Investment trends and application directions in AI: a discussion between Chinese industrial leaders
On April 12, 2024, Tencent Research Institute, in collaboration with Qianhai International Affairs Research Institute and Qingteng, jointly launched the AI&Society Artificial Intelligence + Social Development high-end seminar. The seminar invited industry leaders and renowned scholars in the field of artificial intelligence and social governance from both domestic and international arenas. A roundtable, "AI Investment Trends and Application Directions", was held during the seminar.
The participants agree that AI has great potential in improving efficiency, but its practical applications in marketing are limited. The current focus is on changing workflows, modes of thinking, and data security issues. Personalized customization has become less expensive but may not bring significant competitive advantages. Startups must educate companies about AI, turning it from a buzzword into a practical solution. Industry communication is crucial for startups to understand customer needs and gain more perspectives.
Speakers:
Wu Shichun(吴世春), Founding Partner of Plum Ventures
Fan Ling(范凌), Founder and CEO of Tezign, Professor and Doctoral Supervisor at Tongji University, Director of the Design Artificial Intelligence Laboratory
Zhou Jian(周健), Founder and CEO of Lanma Technology
Yuan Xiaohui (袁晓辉, Roundtable Moderator), Senior Expert at Tencent Research Institute
Here are the transcripts of the roundtable discussion:
Yuan Xiaohui:
My first question is about AI Agents, so I'll ask Professor Zhou Jian. RPA robots have played a certain role in software process automation. With the emergence of large language models, what do you think are the essential differences between AI Agents and traditional process automation robots? And what new possibilities does this bring us?
Zhou Jian:
In the past, RPA was mainly a UI automation technology, with its core task being to solve the problem of data transfer between different information systems in the era of informatization. For example, we may copy files from one application to another on our phones, such as forwarding from WeChat to email, which is the problem that RPA tries to solve. However, in the AI 1.0 era, due to AI's limited language understanding ability, when we automated processes, Agents did not truly understand the data or documents. However, with the arrival of large language models, we now have a more cost-effective language understanding capability. This allows AI Agents to make significant progress in document understanding.
The scenarios we just discussed, whether involving tables or documents, require information understanding and transformation. AI Agents can now help customers with problems such as report generation, report summarization, and information search. I believe this is a change that has already happened because AI Agents can now understand this content. Additionally, today, I also see another possibility: natural language can be transformed into a command. For example, in the application scenario of bank letters of credit, a cell in a table may contain a description requiring the letter of credit to be audited according to specific rules. In the past, this audit process could not be automated. But now, large language models can understand these commands, generate code based on them, and execute them, which has great potential for the future. If this ability is further developed and iterated upon, it will greatly drive productivity explosion. In a sense, natural language is like magic, and AI Agents will be able to understand this "magic" and transform it into automated actions. This will be an extremely powerful development.
Yuan Xiaohui:
Alright, thanks, Professor Zhou. Next, let's discuss a related issue: AI natives. We know that the concept of AI Native, similar to cloud-native or digital-native, is a business model that emerges under the support of the AI era or large language models. I want to ask Mr Wu and Mr Fan if they have encountered this issue in your investment or customer service processes, or do you have any thoughts on AI Native direction? For example, companies like character.ai or the previously mentioned Miao Ya Camera might be business models or entrepreneurial directions that only emerge after AI or large language models appear.
Wu Shichun:
Currently, about 30% to 40% of the AI projects we see are completely based on AI native concepts. Many projects have migrated from past projects, which may have already existed in the AI 1.0 era and are now incorporating large language model capabilities. For investors, these projects demonstrate their revenue-generating capabilities more quickly. That is, the products already have customers. On the other hand, AI-native projects first need to educate the market and acquire new customers. The initial users may be driven by curiosity or self-consumption within the industry. AI native projects are providing new ideas and directions for ChatGPT. We are currently in a phase of rapid iteration, and many applications we see are still relatively basic. We have not yet seen powerful AI native projects, like the strong projects that emerged in the mobile internet era, such as Today's Headlines, Changba, Meituan, and Didi, completely based on mobile applications. I haven't seen such large-scale applications in the United States either.
Yuan Xiaohui:
There are still some companion applications.
