated a lot of materials and created a vertical model specifically for writing these reports.
By implementing these three solutions, Stanford Medical Center reduced the workload of around 100 support team members. So, the right direction is to identify bottlenecks in workflows within traditional industries and use AI to address them, rather than chasing grand narratives.
Today’s AI large models should not be overestimated in terms of their current capabilities. AGI development will not arrive so quickly, but we shouldn’t underestimate its potential either. At present, large models are more suited for breakthroughs in detailed scenarios. As mentioned earlier, the service market in the U.S. is about one trillion dollars, but with AI large models, it could grow to ten trillion dollars.
Zhao: The speed at which AI is transforming the U.S. B2B market is very fast. Some even say AI will rewrite all software. The software market was already huge, but after AI rewrites it, the software market could be 10 times or even 100 times bigger. The B2B entrepreneurship around AI in the U.S. is indeed booming.
Zheng: Currently, AI applications overseas focus on B2C/B2B, while in China, it’s more about B2C and AI for industries. Especially on the industry side, AI is helping companies directly interact with customers by deeply integrating with product systems, providing frequency and depth of interaction, and achieving product-service integration, which can greatly shorten and reconstruct industrial value chains.
Not long ago, I was talking to the founder of the largest engineering machinery company in China, and he brought up a scenario I never considered: the manuals for engineering machinery. Unlike home appliances, an engineering machinery manual can be thousands of pages long. Every time a different batch or supplier provides parts, the manual needs to be reprinted, otherwise, users won’t be able to find the right parts. Now, with large models and generative AI, companies can solve this issue. Users can simply take photos to identify parts and find specific uses. In this small detail, AI has greatly improved industry efficiency and controlled costs.
Zhao: Jianhong has been talking about how sports can benefit from large models and AIGC even before ChatGPT was released. The sports industry recognized early on how large models could transform the sector. The tech industry should pay respect to this.
Liu: I’ll try to answer this broadly and see if I can advertise a bit. Television is a technology- and labor-intensive industry. For broadcasting sports events, you need slow-motion and various data; it’s very complex and requires a large team to accomplish. During this year’s Paris Olympics, I said that television was still growing along the same traditional path, and the new technology for the Paris Olympics was all about adding more hardware and manpower. I believe this will be the last Olympics produced using traditional methods. The next one will definitely change.
We are working hard on this. During the Beijing Marathon not long ago, with 30,000 participants, we produced 30,000 personalized short videos for all the runners on that day. This is something traditional industries couldn’t do. This was just our first product, a beginner AI product that we’ve been working on for six years, and it has matured significantly.
On AI Search: Divergences and Consensus Among Internet Veterans
Zhao: Kevin, you founded Deep AI in 2016 when generative AI was not a mainstream technology. Why did you choose to start a company focusing on this field back then?
Balagona: I believed that one of the best things computers could do was generative AI. Even though the quality wasn't very high back then, it seemed like an impossible task. But generative AI was like magic. I saw the technology that way at the time. Later on, we were able to make the AI perform better, with quality improving over time, just like magic becoming more impressive.
Zhou: Making short dramas is a form of performance art. In fact, it’s to promote my product, Nano AI Search. The AI technologies mentioned earlier are great scenarios, but for many companies, the more realistic approach is to redo existing services. I’ve been in search and browsers for a long time, and it was natural for us to think about using AI to reshape search.
Nano Search has three key elements: 1) Multi-modal; 2) Answer engine; 3) Generative creation capability.
People originally felt that search engines provided too many answers, and the business model often involved ads within the links. People prefer more concise answers. We use large models to read through hundreds or thousands of webpages and give you a direct, straightforward answer, which takes advantage of generative features. Why is it called a creation engine? Search is just a means, not the goal. You are still using search to create, just like when your boss asks you to write a market research report.
Zhao: Now that 360 is working on AI search, where do you see the most commercial opportunities? There’s a unicorn company in the U.S. called Perplexity, which is now considered the biggest challenger to Google. It’s a pure AI search company, and its valuation increased 100-fold in just one year. But they haven’t solved the commercialization problem. The challenge of AI search is figuring out a new business model.
Zhou: We’ve studied Perplexity. The technology is nothing special. The business model is the new challenge for AI search, which presents an opportunity to challenge traditional search. If we solve the business model of AI search, Google and Baidu can easily upgrade their search systems. The reason Google hasn’t acted yet isn’t because they’re foolish. There are many smart people at Google, and turning their massive search volume into a free, ad-free answer engine might not be able to defeat Perplexity. Google might self-destruct if they did that, and they can’t afford that. This might be Perplexity’s opportunity.
