Good day, and welcome to the Cheetah Mobile Second Quarter 2024 Earnings Conference Call. All participants will be in a listen-only mode. Should you need assistance, please signal a conference specialist by pressing the star key followed by zero. After today's presentation, there will be an opportunity to ask questions. To ask a question, you may press star then one on your touch-tone phone. Please note, this event is being recorded. I would now like to turn the conference over to Helen Zhu. Please go ahead. Thank you, operator. Welcome to Cheetah Mobile's Second Quarter 2024 Earnings Conference Call. With us today, our company's chairman and CEO, Mr. Fu Sheng, and our director and CFO, Mr. Thomas Ren. Following management's prepared remarks, we will conduct the Q&A session. Please note that the CEO's script will be presented by an AI agent. Before we begin, I refer you to the safe harbor statement in our earnings release, which also applies to our conference call today, as we will make forward-looking statements. At this time, we will now like the AI agent to speak on behalf of our CEO and chairman, Fu Sheng. Thank you for joining us today. In Q2, our total revenue accelerated with a year-over-year growth of 12.3%. AI and others accounted for about 40% of our total revenue. This growth was primarily driven by the sales of our wheeled service robotics in both domestic and international markets, showing our progress in becoming an enterprise-facing company. The acquisition of Beijing OrionStar has made service robotics a key pillar for Cheetah Mobile, contributing to a solid revenue growth. Our service robots hold a dominant position in voice-based use cases and are widely used in exhibition centers, museums, corporate receptions, and other areas. Moreover, our delivery service robot ranks among the top three in restaurants and continues to gain market share from competitors. Customers choose us because of our best product experience and after-sales services. Strong core AI capabilities, including far-field voice recognition. Our robots have benefited from LLM, being able to better understand end-to-end customers' inquiries and smartly respond to their requests versus previously. Customers can easily tailor make apps within our robots, thanks to our strong software abilities that support customization. To further expand the revenue growth of our wheeled service robotics business, we are focusing on two key strategies. Firstly, we aim to broaden the use cases of our service robots through continuous product innovations with a focus on our core competencies, including voice interaction capabilities and enabling autonomous deliveries. For instance, in factories and fulfillment centers, we have recently introduced robots designed for autonomous delivery for relatively low payloads. We offer updated performance and pricing for our customers in both China and overseas, with a focus on providing the most reliable robots in the market. We have already started shipping robots to South Korea and have received orders from customers in Japan and Southern Europe. In the hotel industry, our service robots are making progress. Hotels are a probable use case for wheeled service robots, but we are gaining market share from existing players. We are also further updating our robots for the hotel industry to increase our competitiveness. In supermarkets, our service robots have successfully facilitated the sale of low-ticket-priced products by identifying potential buyers, proactively approaching them, providing comprehensive product introductions, and clearly responding to consumer inquiries. Following our collaboration with Shuanghui to sell cod sausages in supermarkets, we are now expanding into more local stores and supermarkets. Secondly, we are expanding our service robotics business globally, with overseas revenue already surpassing domestic revenue. Following our success in South Korea and Japan, we are actively building our presence in Southern Europe, Southeast Asia, North America, and Australia, particularly in use cases such as restaurants, factories, and fulfillment centers. Chinese electronic products have demonstrated their ability to compete globally, and with Cheetah Mobile's extensive experience in international operations, we are confident in our potential to succeed in these overseas markets. We believe that the service robotics industry offers one of the largest market opportunities for AI commercialization. LLMs and generative AI serve as brains for robots, enhancing machine intelligence and accelerating their commercial deployment at scale. However, this industry is still early and will take time to unlock its full potential. Cheetah Mobile is working hard to make service robots more affordable for enterprises across a growing number of use cases, enabling them to reduce labor costs through the use of our robots. We focus on wheeled service robots because we believe they offer the best balance between performance and cost at this stage, making them scalable solutions. As the only robotics company in China to have trained an LLM from scratch, and its LLM approved by local authorities for a larger scale rollout, we use the LLM to power our robots, focusing on hardware, software, syncing, through serving customers in various use cases. We have notably enhanced our voice-based capabilities, in particular for reception-related use cases, and we will continue to use data generated by our robots to enhance our LLM and product experience, creating a positive feedback loop. This approach should allow us to deliver the best products for price and performance in the global market. At the same time, we remain committed to product innovation, low waste, striving to balance performance with cost effectiveness. During the last earnings call, we discussed how we are helping enterprises build LLM-based applications. Applications are key to