news

yesterday, at the 2024 tencent global digital ecosystem conference, tencent group senior executive vice president and ceo of cloud and smart industries group tang daosheng accepted a group interview with reporters. tang daosheng said that the big model is still in its early stages and needs to be continuously iterated to meet user needs. he was satisfied with the results of the internal adjustments, which allowed the team to better understand the direction of their efforts.

the reporters' questions mainly focused on large models, followed by to b business, and a small amount on overseas expansion and hardware. tang daosheng answered them one by one in detail.

large model

tang daosheng said that the internet had a bubble at the beginning, and many companies were eliminated, but some companies stood out. the same will be true for ai big models.

tencent attaches great importance to investment in large models, and also to sustainable development. the investment required for large models is much higher than in the past. each business should calculate the cost clearly and set a reasonable price to avoid subsidizing its losses with other profits. "when i do any business, i will think: what is a reasonable business model? what kind of input and output should be required for long-term healthy development? in general, we are a long-term company, and we will persist when we see the right opportunity. i believe that in the end, those companies that can persist and pursue long-termism will be rewarded, rather than just investing in it because the concept is hot. it is difficult for such an approach to fail to persist."

tang daosheng believes that the public is currently unwilling to pay for the big model because it is not yet ready. if it is good enough, they will be willing to pay. the performance of the big model may be 60 or 70 points now, but everyone expects it to be 90 points. in the process of exploration, some new scenarios and new opportunities will be found, avoiding some scenarios that were originally conceived to be beautiful but have limitations in actual implementation.

when ai was popular five or six years ago, many innovative companies invested and won bids. a similar situation is happening now, with some companies competing for bids without necessarily having the capabilities. tencent's previous experience is that customer needs may not be so clear, and there are very personalized needs, so continued investment may overwhelm the company.

tang daosheng is confident in the industrial application prospects of ai capabilities and needs to find user pain points and meet user needs.

tencent's advantage lies in scenario applications, such as quick generation of meeting minutes, ai code assistant, and tencent lexiang. the ai ​​code assistant launched by tencent has achieved coverage of more than 50% of development staff within the tencent group, with a code generation rate of more than 30%, and an average coding time reduction of more than 40%. combined with internal large-scale production experience, r&d efficiency has increased by more than 20%.

he believes that aigc's national-level applications may first be produced in the to c field, and the closest one may be in the information search field, such as yuanbao.

more than just big models

tang daosheng believes that ai is not just about big models or big language models. big models are only part of the big ai track. many other technical routes in the field of artificial intelligence are also worthy of attention. to build a useful intelligent system, big models may be just one of the modules. it is not only players who make big models who are engaged in ai. this is as narrow-minded as thinking that only companies that make mobile phones are important in the mobile era.

image generation, video generation, and 3d model generation are all generation-based. many recognition-based technologies, such as face recognition and the recently popular palm print recognition, have been widely used in the fields of identity authentication and security recognition. ocr document recognition combined with large models has also greatly improved its capabilities.

ai is a very broad technology, ranging from recognition to generation, as well as aig and alphago's reinforcement learning. ai technology also has many different directions, and will flourish in the future. tencent will continue to invest.

to find solutions and solve problems, big models are just one of the ways. although big models have advantages, they also have shortcomings, such as it is difficult to avoid the emergence of "illusions". therefore, the rag (retrieval enhancement) model emerged later, which makes good use of the internal data of the enterprise as the factual basis for answering questions, and then combines the model to understand and give more accurate or more targeted answers.

many companies accumulate various proprietary information. by selecting a basic model and superimposing search enhancement to generate a rag architecture, tencent can create some scenario-based intelligent assistants and intelligent applications to significantly improve efficiency and effectiveness in areas such as customer service, marketing, and content creation.

to b business

tang daosheng introduced that a few years ago, everyone was full of expectations for the saas industry. without careful accounting, they made huge investments, hired many people, and deployed a lot of manpower to do various customized development projects for customers. in the end, they found that the business model was not very tenable. now everyone expects a correction, and the team is reducing some one-time customized development investment and human service investment, and investing more resources in consolidating platform products. the entire industry is also more or less experiencing similar adjustments and changes.

in the future, the industry will return to its commercial essence, and some companies that fail to transform in time will be eliminated in the market. "but i also believe that this will also retain a group of more 'robust' players, and at the same time have a better integration with the entire industry ecosystem, with stronger capabilities and competitiveness. for example, we who make platform products will focus on making platforms; some who make industry applications will go deeper in their own fields; and some service providers will focus on serving customer scenarios and providing operational services. for example, in the financial industry, we have had in-depth cooperation with changliang, shenma, etc. for a long time. they provide the capabilities required by the core banking system or various financial industries, and also need platform software, such as our tdsql database, or some of our audio and video capabilities. as the positioning of each player becomes clearer, everyone's boundaries are also becoming clearer. as everyone cooperates, the roles they play become clearer, and this industry will become more and more mature."

tang daosheng said that he was quite satisfied with the results of the changes in the past two years. the business is more focused, the team knows its shortcomings and will improve in more important areas. they also understand which problems can be solved by tencent products and solutions and which ones need to be solved together with partners. the path is becoming clearer and clearer.

tang daosheng said that tencent cloud has made great progress in reducing losses compared to last year and is gradually approaching its profit target. profitability will not be too far away. in recent years, through continuous improvement of supply chain efficiency and by doing what is right and what is not right in the operation process, the business has become more focused and healthier.