AI might force internet companies to change their traditional advertising model. Originally, internet services had low costs, and revenue from ads could cover these costs. But have you noticed that OpenAI recently announced the GPT-01 version? The price went from $20 to $200 per month, indicating that the cost is too high to sustain. The commercialization model of AI applications is still in its early stages, and the key now is finding the right scenarios. This problem is far from solved.
Zhao: Everyone knows that Zhihu’s product is a Q&A community. With the support of new AI technology, does Zhihu want to become a knowledge search platform?
Zhou: First of all, we’re not undergoing a transformation. As Zhou just mentioned, an answer engine can quickly summarize multiple web pages and give you a concise answer, but many people don’t actually need that.
A community is fundamentally a space for people to communicate, discuss, and exchange. Often, efficiency isn’t the priority. It’s the stories in between that inspire you, rather than the final conclusion. Human interaction isn’t about prioritizing efficiency.
Zhou: I completely agree with your point. After all, we’re not competing. Zhihu has a lot of discussion material that is perfect for short dramas.
Zhou: When it comes to creation, I tend to think about how, with the empowerment of new technology, we can produce high-quality content that helps people achieve results they couldn’t before.
If we break down the content industry, the biggest business model for content platforms is advertising monetization, as Zhou mentioned. However, content can also be turned into a product and then into a service. Turning it into a product is like current short dramas and other video content. This market is relatively small, around 100 billion yuan.
If AI technology can significantly improve product service quality and deliver a user experience that hasn’t been possible before, this has significant potential. As Zhou mentioned earlier, helping people write short essays faster might not be the right approach.
Liu: I believe that whether it's text-based search, image-based text-to-image, or video generation, these technologies face challenges in making a significant breakthrough because the scope is too broad. Artificial intelligence in sports, however, is particularly suitable because of the standardization in sports.
AI is especially suited for sports because of the standardized nature of the games. For example, ping pong tables are the same worldwide, and sports like tennis, basketball, and football follow the same rules. Sports events are already constrained by the rules, with fixed venues and personnel, and relatively fixed broadcasting logic. With artistic treatment added, sports stories can be transformed immediately.
Zhou: Returning to your question, I think AI search is still very incomplete. Traditional search should not only focus on the search box and front-end interaction; it should also consider the search entry point, the indexed web pages, and the ecological chain services that the webpages provide. It’s a closed-loop ecosystem.
The experience of AI search, like Zhihu’s direct answers, will keep users on the platform for a longer time, which is quite different from traditional search. Traditional search takes you to another place, but the new supply chain isn’t fully formed yet. Various agents will help complete the new interactive experience.
Previously, you used a search engine to book flights, which relied on web pages and websites. Now, there’s a new supply chain—service supply chains and content supply chains. If it’s an agent ecosystem, the tasks completed by these agents are workflows, and many of these workflows will emerge in the future.
Zhao: I’ve realized that when humans invented language, we also limited ourselves. When we invented the term “search,” we might have been unable to find a word that could replace it, so we used “search.” But now, we’re limiting ourselves in creating new products. So, how should we name this new way of AI search?
Zhou: Anything that cannot be expressed by language, humans cannot think about. Language is the boundary of thought. “AI search” is a better concept than “AI assistant.” AI, including ChatGPT, has two problems. It promises everything, but sometimes some expectations are met, while others fall short, so the results are uncertain.
The biggest obstacle to AI’s popularity among ordinary users is prompt engineering. If you write bad prompts while using ChatGPT, the result is poor. Search, however, is something users are already familiar with. It’s simple to use and doesn’t require education or writing prompts. What users want from search is simply to find an answer.
Zhihu became popular because people were tired of getting 10,000 results when searching for "Liu Jianhong's entrepreneurial projects," only to find out that it wasn’t about his projects but his scandals. When people used Zhihu, they could get clear and simple answers to their questions.
However, Zhihu relied on human answers. Perplexity did two things: First, it understands your intent very well, even if your words are unclear. Second, it uses AI to generate answers, and these answers are generally reliable.
Zhihu became a community, as Zhou Yuan emphasized the relationship between people. In the future, there might be AI digital experts, as you mentioned, agents who can discuss problems with you further. I think Perplexity is also using AI. The concept of AI search is valid. Zhihu has launched this product, and we’re also working on search. It’s a painful process, and we have to self-disrupt, but we still can’t find a place to place ads right now.