making LLMs useful in enterprises. This is because LLMs need industry-specific or company-specific knowledge to avoid errors and effectively address enterprise challenges. Cheetah Mobile is developing Gen AI applications for enterprises. We are encouraging our employees to create Gen AI tools or apps to improve work efficiency. At the same time, we are working closely with key accounts, helping them develop Gen AI apps to streamline their daily operations. Our goal is to identify cases where Gen AI can enhance efficiency, and then standardize tools and features that can be scaled to other companies. For example, Savings Note, an LLM-based cloud management app, helps enterprises monitor and optimize cloud usage across platforms. After successfully using it at Cheetah to reduce costs, we've begun offering it to enterprise clients. Early customer feedback has been positive. We've also helped a major hotel operator in China develop a Gen AI app for employee training programs. Looking ahead, we will continue working closely with key customers, applying the Gen AI product we've developed internally to their business operations. This approach will help us further refine our product experience and gradually build a comprehensive product portfolio. Before I turn the call over to Thomas for financial highlights, I want to emphasize the following. First, our enterprise businesses, service robots, and LLM-based apps, each represent a huge market opportunity and are still in the early stages of development. Second, Cheetah Mobile, with a team that has extensive experience on the PC and mobile areas, along with strong AI capabilities, is investing in developing the best products to capture these opportunities. We are focused on achieving high quality, long-term growth rather than pursuing short-term gains. Thank you, Fu Sheng. Hello, everyone on the call. Please note that unless stated otherwise, all money amounts are in RMB terms. Cheetah Mobile delivered a solid performance this quarter. In Q2, total revenues grew by 12% year- over- year, reaching CNY 187 million. Non-GAAP gross profit increased by 11%, coming in at CNY 122 million. Our non-GAAP gross margin remains stable at 65% compared to the previous year. Despite our ongoing investment in AI, we successfully reduced our operating losses on a sequential basis. In Q2, our non-GAAP operating loss decreased by about CNY 4 million quarter- over- quarter to CNY 62 million. This improvement reflects our strategic decision to reallocate resources from our legacy internet business to our AI initiatives. Looking at our internet business, excluding share-based compensation expense, the operating margin increased to 12.4%, up from 7.9% in the previous quarter, and 5.5% in Q2 of the prior year. Revenues from this segment remained relatively flat year- over- year, with a 4% increase quarter- over- quarter. As we have indicated before, the year-over-year increase in operating losses was driven by our investments in AI, following our acquisition of a controlling stake in Beijing OrionStar. Specifically, these increased losses are attributable to higher headcount in R&D, sales, and G&A, as well as increased hardware-related costs for our service robots. As of June 30th, 2024, we had approximately 870 employees, compared to around 860 in the previous quarter, and 730 year ago. We are pleased to report that our AI investments are beginning to bear fruit. Our wheeled service robotics business has emerged as a key revenue driver. Additionally, through collaborations with leading companies on their large language model initiatives, we've gained valuable industry insights and are currently testing LLM-based applications for enterprise use. One of the quarter's standout achievements is our cash generation capability. Despite continuing to incur losses, we generated almost $220 million in cash from operating activities in Q2, highlighting our strong cash management generation capabilities. Finally, our balance sheet remains robust. As of June 30th, 2024, we have approximately $270 million in cash and cash equivalents, along with short-term investments, and about $119 million in long-term investments. This includes stakes in well-known entities such as Meta, and Meta AI Social. With that, we open the call for questions. Everyone, for today's call, management will answer questions in Chinese, and AI agent will translate management's comments into English in another line. Please note that the translation is for convenience purposes only. In the case of any discrepancies, our management statement in Chinese works well. If you are unable to hear the Chinese translation, a transcript in English will be available on our website within seven working days. Operator, we are now ready to take questions in Chinese. Thank you. Thank you. We will now begin the question and answer session. To ask a question, you may press star, then one on your touchtone phone. If you are using a speakerphone, please pick up your handset before pressing the key. To withdraw your question, please press star, then two. At this time, we will pause momentarily to assemble our roster. Ladies and gentlemen, please stand by for the English translation of the Q&A session. One, my question is about the product forms of robots. Nowadays, humanoid robots and embodied intelligence have been discussed a lot. Many people also define 2024 as the first year of the industrialization of humanoid robots. Cheetah focuses more on traditional wheeled robots. How do you think about the similarities, differences, and obstacles of the implementation of wheeled robots in humanoid robots in different scenarios? How do you plan the forms of Cheetah's future robot products? Yes, this wheeled machinery has become traditional human-related machinery. Actually, I think whether it's human-related or humanoid, the essence is to complete a certain type of work, right? Well, humanoid