Tencent Hunyuan large model was officially unveiled, and we took the lead in trying its productivity

The first batch of domestic large-scale model registrations were approved last week and services began to be opened to the whole society, marking that large-scale models have entered a new stage of large-scale application. Among companies that have released applications before, some technology giants seem to have not yet taken action

On September 7, 2023, Tencent officially unveiled the Hunyuan large model at the Tencent Global Digital Ecology Conference and opened it to the outside world Tencent Cloud

As a large model with over 100 billion parameters, Hunyuan uses more than two trillion tokens in pre-training corpus. It has gained strong Chinese creation capabilities and complex context by virtue of a number of unique technical capabilities. Excellent logical reasoning ability and reliable task execution ability.

Jiang Jie, Vice President of Tencent Group, said: "Tencent Hunyuan's large model was trained from scratch starting from the first token. We have mastered everything from model algorithms to machine learning. framework, and then to the full-link self-developed technology of AI infrastructure."

Open the large model, everything is productivity

Tencent has always stated that it is related to the large model The direction has already been laid out, and special research has been progressing in an orderly manner.

What is the level of this large model if it's not "new technology"? At the conference, Jiang Jie revealed some basic information by directly asking the Hunyuan model. Its parameter volume has reached hundreds of billions, and the data used for training is as of July this year. In addition, Tencent also stated that the knowledge of the large model will be updated monthly

The contents displayed on the scene include the Tencent Hunyuan large model applet, the AI ​​assistant in Tencent documents, and the functions of the Tencent conference AI assistant

This site obtained the test qualification for the first time and tried it. The first was the WeChat applet.

When we entered the mini program, we were pleasantly surprised to find that the content was no less than other large applications. Here, we can find some inspiration and see what functions Hunyuan can provide

From productivity, life, entertainment to programming development, its open capabilities can be said to be very comprehensive , consistent with the identity of a 100-billion-level large model. So can Hunyuan really complete these tasks effectively?

I need to prepare a PPT. I have decided on the topic, but I don’t know where to start. I asked the Hunyuan model a question, and it only took a few seconds for the AI ​​to give an outline containing seven parts. Each part also included the key points of subdivision

Enter the abstract and introductory part of a paper "RLAIF: Scaling Reinforcement Learning from Human Feedback with AI Feedback" submitted by Google on arXiv in September. It is several long paragraphs. Many large models do not support so much input content at all. It is confusing. The Yuanda model was directly summarized and translated into Chinese.

Probably means that artificial intelligence can replace the role of human reinforcement learning feedback (RLHF) in large-scale model training

A technology that has reached the practical stage Large models can help us write code. Now let's give the AI ​​a piece of code and let it explain the content that is incomprehensible and not clearly commented:

It explains in detail the meaning of the numbers in the square root reciprocal algorithm (but I don't understand the annotations very well). Maybe it won’t be long before we can’t do development without large models.

Then there are Tencent documents. Many people have used large model tools such as GPT-4 in their own workflows, and the Hunyuan large model has been used in the intelligent assistant function launched by Tencent Documents. Create a new smart document on the PC and enter "/" to realize content generation, translation, polishing and other operations according to your needs.

Then enter the natural language command, and the large model’s generation ability can help you summarize the long text in Tencent documents:

It seems to be very useful when writing a paper

Of course, if you give a topic, it can create text, and then you select a part of the generated content, and AI can also do more Sub-refined editing. After writing, you can also translate it with one click:

In addition, table data calculation and chart generation are all done in one sentence.

#These functions are currently in the internal testing stage and will be open to users when mature.

In Tencent meetings, the application of Hunyuan large model can help you no longer be distracted during the meeting. For example, you can ask the AI ​​assistant at any time what was just said, or what the argument between those two people was about. AI can quietly summarize the content into a few short sentences and clearly list one, two or three

Of course, after the meeting, the Hunyuan large model can also Summarize meeting content more quickly and comprehensively, and mark to-do items

Has covered more than 50 businesses of Tencent

Jiang Jie summarized the Hunyuan model Three major characteristics: strong Chinese creative ability, logical reasoning ability in complex contexts, and reliable task execution ability.