On the Relationship Between Generative AI and Humans: Efficiency Should Not Be the Only Measure
Zhao: This leads to my next question for Zhou Yuan. At this year’s World Internet Conference in Wuzhen, you mentioned that in the next decade, IoT will evolve into the concept of IoA (Intelligence of All Things). The core of community is human interaction. When products like Zhihu evolve into an intelligent community under AI, how will AI manage relationships between people? Will Zhihu’s community foundation also be affected?
Zhou: Many of humans’ fundamental needs are stable and long-lasting. For example, things like birth, aging, illness, death, and social interactions, as well as the building of cognitive systems, are common across generations. Human-to-human interaction and AI-to-AI interaction are two different paradigms, and we shouldn’t simply divide them by efficiency. This relates to each person’s reward system and whether their motivation has been significantly changed.
Earlier, we talked about sports technology. In the future, what will become valuable and real? I believe competitive sports, the act of “humans running,” cannot be replaced. It’s driven by human curiosity.
Zhao: If we go deeper, it might lead to philosophical discussions. Zhihu is a customer of Lighthouse. Lighthouse has many internet company case studies from the last internet wave. Now that AI is here, do you think the technology we’re seeing right now will be replaced by something else, or will it be fully integrated?
Zheng: The biggest difference between this generation of AI and previous ones is its generalization capability. It forms the ability to generalize while solving general problems. Previous AI couldn’t discuss daily active users (DAU), but this generation of AI is qualified to discuss it.
As AI continues to evolve, the problems it solves are different from those of mobile internet. Mobile internet is essentially about connection—efficiency, spatial and temporal scale, etc. It connects things that weren’t online, making them online and accessible.
But the essence of AI is the integration of perception, decision-making, and execution. It solves problems directly by generating the service itself and producing the results users want. Since the problems it addresses are different, things that mobile internet used to handle will be done again in an end-to-end manner. Additionally, AI has a great opportunity to address problems that mobile internet couldn’t solve before. Many previously unimaginable scenarios might be realized through AI.
The biggest differences between the Chinese market and other places can be summarized in two aspects:
1. A huge consumer market. Chinese products, especially B2C products, theoretically have greater daily active users, larger scale effects, and network effects than products in other parts of the world, leading to great commercialization prospects.
2. Better industrial soil in China. China is the world’s largest industrial nation, with the most complete industrial sectors and supply chains. In China, we focus on AI for industry and AI for consumers, while in the U.S., AI for business is more emphasized. Therefore, the logic of earning money from industries and consumers in China is perfectly valid.
The three things Chinese companies care most about:
1. Intelligence—how to use AI to create new products or scenarios and disrupt or transform their industries. For startups, it’s about creating industrial scenarios and applications.
2. Globalization—Chinese entrepreneurs are exploring overseas markets because Chinese companies have the foundation to compete globally. Even though China’s market is the largest, the supply chain dividend in China is not yet large enough for Chinese engineers.
3. Industrial innovation—entrepreneurs are focused on how to avoid internal competition and innovate more effectively.
Among these three, the most important is the wave of intelligence, i.e., AI. As the two previous speakers mentioned the issues around industrial scenarios and applications, it shows that AI has already made significant progress.
Zhao: Jianhong, we are both professional media people, and we have an obsession with professionalism that cannot be challenged. I used to be an investigative reporter, and I had very strict requirements for my writing. I was responsible for the content I produced, ensuring its truthfulness and professionalism, with zero tolerance for typos.
I noticed a big problem with AI-generated content: users often receive content that is misleading, false, or not entirely real. When AI generates sports content, if there’s a mistake, the platform might not care, but as professional journalists, we see this as a serious issue. How do you deal with this error rate?
Liu: We started by recording and reflecting human sports activities objectively. The first thing we have to do is record things truthfully, and I believe this is now fully achievable. We’ve done more than just marathons; we’ve also fully automated the intelligent broadcasting of snooker and table tennis events, with the level of AI automation in these events far surpassing marathons. Broadcasting sports events is a multi-person, single-task operation. Everyone works towards the goal of broadcasting the event, and there are patterns to follow.
Zhao: It’s one thing if the job gets automated, but as I mentioned earlier, I need to be responsible for my content, and the AI-generated content under my name. Just like a commentator—people remember the personalized moments of Huang Jianxiang.