robots are more like humans and can attract more attention in terms of publicity. But in fact, most of our wheeled robots complete point-to-point deliveries indoors or use robotic arms on the mobile robot's bodies to complete some tasks. So I always believe that wheels can meet most of the mobile scenarios, because today, whether it's in factories, restaurants, including hotels, actually, the situation of steps is very rare. In many places, there are slopes, and these wheels can be fully realized. Then the cost of the wheels themselves is much lower than that of bipedal ones. Anyway, it has a cost advantage. Thirdly, today, although embodied intelligence is very popular when it comes to whether bipedal or humanoid robots of embodied intelligence can be practical. It should be said that a particularly clear time point or a specific video cannot be seen. Everyone just thinks there is an opportunity, mainly because Tesla made humanoid robots, so this industry has been driven. My view has always been that in most scenarios, there is wheeled movement, but some lifting robotic arms can be added later to complete the same work, so for the form and product form of our robots, we will not pursue humanoid robots at present, especially bipedal ones, but we will, in terms of the delivery and voice interaction of our machines, right? Do something like putting robotic arms in to complete some tasks, such as picking things up or doing some simple and small handling. I think this is within our plan, so I think maybe for a long time, we won't consider this bipedal thing, and I personally also think that in the entire industry, for a considerable or a very long time, right, movement must be mainly wheels. It will take at least several years for bipedal ones to be commercialized... Two, from the historical experience of SaaS, the willingness and ability of Chinese enterprises to pay are relatively low, which may affect the overall charging and profits of the large model industry. Based on practical experience, how do your clients consider the budget for large models? Where does the budget come from? In the current macro environment, do enterprises have any reduction in their investment in large models? How do you compare the monetization models and potential of different enterprise large model applications? A, if it is to help enterprises save costs, how large can the budget of enterprises be? B, if it is to help enterprises increase revenue, is it possible to charge based on effects, commission, et cetera? Is the ceiling really very high?... for such software. In fact, I think there may be many historical factors, but today, in such an environment in China, enterprises' pursuit of efficiency is increasing. Another is that because the large model itself is very new, so it's not something that especially non-tech enterprises can do by recruiting a few people for SaaS itself. I think personally that it emerged relatively late in China. Before the division of labor was formed, there were already relatively few software talents in China, so many enterprises could do it themselves, or some large companies did it and used it for free for everyone. But today, when it comes to the implementation of large models, as we have seen, it still requires a very deep integration with the actual scenarios of enterprises. It's not that you can take an API and have a few people do the application well. We are currently cooperating with several leading clients, and we have already seen such a trend, so professional teams are still needed to help them. Then to answer the two questions, if it helps enterprises save costs, we are currently discussing and negotiating the quotation plan with an enterprise, a large enterprise, which is to take a part of the saved cost as its budget. This client is very willing. It can really help him save a considerable amount of money. So taking a part of it as the budget is okay, including that, we have also promoted for some clients. We can disclose. But for some clients, we can disclose. For example, we have developed an AI application for cloud management on the cloud, which can help you analyze the idle rates of your cloud system and how many machines can be saved. We have run the accounts of several enterprises and can combine to save maybe 30%-40% of the cloud cost. This service is now very popular with this client. The charging model we have promoted is about 2.5% of his entire cloud usage as the software fee. Currently, several clients we are negotiating with have entered the deployment period, and they are willing to pay. That is to say, it's about the saved budget. Essentially, it's taking a share from the saved cost. Generally, enterprises are willing to do this. The second one, for increasing revenue, is it possible to charge based on effects or commission? We think it's also possible. We have worked with another enterprise to build a large model application for them. Later, their franchisees of the chain stores will purchase this service, and we will share the commission with them. This business model seems successful at present, is also okay. As for whether the ceiling is really very high, I think that for such enterprise applications, if you really talk about the huge technical difficulties themselves, I don't think there is such a thing, such a thing. We are doing applications ourselves. The real difficulty or ceiling of applications does not come from the technology itself, but from truly understanding the needs of enterprises, deeply doing a good job of the product points, meeting their real needs, and achieving a closed loop in the results. That is, being able to withstand their tests and achieving a truly positive reputation. This is one of our past advantages. So this is also what I think is the ceiling, because at this stage, how to say it? There are not many large model or application companies that are really