Wu Shichun:
These are all easily replaceable by the next generation of upgrades because the large models themselves are constantly evolving and upgrading. So, I believe ChatGPT might destroy you, but that's not your concern. Forming independent logic and completing a major user experience service loop is difficult.
Yuan Xiaohui:
Mr. Fan, what are your thoughts? You also deal with content plus AI on your end.
Fan Ling:
I don't indulge much, so I think the likelihood of native apps emerging in the West is higher. As Mr. Wu mentioned, it's been a long time since we've seen new consumer-facing apps.
Wu Shichun:
In the United States, many AI applications initially make money by meeting users' niche needs. But in China, these needs are not allowed. A forest needs various trees, flowers, and grass to thrive. But currently, that's not allowed; only high-tech is permitted. So, I think it's very challenging for many companies to survive in the early stages.
Fan Ling:
Indeed, many AI native applications are targeted at overseas markets, often related to gaming and specific content. In a B2B scenario, the core problems to solve remain the same. It's all about finding more efficient solutions. This doesn't rely on the inclusion of AI but the integration of AI with other domains, such as industrial IoT and marketing technology. While the concept of AI natives has garnered a lot of attention in China, leading to a surge in the stock prices of many concept stocks, in the United States, companies that make money in the B2B field, like CRM and marketing automation companies, see improved user experience and smoother business processes after integrating AI, resulting in better stock performance. I believe truly native AI applications may still be too "fragile" at this stage and need time to mature and improve.
Yuan Xiaohui:
From the perspective of AI natives, AI Agents might also be a direction. We've also heard that GPT-5 in the AI Agent space could be very powerful. Professor Zhou, are you concerned that the emergence of GPT-5 might surpass existing related services?
Zhou Jian:
For us, we're not worried about being surpassed. When GPT-4 was released in March, we did feel anxious, but now we're increasingly less concerned about this issue. Sora talks about the world model, but I think it's more of a marketing tactic. In an enterprise environment, organizations have their processes and personnel, just like Einstein couldn't become the CEO of a publicly traded company without learning the organisation's internal environment. We believe that enhancing large models will strengthen our AI Agent's ability to serve customers within the enterprise because we have proprietary data within the organization. AI Agents need to gather information from the environment and change it. I think this requires drawing on some principles of bionics, such as how the human brain works. Recently, I've been pondering the issue of short-term and long-term memory for AI Agents, and I'm starting to think that memory is not just a database but a more complex concept. In the human brain, the hippocampus has two types of cells related to spatial memory. I believe constructing a world model that can understand spatiotemporal similarity is challenging and not a problem that can be solved simply by publicly available data. It's crucial to implement Next Action Prediction in AI Agents within the enterprise. This might be a prerequisite for creating engaging, intelligent agents.
Yuan Xiaohui:
What you mentioned might involve software aspects, hardware, and interactions with the environment, including how it gathers more information through interacting with the environment and iterates on its process. All three of you mentioned contextualization, including our industry's interest in AI. Which industry sectors and segments are the quickest to implement or more willing to purchase existing AI services? Can you share some insights?
Zhou Jian:
Implementing AI technology should start with traditional industries, especially the finance and energy sectors. Taking the finance industry as an example, major banks like ICBC have already deployed over two thousand GPUs, while securities firms, although slightly behind, are catching up within three to six months. On the government side, many local governments are building large data centres with considerable computing power. However, due to lower profits, the retail and manufacturing industries are not keen on investing in projects that may not change the current economic environment. This is an observation I've made within the customer base.
Fan Ling:
In the current entrepreneurial environment, the application of technology needs to be very specific and directly beneficial to businesses. It should help increase revenue within a year or improve efficiency within three months. If the CEO's tenure is longer, technology should also help them enhance profitability within three years. From my perspective, the ultimate goal of all technologies revolves around these three targets.
If we were to prioritize these three goals, "improving efficiency" might be the most critical. A key way to "improve efficiency" is by reducing costs and enhancing sales efficiency, particularly in industries like fast-moving consumer goods, beauty, and apparel retail. The pharmaceutical manufacturing industry is now also focusing in this direction due to changes in centralized procurement policies. In the long run, the growth potential may lie in operating new channels, such as integrating online and offline channels. AI can play a crucial role in this process, such as providing personalized services when selling complex products. We have clients in the cosmetics industry with an average order value ranging from 3000 to 4000 yuan, where personalized service from sales staff is crucial. Now, with AI technology, it's like providing sales staff with an assistant, making their communication with consumers more efficient.