Liu: McDonald's serves the masses to fill their stomachs, while Michelin offers personalization. In my industry, achieving the "McDonald's" level of service is already easy. I can clearly see the path. The future sports “Michelin” won’t be Huang Jianxiang or Liu Jianhong; it’s the logic of the future.
In the future, sports events broadcasted by AI might be fully data-driven. AI’s reasoning and associations could follow a different logic. Even the best commentators can’t know all the spectators, but AI can build connections, like the relationship between Messi and his first coach, or between Lao Zhou and his rumored girlfriend. AI can establish these connections and express them through its reasoning capabilities. People will see sports differently, and even the sports humans engage in might change.
Looking forward to 2025: workflow, embodied intelligence, AI for science, and more
Zhao: Due to time constraints, I’ll ask the next question to all five guests. As we approach the end of 2024, what are your predictions and expectations for AI trends in 2025?
Zhou: I’m still very clear about two major opportunities: the base models and scenario-specific applications. Looking at Nvidia’s quarterly revenue of over $30 billion from base models, that’s the cost of these base models. In China and the U.S., about 10 companies are spending large amounts to scale these models.
Of course, Nvidia also has a consumer business. If we halve that, it's about $1.5 billion, and the total spend per year is about 20 billion RMB. The number of players is limited, so with the open-source technology and applications already in good shape, the open-source application scenarios, including a lot of workflows, will be in high demand. Agents will be customized solutions for problems.
Balagona: I believe AI will continue to flourish, and we’ll see more applications, higher quality, and faster developments. In 2025, people might become somewhat fatigued because AI will be everywhere, much like now—pervasive in gaming models, music models, etc. The key will be how we define AGI. You could argue that we already have AGI, and that it will just keep improving in the coming year.
Zhou: If AGI is defined as general or super intelligence, I believe that AGI might not appear in the short term. There are still many obstacles with Scaling Law. When it comes to big models applied to real-world scenarios, it will form two systems. Worldwide, no more than ten companies will continue to compete in AGI by scaling computing power, data, and training. I expect that by 2025, we’ll see big models as capabilities rather than products.
There are six key scenarios to focus on, which have already started in 2024:
1. Universal AI, how to turn AI into a productivity tool for everyone in the 2C field.
2. Intelligent everything. All smart hardware, including industrial equipment, will consider edge models, and your home appliances will start talking to you.
3. The “digital transformation to intelligent transformation” of traditional enterprises will lead to great progress.
4. Helping emerging industries, including embodied intelligence, autonomous driving, and low-altitude economy.
5. AI for science—this year’s Nobel Prize already demonstrated this.
6. AI security. Both AI solving security problems and ensuring AI security will become increasingly important.
Zheng: I completely agree with what Zhou mentioned. Let me focus on the opportunities for startups.
In the 2C space in China, AI combined with specific scenarios to generate services, content, and results will likely present opportunities for startups, especially in fields like gaming and education.
I’m also optimistic about embodied intelligence, particularly in China. While China started later and has fewer computing resources, it has the most powerful hardware supply chain and the best application scenarios, such as in industries, logistics, and more. China is in a strong position to lead this area.
From a software perspective, China is not falling behind. Many people working on embodied intelligence in both the U.S. and China are from the same batch of graduates, mostly from Tsinghua, Peking University, and Stanford. The projects started around the same time, and we’re all at the same starting point.
That’s why I believe China will become a leading ecosystem in embodied intelligence, and LightSource is actively investing in this direction.
Why do startups have opportunities? Traditional tech giants aren’t enthusiastic about hardware products, and the gap between their software capabilities and the demands of embodied intelligence is large. Embodied intelligence requires software to define hardware, and startups have a great opportunity to surpass them.
As Zhou mentioned, moving from “idiot” to “smart,” once the cost issue is resolved, AI will become a game-changer in fields like industrial and logistics applications.
AI for science is also very promising, as seen in the Nobel Prize trends. In fields like materials science and life sciences, AI is no longer just a connector of productivity—it has become the productivity itself, directly participating in the process and even replacing some human roles.
This will be one of my focus areas in the coming year.
Liu: When we talk about breakthroughs, it’s not about the overall picture, but rather specific points or areas. In 2025, there might be breakthroughs in projects like marathon and snooker broadcasting, where AI will surpass human capabilities. I can’t wait for that day to come. I look forward to AI's progress providing strong support for us to achieve breakthroughs in these areas.
Zhao: A special thanks to all the guests for their wonderful insights and discussions. Thank you, everyone.