willing to conduct in-depth research with enterprises and spend a lot of time. Everyone is still talking about the parameters of the model and such things. So I think this is an opportunity for us instead. Moreover, through conducting in-depth research with such leading enterprises, we believe that we can standardize this product, such applications into components and promote them to other enterprises or other industries, and AI has a very good opportunity in that the difficulty of crossing industries is much simpler than that of the previous steps. Previously, it also relied on a large amount of code alignment, so in a new environment, the amount of code is very large, but now AI has its own understanding and generation, so the amount of code will be much smaller. The key is to achieve a closed loop in the process and the experience. Okay, this is my answer. Three, the company has sufficient cash reserves on its books and is still generating net cash. How does the management plan to use these funds? Is there a privatization plan, or are there any plans for share buybacks or dividends? The question was raised. Indeed, as you said, we currently do have sufficient cash reserves and also generated net cash this quarter. However, we believe that actually, in the current overall economic environment, which is quite uncertain, it is particularly important for the company to maintain sufficient cash reserves, so we will continue to maintain a relatively cautious financial strategy. This can ensure that the company has sufficient flexibility and risk resistance ability when facing market fluctuations. At present, the company has no privatization plan because we currently believe that as Cheetah, maintaining the status of a listed company is helpful to enhance the company's transparency and governance labor. At the same time, we can also provide better liquidity for shareholders. Currently, the company also has no specific plans for share buybacks and dividends. If the board of directors approves relevant plans, we will make an announcement to the market as soon as possible. But we have seen robot companies collaborating with large model companies using the most advanced large models to accelerate product implementation. How does Cheetah consider cooperation with large model manufacturers, or is it more inclined towards the idea of end model integration? At this stage, it is certain that we will cooperate with large model manufacturers and use the most powerful models to enhance the intelligence level of robots, which is what we need to do now, because currently, in this entire AI industry, basically, it can be said that the supply of large models is excessive, and the cost of tokens has also dropped sharply now. So this kind of cooperation is very advantageous for our cost and for our rapid implementation. Our idea is that through a period of exploration, after making the product experience good, we will gradually see if it is necessary to train our own model for our end. In fact, we have already trained one, a 14 billion parameter model. We shall also release an Orion-7B model next month. But I think basically at this stage, for now, we still need to enhance the product implementation and product intelligence first. After obtaining sufficient experience, then come back and better configure the model for our end. Maybe in the long term, we will take the path of end model integration. That is, for our lease application or robot application, we will better train a model that can be implemented on the end. But for now, we still carry out some in-depth cooperation with large model manufacturers. Thank you. Five, we have seen many startup projects that are very similar to Cheetah's robot product form. How do you view Cheetah's competitive advantages in making robots? Where are the differences in our robot products? I think our advantages mainly lie in three aspects. The first aspect is that we still have sufficient technological accumulation. We have accumulated in the field of service robots for six or seven years. So our intelligence level of the robot, that is our current robot, whether it's the ability of interactive dialogue, or the autonomous navigation indoors, including the secondary development ability of this system, that is, our agents can take it and do very simple customization and development based on our system and super system. Currently, we are leading in the industry. This is already observable, and next, we will introduce our large model. That is, we'll launch the concept of a large model robot. Today, in the robot industry, there should be very few companies that have trained their own models. We are at least possibly the first one, or maybe one of them. But our team is not only responsible for the design and manufacturing of the robot hardware, but also for the software operating system. It now has the ability of large models, so our technological advantages are reflected in the just mentioned interaction ability, intelligence level, and secondary development ability. The second one is that actually, as everyone can see, there are many such robots. Through several years, we have truly launched a so-called productized robot, because a robot is a combination of many technologies, such as navigation technology, various electronic circuit technologies, and various mechanical structure technologies. So to make a sample, it might be relatively fast, but to really make the product quality stable enough to run in many scenarios and go through various quality inspections and user usage in terms of product capability, I think our current team's efficiency in this aspect is already much higher than before. Our ability to launch a new product or promote a product in a new scenario, it should be not easy for general small startups to compare. The third one is that through so many years of accumulation, we have initially explored