Wu Shichun:
I believe that analyzing industry pain points can be represented by three circles. The first circle is which industries have capital, and the second circle is the pain points within these industries, such as cost reduction, efficiency improvement, and saving or making money. The third circle is alternative solutions, whether using large models is better or more efficient than traditional solutions. The intersection of these three circles is where companies can gain benefits. Industries like finance, security, oil, and power still have ample capital. Large brand retailers also fall into the category of industries with abundant funds. A company we invested in, originally in operations, has now introduced new services combining large models, mainly targeting critical applications for key customer services. These are the areas companies are most willing to invest in. China's ToB market is relatively small, with the 2C market potentially comparable to the United States, but the ToB market may not even be one-twentieth of the size of the U.S. market. Large enterprises develop their technology in this market, medium-sized enterprises use open-source technology, and small enterprises do not. Large models could be a new way out for the SaaS industry, and we have also invested in several SaaS projects. In China, the SaaS industry faces challenges such as poor payment capabilities and high demand for customization.
Yuan Xiaohui:
Some opinions suggest that AI may bring new hope to the SaaS industry. AI can provide services to new customers in delivery, personalized customization, and code generation and bring new hope or provide personalized services with fewer human resources.
Fan Ling:
In concept, everyone generally agrees on the value of AI. Still, in practical applications, especially in the marketing field, we find that although the efficiency of content production has increased, the time and cost it occupies in the marketing chain are not high, maybe not even 10%. For example, using Midjourney technology can quickly create a poster, but the initial communication work is very complex and takes up 90% of the time. Although AI can rapidly generate many images, management often gives specific modification requests for individual images, such as requesting larger font sizes, leading to the workflow reverting to manual operations, and efficiency cannot be improved. If the client can accept multiple options provided by AI instead of giving specific modification requests, then the potential of AI to improve efficiency will be fully realized. Currently, when companies discuss AI, the focus is no longer on technical issues but on how to change workflows and modes of thinking and data security issues. For example, McDonald's will not provide its core data to GPT. Currently, our main discussions revolve around productivity issues rather than the capabilities of large models.
Zhou Jian:
My view is slightly different. In the past, we've seen many RPA process products in the market, although ultimately, only a few companies could achieve positive cash flow in the RPA era. But now, with the significant reduction in the cost of personalized customization, we are starting to replicate these applications in large banks and have the opportunity to reduce the cost of personalized customization to 10%. We emphasise expert knowledge, such as rapid analysis and evaluation of a company's financial situation. These knowledge and evaluation standards are often dictated by industry norms or government regulations so they can be replicated. The personalized aspects, such as specific rules and requirements for each bank, although different, have similar sources of information needed. In the current scenario, we can see the capabilities of open-source models are continuously improving, and large model companies are also making progress. For example, we can already calculate the breakeven point in intelligent customer service.
Yuan Xiaohui:
Sequoia Capital recently issued a report saying they found this PMF in the customer service scenario.
Fan Ling:
I wonder if Mr. Wu has ever invested in such companies. There may not necessarily be so-called AI-native applications, but companies will be built on the native concept of AI. These companies are founded on AI to optimize collaboration between people and improve efficiency. For example, an AI Agent might not just be for human use but could play different roles and collaborate to get work done. We might see AI-native chain stores or cross-border e-commerce companies whose value could far exceed a single AI application.
Yuan Xiaohui:
Are you suggesting that various tools are used to optimize internal processes to the extreme?
Fan Ling:
Yes, now we need to consider AI Agents or AI as an indispensable part of the organization, thinking about organizing them to collaborate with humans to solve problems. AI-native business organizations have been built on this premise from the beginning.
Yuan Xiaohui:
So, from the beginning of the company's establishment, a strong team composed of digital employees was built.
