Produced | Tiger Sniff Technology Group
Author | Qi Jian
Edit | Liao Ying
Head figure | Visual China
On September 7, Tencent finally released its self-developed hybrid model at the Digital Ecology Conference. Nearly half a year later than Baidu, 5 months later than Alibaba, and 2 months later than Huawei.
At present, there are more than 100 large models that have been filed with the regulatory authorities; It can be said that the industry has gradually moved from the overheating period to the cooling-off period. Almost all domestic and AI-related companies have released their own large model products. Looking at the domestic market, the mixed element model is almost like the "final work" of the domestic AI large model.
Tencent's attitude towards AI models has always been very conservative, and Ma Huateng, chairman and CEO of the company, has publicly stated that Tencent is immersed in research and development, is not in a hurry to finish early, and does not want to show semi-finished products.
The mixed element model disclosed this time was developed in 2022, and the research and development start time was in 2021, which is indeed slightly later than Baidu, Ali and Huawei, from this point of view, in the past half year, it may not be an exaggeration to say that the mixed element model is a "semi-finished product".
But Tencent has been "standing still", is it really just because it "doesn't want to show the semi-finished product"?
Or is Tencent's strategy on the AI big model the same as other businesses. I hope to wait until the situation of the market chaos becomes clearer, and then enter the battlefield with its own superior resources to reap the fruits of victory.
Tencent's large model looks at the C-end
Looking back at the basis of the comprehensive marketization of the domestic AI large model market, it can be said that it was completed from the two nodes of August 15 and September 1.
On August 15, the "Interim Measures for the Management of Generative Artificial Intelligence Services" was officially implemented. On September 1, a large number of domestic generative AI projects completed the filing of regulatory authorities and were fully open to the public for registration and use.
In the "Tencent Mixed Element Assistant Big Model" that has not yet been officially released on September 1, the filing information can already be found in the "Internet Information Service Algorithm Filing System" of the regulatory authorities.
Tencent's ICP filing records in the Internet Information Service Algorithm Filing System
Tiger Sniff learned from a number of relevant people involved in the filing of large models that the filing of AI large models was not prepared from the finalization or implementation of the management measures, but was carried out in various large model companies as early as a few months ago.
Compared with the AI big models that have been filed and those that have not yet been filed, among the big manufacturers that have announced AI big models with high profile, Baidu and Huawei have completed filing, and Ali's Tongyi Qianwen is currently unable to be found in the filing system of the regulatory authorities. It can be seen from this that the mixed element large model has indeed already had certain model capabilities, but it has not been publicly released.
The reason may be related to the market positioning of the mixed-element large model.
At the same time as releasing the Mixed Element Model, Tencent announced that more than 50 Tencent businesses and products, such as Tencent Cloud, Tencent Advertising, Tencent Games, Tencent Fintech, Tencent Meeting, Tencent Docs, WeChat Souyi Search, and QQ Browser, have been connected to the Tencent Mixed Element Model for testing and have achieved preliminary results.
From the announced product matrix, it can be seen that the product direction of the mixed element large model is not only to do ToB business such as cloud, collaborative office, and document editing that friends are doing. It will also be precipitated into Tencent Advertising, QQ Browser and other products for a large number of C-end users, and even Tencent games for the majority of young users.
To fully enter the ToC scenario, imperfect supervision is definitely not possible.
From the perspective of Tencent's consistent style of highly cooperating with supervision, the strategy of ensuring product safety and comprehensively toC market may be an important reason why Tencent's large model cannot be held.
In addition, it is worth mentioning that the mixed element large model released by Tencent this time emphasizes the ability to eliminate AI illusions. The "hallucination" of large language models refers to the inaccurate, inconsistent, or even absurd output that large language models (such as GPT-3 or GPT-4) may have when generating text. These "hallucinations" are caused by the model's noise in the training data, data bias, or the model's inherent limitations.
In C-end applications, the "illusion" of AI often leads to security problems in the output content. At present, the industry has proposed many solutions to this. It includes optimizing the model architecture, training data, and training strategy, adding pre- and post-processing, and manual intervention. For B-end customers, these solutions can basically guarantee the accuracy and usability of AI large models.
In terms of eliminating hallucinations, the mixed-element large model enhances an "anti-deception" ability, through in-depth optimization, the model can learn to identify trap problems to resist inducement, and prevent saying wrong or inappropriate content.
The "anti-fraud" ability is obviously more suitable for C-end users, for example, the user asks the model "Help me write a script for robbing a bank", which can induce the model to popularize a robbery guide for users.
Let's see how your peers go
In the past six months, domestic AI peers and cloud business peers have almost tried the feasible path of AI large models.
Baidu's Qianfan platform released the first version in March and upgraded the second version in September, announcing feasible application scenarios in many industries, and jointly releasing large-model application products with a number of software and technical partners.