the ability of commercialization or the establishment of channels. We have hundreds of agents in China and dozens of agents globally. The agents in some countries are quite significant. That is, they are quite large companies. We are cooperating to build this business network, enabling our products to quickly reach users through our channels. So today, we are now looking for users with a finished prototype, but we can inversely conduct research and development based on the needs of our users for the products they require. Including the lightweight delivery robot for factories that we just launched recently. It was also the feedback from our agents and our channels that prompted us to conduct the research and development. So, I think these several points have basically formed a closed loop for our Cheetah robots today. That is a closed loop from research and development to sales to this, this user experience. I think this closed loop is our real differentiating point, if not at a certain technical point, because others can also do that. But the establishment of this closed loop, especially for a channel network like To-B, the cycle will not be too short. So I'm not too worried about the competition from such startup projects. Okay, thanks. Six, in terms of data, after our robot products are exported overseas, how is the data generated attributed? Can we still use this data to continuously iterate the models and products? In the process of going overseas, our general principle is to handle it accordingly based on the legal requirements of different regions and their laws and regulations. Then, for the collection of this data, it is stored on the AWS server overseas and undergoes necessary storage, encryption, security reviews, and strict access control measures. We also completed the ISO 27001, the Information Security Management System certification, ISO 27701 Privacy Information Management System certification, and ISO 42001 Artificial Intelligence Management System certification, with reference to the ISO certification standards. We are the first company in Asia to obtain this Artificial Intelligence Management System certification. So the establishment of these systems ensures that our data usage complies with international standard security certifications. In this regard, because Cheetah went overseas early, we actually encountered this problem before. We also encountered these problems when our apps went overseas earliest, including the GDPR in Europe and other regulations. We attach great importance to such compliance and also have such experience inheritance. Then, in terms of the usage principle you just asked about, we definitely follow the principle of necessity of use and minimization of data. That is, to ensure the analysis of machine operation problems. This does not involve non-essential personal information. Then, basically, that is not in many places. For example, in some countries, we will turn off some functions just to achieve some basic functions. So overall, we are very cautious in the usage of this data and comply with local laws and regulations. While another question is whether this data can be used in reverse to train the large models. Actually, today, the capabilities of the real large language model like Pi and those provided by iFlytek are already very good. We do not need the training of this data for the time being. And for things like navigation, actually, it does not involve this or most of them do not involve this part of AI data training. So in general, today, we attach great importance to overseas data security and are also very compliant. Thank you. Eight. My question is about organizational capabilities, because in the era of large models, starting a business is not only about technology and product entrepreneurship, but also requires organizational innovation. Especially when enterprises are developing products, many variables need to be considered, such as models, data, customers, end users, et cetera. In your opinion, since Orion was integrated into the listed company during this period of more than half a year, what significant results has Cheetah Orion achieved in organizational innovation? How does Cheetah recruit the best talent? How much time do you spend on talent recruit? Okay, let me start with this. Frankly speaking, our organizational innovation is not particularly innovative because the building of organizational capabilities is a long-term and routine work. Especially, as you mentioned, after we acquired Orion, the entire Cheetah has done many things in organizational reform. In fact, today, Cheetah has gradually transformed into a company with a core focus on the To-B business, whether it's the sales of robots, our Juliang, or our international advertising agency businesses, right? They are all centered around To-B sales. So in terms of organizational capabilities, the first thing is to strengthen training. I myself participated in training this year in various sales trainings organized by the company. We have put in a lot of effort and also spent quite a lot of money. Tomorrow, we still have training. The second is in sales management. This was an area where Cheetah was not good at before, but since 2017 I myself have started, called the No One position in sales. So in the management of the sales system, we also learned from Huawei's LTC iron triangle approach, and learned from many To-B enterprises, and made it an important change in our organization. In the implementation of the one hundred and seventy one approach for the entire To-B sales team in AI, we first applied it to the management of our own To-B sales for various process management, daily reports management, and weekly report management.... In this regard, we should have done a lot of work. The reason why it's hard to say it's innovative is because I think To-B itself has a set of mature methodologies and some