Wu Shichun:
At the organizational level, due to inertia, companies may not quickly adapt to changes brought by new technologies. We may need to wait for more cases to emerge, perhaps in Silicon Valley and other places. A common pattern is that the United States invents new technologies, and China can scale them up for mass production, even turning them into low-cost goods. This trend has already emerged in fields like new energy vehicles. The AI field might follow a similar path; once a new technology emerges, China will likely quickly popularize it, even if it doesn't initially make money.
Yuan Xiaohui:
We recently conducted a study, interviewing over 100 experts from different industries to explore which industries are more proactive in using AI. Due to its strong digital foundation and high employee acceptance of new technologies, we found that the financial industry is clearly leading the way. Digital native industries, such as software, advertising, tools, and SaaS, are also showing a faster pace of AI application. In contrast, industries with heavy assets like agriculture and manufacturing, or those closely interacting with the physical world, have a relatively slow pace of AI application. In different parts of the industrial chain, we can see that design and development and sales and customer service, the ends of the smile curve, are adopting AI applications more rapidly, while the production and manufacturing processes are relatively lagging behind.
Wu Shichun:
The more transparent the stages that use AI, the harder it is to form a competitive advantage. On the contrary, once they establish a competitive advantage, industries that need to overcome many usage obstacles will have a more solid market position. Industries that are initially easy to make money in may become increasingly difficult to profit from over time. Whereas industries that initially seem challenging to profit from, if you can gradually overcome usage obstacles, may eventually establish an invincible market position.
Yuan Xiaohui:
Indeed, we should stick to doing the right and difficult things, as these difficult tasks often help establish competitive barriers.
Wu Shichun:
Yes, industries that don't make money initially may lack attention and understanding, but they can build high barriers over time. On the other hand, those industries that initially seem very profitable and transparent often end up not making money because someone will always be offering services at a lower price.
Yuan Xiaohui:
Thank you, Mr. Wu. Earlier, we discussed industries or processes that are relatively easier to implement. Now, let's talk about the pain points and bottlenecks that enterprises face in driving AI applications and conducting POCs with customers. Are there issues related to computing power, awareness, talent, etc.? Do you have any experiences to share?
Fan Ling:
We mainly deal with large enterprises. In reality, very few enterprises are truly willing to commercialize their computing power. Most companies use AI for applications and reasoning without building models or selling hardware. The main bottlenecks we encounter currently are whether the company's processes are AI-friendly, whether they are willing to make necessary reforms for AI, and whether the company can overcome concerns about compliance and data security. Of course, many technologies are available to help solve these problems, but companies also need to learn how to apply these technologies. As Wu mentioned, we have also found that the more transparent the field, the more willing companies are to invest, but these fields often do not create significant competitive barriers. On the other hand, companies tend to be more cautious in more challenging areas. Currently, most Chinese companies allocate their AI budgets to innovation, which means they are marginal budgets that can be cut. For example, in China, some pharmaceutical companies invest heavily in AI for new drug development because a new drug could potentially have a market worth billions of dollars so that they can invest significantly. However, they have not invested this money in China. In China, most companies allocate AI to relatively transparent tasks, such as video production or improving human resources efficiency.
Yuan Xiaohui:
Are there any breakthrough strategies?
Fan Ling:
There are two breakthrough strategies: first, we need to educate companies about AI, turn AI from a buzzword into a real possibility in their minds, and embed the native concepts of AI deeply. Everyone is talking about Sora and ChatGPT, but the number of people who truly understand and use these technologies is quite limited. Therefore, enhancing the awareness of AI among companies is crucial. Secondly, we need to understand the needs of companies truly. Companies aim for growth, both short-term and long-term cost reduction and efficiency improvement. If we can create specific AI applications, companies will usually be willing to adopt them. Pure innovation without practical applications often has a short lifespan.
Zhou Jian:
I want to add some observations from last year. Initially, we thought AI would be easily applicable in the legal and headhunting industries. Still, we later found that these two industries' informatisation levels were not high. They are concerned about "teaching the apprentice and starving the master." For example, initial phone calls and face-to-face meetings with candidates are crucial in headhunting. Headhunters usually maintain their own Excel spreadsheets to manage lists of a thousand or two thousand candidates and are unwilling to share them with companies. The situation is similar for lawyers, especially in China, where the level of information in law firms is generally low. Once a lawyer leaves, their ongoing projects may come to a halt.