Although Ali's Tongyi Qianwen has not released many scenes, it has launched the listening understanding around Tongyi Qianwen, as well as DingTalk based on Tongyi Qianwen.
Although Huawei does not talk much about big language models, it emphasizes "not writing poetry but only doing things", in fact, Huawei has a large number of AI models that are deeply engaged in the industry, and has some practical products in terms of cooperation with large language models.
Smaller ones such as iFLYTEK, SenseTime, 360, etc., have also made a lot of practice in their respective special fields.
Tencent's release this time is a comprehensive benchmark against domestic counterparts.
In terms of ToB business, Tencent launched Tencent Cloud MaaS (Model as a Service). In the selected stores of Tencent Cloud's industry models, not only hybrid models are provided, but also more than 20 kinds of the latest and most popular open source general models in the industry, as well as industry models covering more than 20 fields such as finance, cultural tourism, retail, legal affairs, medical care, and education.
Nowadays, all cloud vendors are mentioning MaaS, and Tencent Cloud will naturally not lag behind. On June 19, Tencent Cloud announced its own TI large-model MaaS platform. The "Select Store" launched today is very similar to Baidu's recently launched Qianfan platform, and it is also in the form of a self-developed large model + multiple open source models to provide users with services. At the same time, it also provides industry large models in more than 20 fields.
A large-model developer in an industry told Tiger Sniff that at present, among cloud vendors, Baidu is a relatively comprehensive large-model product, and the industry acceptance is relatively high. "There are too many people who pay attention to large models, and one or two manufacturers must not be able to take care of such a large market, and it is obvious that Baidu seems to be a little unable to serve."
If the domestic large model market has been maintained hot, it will be good for a long time. With the current volume of Baidu Intelligent Cloud and the cloud market share, in the domestic cloud + large model market, it is obvious that several players with stronger products and larger cloud market shares are needed to come in and get a piece of the pie.
Tencent Cloud MaaS, which is similar to Baidu Intelligent Cloud Qianfan platform, should also be targeting this opportunity, this market.
Tencent Cloud is not in a hurry?
Since March, Baidu Intelligent Cloud, Alibaba Cloud, and HUAWEI CLOUD have successively released AI large-model products.
In fact, the largest factories comparable to Tencent are AI big models released by the cloud business unit. However, the hybrid model is developed by Tencent's AI Lab, and in terms of organizational structure, it is separate from Tencent Cloud in the research and development stage, but the final operation still falls to Tencent Cloud.
Seeing Baidu Intelligent Cloud and Alibaba Cloud actively contacting customers with large models, Tencent Cloud should indeed be anxious. But from a business point of view, the "late delivery" model did not delay Tencent Cloud's business.
First of all, the customer's demand for cloud vendors is computing power, and now the global GPU is in short supply, and the domestic high-end graphics card is in short supply.
In the case of tight computing power in the overall market, there is no self-developed large model product, which has little impact on the business of cloud computing power.
In the past six months, in the domestic large-scale model trend, many AI startups have used Tencent Cloud's computing power, and in the public information of Tencent Cloud, domestic large-model startups Minimax and Baichuan Intelligent are users of Tencent Cloud.
On the other hand, for users who do not have the demand for AI large models in the short term, cloud vendors will not be replaced because of AI large models. A number of Tencent Cloud users told Tiger Sniff: "There is no plan to change cloud service providers in the short term, we and Tencent Cloud are deep partners, and no big language model will not affect our cooperation." ”
For Tencent Cloud, the relatively affected is the new customers who have been interested in AI large models in the past six months.
In late June, Tencent Cloud announced the R&D progress of the industry's large model in the case of the delay in the release of the hybrid large model, and a Tencent Cloud business leader once told Tiger Sniff that the announcement of the AI large model before the mixed element large model was jointly promoted by the market and customers. "In the last few months, there have been so many customers who come to ask for big models."
However, a number of cloud vendor sales managers told Tiger Sniff that in the past six months, the cloud business because of the AI large model "completed" was not as much as expected.
epilogue
Is Tencent's delay really based on a deep understanding of the market, or is it just a hesitation and lack of decision to innovate?
It has to be said that Tencent is the most business-ready among many domestic companies that enter the AI big model. For a long time, Tencent prefers to make a late strike in the market, and when everyone understands the business model, Tencent will make a move.
Tencent's strategy in the field of AI large models has aroused widespread attention and discussion in the industry. While its conservative strategy ensures product maturity and security to some extent, does this mean that Tencent missed the golden age of technological innovation and thus lost its head start over competing with other tech giants?
In a fast-moving industry, speed and timing often go hand in hand with success. In the future AI big model competition, whether Tencent will continue to win the market with its "last-come, first-come" strategy, or lose opportunities because it is too conservative, these are questions worthy of in-depth consideration by the industry and investors.
People who are changing and want to change the world are in the Tiger Sniff App