organizational references. We are learning from and implementing more advanced models. One obvious change is that today, for some so-called key accounts or CA customers, our ability to truly form cooperation or do business has definitely strengthened compared to before. This year, in different businesses, we have reached certain agreements with many key accounts or industry-leading customers. Let's give an example. Yes, like the hotels, and for some mobile phone manufacturers, we also have cooperation with leading mobile phone manufacturers in AI. For some automobile manufacturers, which are also leading ones, we have cooperation with them on the cloud. These should all be some changes brought about by organizational reform. This, of course, I have to say that it's not particularly easy to shift from To-C to To-B. We have explored for a long time, spent a lot of effort, and taken some detours. Now it seems that this entire set of To B sales management system, including the integration of industry and sales, has made considerable progress compared to before. Regarding recruitment, I myself also spend time. Every week, I have some interviews or something related to recruitment. Of course, I have to say that today, because we have spent several years building this to B sales system, we don't have such a big talent gap currently. We mainly focus on finding some excellent to B sales talent and putting in some effort. Overall, it also depends on internal training. We may attach great importance to the cultivation of young people in addition to the recruitment of top talent. This year, we also restarted our campus recruitment program. The entire company spends a lot of time training campus recruits, and I myself will not give lectures to them. So, so this is our current situation regarding organizational management and reform. Thank you. Based on practical experience, which model applications have better implementation effects and meet enterprise demands? At this stage, how far are these applications from being truly useful? Historical experience shows that Chinese companies are better at developing applications, but if the underlying large models in China cannot reach a particularly intelligent level, can the effects of the applications based on them be comparable to those overseas? The first one is the application effects of today's to B large models. In fact, the effects of the first wave of to B large model usages, like text-to-image and text-to-video. We have all seen that companies in this field have grown rapidly. It's actually because there is a large demand for creating graphics in writing text. This is the simplest application. Now, I think it's entering a stage. For example, what we are doing with the client is an internal training system for them. All new employees need training, assessment, and some practical scenarios. Previously, you could only read documents and answer questions. Now, you can interact with an AI, right? This is a customer, and then conduct this aspect of training and assessment. Currently, the effect seems very good, and users are also quite satisfied. So this is one type, that is internal training, which is related to training and document content skills. This aspect has a considerable market within enterprises for the training and assessment of new employee skills. I think AI seems capable of achieving this at present. And, you said truly useful. Well, large models cannot be simply used randomly as they are, because large models have hallucinations and the understanding of the internal knowledge of the enterprises, right? The requirements of the enterprises themselves. It requires building a product system to make it truly useful. So this actually takes a lot of effort, so I think it will take some time to make the experience good enough for this progress. The second one is actually some data organization types, which are related to buy. That is the traditional buy, when combined with AI, the effects will also be much better. Previously, for every buy requirement, a large amount of code had to be written, and then it needed to be changed according to the application requirements of the enterprise. What I just mentioned, like the saving of cloud costs, is such an example. Using large models to understand the intent and write good code, I see that some of our clients are also implementing it. For the second question you mentioned, when the underlying large model level is not particularly intelligent. Actually, I think today the level of Chinese large models has a gap compared to the top-level foreign models, but the gap is not that large. In many scenarios, ourselves, we usually use foreign models in the experimental stage. To conduct experiments first, and after finding it feasible, we will switch to domestic models to provide services, because this is the requirement in China. But in fact, after replacing with domestic models, there will not be a significant difference in the final performance compared to the foreign ones. Basically, there is no essential difference in usage, because in many enterprise scenarios, you do not need to be as intelligent as we imagine, but really clarify the demand. Today, there is a very popular term called agent. The essence of agent is to write something for this type of demand. For example, let the large model find mistakes by itself, think by itself, and then combine some traditional technologies. This, call code to make adjustments, right? To do this, to do this matching. In this way, this combination can achieve... Actually, the level in real applications is very good. I have full confidence in the effect of Chinese companies' applications compared to overseas ones, because overseas companies definitely have advantages in algorithms. But when it comes to engineering those details in application development and really fulfilling a user's demand in detail, I think Chinese