Additionally, there are significant differences in corporate culture between China and the United States. Management in the US is more standardized, while Chinese bosses often believe they are unique and are unwilling to share their business processes. This mindset, to some extent, limits the application of AI. Furthermore, establishing an AI team effectively is necessary if companies want to utilise AI. This team needs to have business experts who may not need to delve into the technical details of AI but must be sensitive to data. For example, in the recruitment process, if no one internally can understand the resume data received and define matching rules, we, as service providers, will also find it very challenging. AI product managers who understand business knowledge are currently very scarce, which is a key factor in successfully implementing AI applications.
Wu Shichun:
My advice to entrepreneurs is that most scenarios where you can make money on your phone are dominated by big companies. In the US, despite the rise in the stock prices of AI concept stocks, apart from OpenAI, there haven't been any particularly profitable or outstanding companies emerging. Therefore, entrepreneurs should focus on doing the difficult and right things that big companies are unwilling to do or cannot see. Don't just focus on seemingly opportunistic areas; instead, delve deep into areas seemingly devoid of opportunities. This is the path to potential entrepreneurial success. In the past five years, there haven't been any new multi-million-dollar applications based on mobile phone scenarios. All the good application scenarios have been almost monopolized by BAT, ByteDance, JD, and other big companies. I hold a rather pessimistic view of new entrepreneurial opportunities.
Yuan Xiaohui:
During the selection process of entrepreneurial projects, investors often ask founders whether they have viewpoints or directions that they believe to be correct but others do not agree with. For example, Airbnb wasn't highly regarded by many investors in the early stages but eventually succeeded due to investors' persistence.
Wu Shichun:
The transition from computers to smartphones as carriers has given rise to new applications. Currently, we still see many applications based on smartphones, but in the future, new opportunities may emerge based on VR, machines as carriers, or vehicles as carriers. For applications based on smartphones, if we are still just repeating what already exists, much of our efforts may be in vain.
Yuan Xiaohui:
The final question, circling back to the theme of this event, is about the entrepreneurial ecosystem in the era of big models. All three guests are from Beijing and Shanghai. Looking at the entrepreneurial ecosystem, what would it be if you had to choose one area for improvement?
Zhou Jian:
For startups, the most crucial aspect is communication and matching needs. Large model manufacturers often say the scene isn't right, so we hope for more opportunities for such connections. Because we did well in brand building this year, many needs naturally came to us. At this stage, we particularly need professional scene connections, such as banks or the retail industry, to understand better how customers think and gain more customer perspectives.
Yuan Xiaohui:
So, industry communication is very important.
Fan Ling:
I believe that in the current highly competitive environment, maintaining openness is crucial. I'm interested in participating in this forum because I need to go everywhere, experience different atmospheres, and gain different energies. For the enterprise services we provide, especially AI-related when we reach a certain scale, efficiency will decrease if we don't break through. Besides going public and failing in entrepreneurship, the US has various exit mechanisms like mergers and acquisitions. In the current Chinese environment, founders need more collaboration methods, including mergers and partnerships. In the past, when funds were abundant, people didn't care much about product quality, but now everyone knows that only good products can reduce customer acquisition costs. There should be more choices for product boundary expansion and exit mechanisms.
Yuan Xiaohui:
Mr. Wu, what are your thoughts?
Wu Shichun:
I believe that the new tech leaders in the US, like Musk, are not standard scientists but talents with comprehensive abilities, even product-oriented CEOs. They can integrate research and development, supply chain, products, and sales and have a cross-disciplinary perspective. In China, many leading CEOs come from a scientific background quite different from the US. We hope to find talents who can transform technology into business outcomes, possess foresight and performance abilities, and describe ordinary things very appealingly, like Lei Jun.
Yuan Xiaohui:
Zhou mentioned communication opportunities, Fan talked about openness and diverse exit mechanisms and service systems, and Wu discussed versatile talents. These are all crucial factors in the entrepreneurial ecosystem in the era of big models. Many thanks to the three teachers for bringing new thoughts to our roundtable in the second half. Thank you!