cloud vendors re-recognize the capabilities of large models, tencent's tang daosheng said "be more patient"
2024-09-08
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
a year ago, tencent released the hunyuan big model at the global digital ecosystem conference. a year later, the "100-model war" has come to an end, and the pattern of giants and unicorns has been initially determined, but the market's expectations for big models seem to have declined, and the industry is re-recognizing the capabilities of big models.
in a recent conversation with yicai global and other media, tencent group senior executive vice president and ceo of cloud and smart industries group tang daosheng answered questions about the decline in the popularity of the ai market and said that it was just like the cycles of new technologies in the past. at the beginning, everyone rushed in and invested, even creating bubbles, and then found that new technological changes needed time to settle, and even had to wait until the first wave of capital-driven, not very professional players were eliminated before returning to a rational and pragmatic state. "it's the same with ai big models now. everyone may have high expectations at the beginning because they haven't experienced the polishing time. (now) the pendulum has swung to the other direction. this is my feeling."
in the early stages of the development of big models, the cloud and big models were very close, and computing power consumption reflected the demand and popularity of big models. li qiang, vice president of tencent group and president of government business, told china business network that the emergence of big models in the past two years has brought about a huge increase in gpu computing power and the rapid growth of products related to supporting big model training. but from the application side, the commercialization of big models on the to b side is far less prosperous than everyone thinks. relatively speaking, big models are currently better used in scenarios with relatively high fault tolerance. tencent is still providing flexible and diverse adaptation solutions for big model services for enterprise customers, and exploring the combination of big models and cloud products.
gpu computing power still accounts for the bulk of ai revenue
tencent's latest financial report shows that in the second quarter of this year, tencent's financial technology and enterprise service revenue increased by 4% year-on-year, of which the enterprise service business increased by more than ten percentage points year-on-year. tencent cloud's customers include more than 80% of the top large model manufacturers in china. however, this tencent financial report does not talk much about the changes brought about by ai.
as for how much of tencent cloud's revenue comes from ai, tang daosheng told the first financial reporter that the data may be difficult to quantify now, but it is increasing. for example, the cooperating autonomous driving manufacturers are continuously increasing their investment in model training based on vehicle perception data. in terms of revenue structure, li qiang told reporters that the majority of tencent's ai-related revenue is still gpu computing power.
tencent not only provides computing power to large model manufacturers, but also promotes the commercial use of its self-developed hunyuan large model. after the initial market enthusiasm for large models declined, tencent is also evaluating the actual increase that these two businesses can bring. in a previous interview, tang daosheng said that if the large model enterprise invested in is successful, it will require continuous cloud consumption, which is an excellent source of income for the cloud business. but he also said that in the early days of new technology, a large amount of capital drives the wild growth of startups, which may lead to over-investment, and many players may be part of the bubble. "if cloud revenue is too dependent on capital-driven startups to consume, once the bubble bursts, some customers will disappear, performance will drop, and it will be painful when it drops."
in addition to the gpu computing power consumption brought by large models, cloud vendors have also experienced a new understanding of the capabilities of large models in promoting the b-side commercial use of large models.
li qiang said that from the overall market perspective, the proportion of ai-related revenue that actually comes from the commercial output of the big model itself is still relatively low. he described this process as the market moving from "fanatical" to "rational." specifically, the barriers to b-side industries are relatively deep, and the application of ai in the industrial field and traditional industries is more complicated than to c, while the big model has not yet met the requirements of traditional industries. for example, the industry has more stringent requirements for fault tolerance.
li qiang explained that key applications or links that affect safe production and important decisions are less likely to accept unexpected situations. at this time, it would be better if ai can assist in decision-making. in some traditional niche areas, general large models may not be the best choice. for example, strict iphone mobile phone quality inspection requires taking photos and enlarging them more than a hundred times. it is meaningless to use large models trained with general knowledge in such scenarios. the efficiency and cost are not as good as small industry models. large models may not be applicable in all niche industries and fields. just like there is no need to train children to become undergraduates from prestigious schools and then put them in a position dedicated to screwing screws.
in addition, large models, especially those with huge parameters, need to be trained on large computing clusters. manufacturers that use these large model capabilities also rely on large computing clusters to provide reasoning. previously, this was considered an opportunity for cloud vendors to develop public clouds. however, b-side enterprises still have concerns about data security, which makes the process of providing large model capabilities based on public clouds not as smooth as expected.
"domestic companies have relatively high requirements for the confidentiality of their own industry data, and those with core businesses are more willing to do so in the form of private deployment. however, the privatization route will affect the integration of big models with the industry. in a sense, this has formed today's bottleneck." li qiang said that the current deployment methods for big models to b include both private deployment and api access, and when it comes to core applications, the industry is more considering privatization.
where are the big models really used?
in addition to tencent, some major technology companies are also betting on ai, and some of them mentioned the increase in revenue from ai in their latest financial reports. for example, in the quarter ending at the end of june 2024, alibaba cloud's ai-related product revenue achieved triple-digit growth, and baidu smart cloud's revenue increased by 14% year-on-year in the second quarter of this year, of which ai contributed 9% to baidu smart cloud's revenue. among international cloud vendors, google's cloud business revenue increased by 29% year-on-year in the second quarter of this year, and cloud revenue was boosted by ai demand. in the latest quarter, microsoft azure and other cloud service revenue increased by 29% year-on-year, and ai contributed 8% to azure's revenue growth.
judging from the data disclosed by some manufacturers, the revenue growth rate brought by ai is in the single digits, or the proportion of ai in cloud revenue is in the single digits. compared with the market's previous huge expectations for ai, the boost of ai to cloud revenue seems to need to be further strengthened.
li qiang told reporters that customers are gradually realizing that big models cannot "cure all diseases" and are becoming more rational in their choice of scenarios. compared with industrial production scenarios with strict fault tolerance, big models have better application space in other scenarios, including knowledge management, marketing, customer service, code, intelligent risk control, and areas with relatively low professional requirements, such as field inspection scenarios. customers will choose more high-level application scenarios to cooperate. for example, the medical big model that tencent cooperated with zhongshan hospital is advancing in medical diagnosis and diagnosis book writing, which is considered auxiliary diagnosis and treatment. in decision-making scenarios, big models are more used to assist decision-making. in knowledge-related scenarios, big models are more commonly used in customer service and employee training scenarios.
regarding the challenges of large models, tang daosheng told reporters that this includes both the challenge of relatively scarce high-quality public data and the challenge of large model implementation. the implementation of large models involves a series of issues such as data confidentiality, implementation costs, accurate results, and scenario selection. in addition, another challenge is that the industry is prone to zero-sum games when the environment is stressful. he believes that if everyone's anxiety becomes stronger, it will not be a healthy state to maintain market share by making losses.
"i watch the income statement very closely. every business should calculate the cost clearly and set reasonable prices to avoid using other people's profits to subsidize its own losses. everyone needs to be more patient. today's technology may only achieve 50 or 60 points in some scenarios. it takes time to reach 90 points. at first, many people believed that models could quickly change the world, but recently some people have become pessimistic and feel that large models look good but are not easy to use. in fact, 'overestimating progress in the short term and underestimating results in the long term' are both undesirable," said tang daosheng.
(this article comes from china business network)