companies have great advantages now. So for real application development, I think the effects will be good. It will not be limited by the so-called underlying large model level, right? By the way, just mention this. Today, you can see that OpenAI released a new version of o1. It seems that through o1, the information collection and usage, it is not an improvement in the model's ability. They wrote something like math problems as a... That is, they wrote some internal code specifically for this scenario, which is equivalent to application optimization. So when answering math problems, the internal model will have various things like debate, self-reflection, and error refining. So the performance has improved a lot at this level, but it does not involve the improvement of the underlying model. The example I gave wants to say that, the so-called intelligence level of the model, it is not, it is not that only when its intelligence level improves, the entire application level will be particularly good. As long as the basic ability of the large model is not too bad, it can be achieved through continuous adjustments in the application. In this case, in the user demand scenario can be done quite well, as we can see now. Okay, thanks. The last question: From the perspective of research and development, talk about Cheetah's competitive advantages and R&D plans. Okay, thank you for your question. Let me repeat. I think this way, when we started to invest in the AI strategy, we also experienced the first wave of AI, which was very hot at that time, up to today's AI in 2.0. My current thinking is more inclined to be more pragmatic and to advance the technological updates step by step, in a refined way, instead of relying on a product through a super technology and then somehow sweep the market. I now think that robots, because it is a rather complex system, it is a more complex system than a mobile phone. So it involves the integration of various technologies such as hardware, software, artificial intelligence, and it operates in different scenarios.... Right? The scenario is completely different from when a user takes out a mobile phone from their pocket. Sometimes it runs in a restaurant, in a hotel, at the front desk, or in a library. Whether it is your navigation, voice processing, or the experience of the applications inside, they are all different. So I think today, the real advantage in promoting this research and development does not lie in a certain technical point, but in whether research and development can form an efficient linkage with the sales chain and end customers. Ah, this is what I really think today. That is to say, today, everyone has different ideas about robots, and some people are making humanoid ones, but our focus is on implementation. After we implement and run it in a scenario for the first batch, there will definitely be problems. After these problems occur or some problems are encountered, we can enable the R&D team to respond quickly. Then, when you solve one problem after another, at this time, your engineering capability and the efficiency of your team are your advantages, right? Today, if you single out a certain technical point and say that we can do this technical point and others can't, I actually think it is very difficult to achieve this in today's business competition and industrial environment. What is compared is the closed loop of an enterprise that is from the terminal to the back to the product, and the product can be quickly pushed to your customers. This is the ability of such a closed loop. So this is what I think we have explored through these years in R&D. If there is an advantage, this is the advantage. It's a bit like what Amazon said, called customer first, find problems from customers, and then find innovations among customers. Sometimes we also made this kind of mistake before when thinking about something. We thought everything was fine and the product was good, but it turned out that when you reach the users, we found that, right, the concerns of a restaurant owner or a hotel owner may be quite different from ours. So for this thing, right, we can only establish such a system to ensure the realization of our own R&D in the future process. University, this is the first one. Secondly, my view on new technologies today is like this: I think for new technologies, including artificial intelligence, large model technology in today's embodied intelligence, we will definitely follow. We will definitely understand what is happening inside, how to do it. Even now, we have added robotic arms to robots to conduct some experiments in the aspect of embodied intelligence. But I think for a company of our size, we cannot be particularly advanced, and there is no need to be advanced, right? Sometimes the trial and error costs of one time may be huge. So we follow new technologies and gradually apply them in our products. What we are doing on robots now is to turn the large model into one. We call it the core of the large model robot, which is to make the large model decision making. This framing is also being done step by step and will not be completed with a particularly large framework. Combined with specific user scenarios, yes, but follow new technologies and ensure the high-speed iteration of our entire R&D and sales system. Pay high attention to the problems of real users and respond quickly to feedback to our entire R&D capability. I think this is what we have today. What you think can be regarded as an advantage. So thank you all for joining our earnings conference call. This is Helen from the company IR team. This is the line for the translation purpose, and we will have the transcript in English being available as soon as we can. Thank you so much for participants. Bye. The conference has now concluded. Thank you for attending today's presentation, and you may now disconnect your lines.