Currently, many large models in the industry still have limited application in scenarios. The main problem is that they have high fault tolerance rates and are only suitable for casual scenarios with simple tasks. Tencent has conducted a series of self-research innovations at the algorithm level to improve model reliability and maturity.

Tencent Group Vice President Jiang Jie appeared at the event

In view of the problem that large models are prone to "talk nonsense", Tencent The pre-training algorithm and strategy have been optimized, and through self-developed "truth detection" technology, the "illusion" of Hunyuan large models has been reduced by 30-50% compared to mainstream open source large models.

"The industry's approach is to provide search enhancements, knowledge graphs and other "plug-ins" to improve the ability of large-scale model open-book exams. This method increases the knowledge of the model, but there are many limitations in practical applications." Jiang Jie said. "In the early stages of development of the Hunyuan large model, we considered a method that did not rely on external data at all, and conducted a lot of research attempts. The pre-training method we found has largely solved the problem of hallucinations."

Tencent also The reinforcement learning method is used to let the model learn to identify trap problems, and through the optimization of positional encoding, the effect and performance of the model in processing ultra-long texts are improved. In terms of logic, Tencent has proposed a new strategy of thinking chain, which enables large-scale models to reason and make decisions based on actual application scenarios like humans.

Tencent Hunyuan large-scale model can understand the meaning of context and has long-text memory capabilities , can smoothly conduct multiple rounds of conversations in the professional field. In addition, it can also create content such as literary creation, text summaries, and role plays to fully understand user intentions and provide timely responses efficiently and accurately. Only when such technology is implemented can productivity be truly improved.

The content that needs to be rewritten is: write a 4000-word article. GPT-4 cannot meet the requirements, but the Hunyuan large model can do it

In the standard compliance test of the China Academy of Information and Communications Technology's "Evaluation Methods for Large-scale Pre-training Model Technology and Applications", the Hunyuan large model evaluated a total of 66 ability items. In "Model Development" and "Model The comprehensive evaluation in the two areas of "ability" obtained the highest score currently. In the mainstream evaluation sets MMLU, CEval and AGI-eval, the Hunyuan large model has excellent performance, especially in Chinese science, college entrance examination questions and mathematics.

The significance of building large models lies in industry applications. It is understood that more than 50 businesses and products within Tencent have accessed and tested the Tencent Hunyuan model, including Tencent Cloud, Tencent Advertising, Tencent Games, Tencent Financial Technology, Tencent Conference, Tencent Documents, WeChat Souyisou, QQ browser, etc., and achieved initial results. Tencent's programmers have begun to use large model tools to improve development efficiency

In addition, Tencent has also developed its own machine learning framework Angel to double the training speed of the model compared to the industry's mainstream frameworks, and the inference speed is faster than The industry's mainstream framework has improved by 1.3 times.

The infrastructure for building large models has not been pulled down either. Tencent has previously stated that it has built a large-scale computing center at the beginning of this year. Recently, MiniMax and Baichuan Intelligence's large models have used Tencent's computing power.

Tencent is also working hard to combine industry data with its own capabilities, using industry-specific data from external customers to solve problems in specific industries, and integrating with the real industry to continuously promote the social, economic benefits and business of large-scale models. value

"According to public data, 130 large models have been released in China. There are both general models and professional field models. Hunyuan, as a general model, can support most of Tencent's internal businesses. Today I show several in-depth interfaces All the businesses we enter have a large number of users. Large models have been deeply applied in our core areas," said Jiang Jie. "My big model first serves the enterprise itself, and secondly it is opened to the outside world through Tencent Cloud."

When it is opened to customers, the Hunyuan Big Model will serve as the base of Tencent Cloud Model as a Service MaaS. Customers can either directly call Hunyuan API or use Hunyuan as a base model to build exclusive applications for different industrial scenarios.

It can be seen that Tencent’s strategy in the field of large models is focused on stability: focus on laying a solid foundation and not rush to show off half-finished products. And this move showed excellent strength.

The development of large models is still ongoing. As Jiang Jie said: "It is no exaggeration to say that Tencent has fully embraced large models. Our capabilities have been continuously improved." Evolution, I believe the potential of AIGC is unlimited, and we have already embarked on this path."

The above is the detailed content of Tencent Hunyuan large model was officially unveiled, and we took the lead in trying its productivity. For more information, please follow other related articles on the PHP Chinese website!