"China Science News" reporter Zhao Guangli
"In the past period, artificial intelligence technology represented by large language models has achieved shocking results, and these have allowed us to see the dawn of general artificial intelligence."
RECENTLY, AT THE WAVE SUMMIT DEEP LEARNING DEVELOPER CONFERENCE HOSTED BY THE NATIONAL ENGINEERING RESEARCH CENTER FOR DEEP LEARNING TECHNOLOGY AND APPLICATION, WANG HAIFENG, CHIEF TECHNOLOGY OFFICER OF BAIDU AND DIRECTOR OF THE NATIONAL ENGINEERING RESEARCH CENTER FOR DEEP LEARNING TECHNOLOGY AND APPLICATION, FIRST PUBLISHED HIS UNDERSTANDING OF GENERAL ARTIFICIAL INTELLIGENCE. He also said that he hopes to build an open source community with all developers and go to the sea of stars of general artificial intelligence.
WANG HAIFENG AT THE WAVE SUMMIT DEEP LEARNING DEVELOPER CONFERENCE. Image source: Baidu
Four keywords that reflect the core capabilities of artificial intelligence
Among the words related to artificial intelligence, Wang Haifeng chose four words: understanding, generation, logic and memory. In his view, the meaning of these four words is not only the core ability of artificial intelligence, but also the most basic ability that a general artificial intelligence system should have.
"For example, if a large language model wants to create a piece of content, it needs to 'understand' the theme of creation, clarify the 'logic' of creation, search for materials in 'memory', and integrate the understanding of the materials, and finally 'generate' a qualified manuscript." For example, Wang Haifeng said, the same is true for programming, problem solving, planning, etc.: "If an AI system has strong understanding, generation, logic and memory capabilities, it can complete many different tasks." ”
Obviously, the meaning of the four keywords of understanding, generation, logic and memory is also the ability that runs through the human thinking process. Wang Haifeng believes that today, large language models have initially possessed these capabilities, and as these capabilities become stronger and stronger, "it will allow us to move faster towards general artificial intelligence."
The code capability of the large model may allow people to see more realistically its progress along the road to general artificial intelligence.
Wang Haifeng said that language is a tool for human communication and communication, and it is also a carrier of thinking. But computer programming languages are different from meaning-rich natural languages, they have strict syntax, line formatting, and each line of code can only be interpreted and executed. Therefore, the process of writing code by human programmers is actually a process of expressing human thinking into programming languages, which can be executed and interacted with.
When big language models can also "write" code quickly and well, things start to get worse.
"When I was in school, a foreign language teacher once told me that if you can think in a foreign language, it means that you have learned the foreign language." Therefore, Wang Haifeng said, when the big language model can "write" code, it also means that it can write the needs expressed by human natural language in code.
Wang Haifeng expressed his opinion: "This is not only to help people's software development, but also to build a bridge between machine 'thinking' and 'execution' - which is of great significance for the development of general artificial intelligence." ”
"Knowledge is power" also applies to AI
When people see the play of some large language model products "sometimes gods and sometimes ghosts", they can know that artificial intelligence's cultivation of the ability of "understanding, generation, logic, and memory" is not yet at home.
How can AI continuously acquire and enhance its capabilities in these areas? Wang Haifeng's answer is "knowledge augmentation big language model".
In March this year, Baidu released the knowledge enhancement big language model "Wen Xin Yiyan". In the past five months, Wen Xin's ability to speak has made great progress. A Few-Shot review of the base model in June showed that the latest version of the Wenxin Model 3.5 scored better than ChatGPT in multiple test sets. Wen Xin's progress was faster than expected. Behind this, there are many elements worth mentioning, such as a huge amount of high-quality data, the optimization of multiple strategies, the "long text modeling" of the basic model, the supervised fine tuning of multi-task adaptation, the reinforcement learning of multi-level and multi-granularity reward models, and the joint optimization of Wen Xin and Flying Oar... And so on, and so on. But in Wang Haifeng's view, it is important that Wen Xin not only learns from massive data, but also draws nutrients from the huge knowledge map.
As the name implies, as a "knowledge enhancement" big language model, Wen Xin is inseparable from the learning and reinforcement of "knowledge". The phrase "knowledge is power" also applies to artificial intelligence.
Wang Haifeng said that Baidu has a knowledge graph of more than 550 billion knowledge that took more than 10 years to build.
In the training process, how to make good use of this knowledge graph of the Wenxin large model? Wang Haifeng replied that there are two ways: knowledge internalization and knowledge external use. Knowledge internalization means that in the training process, through semantic unit-based learning and constructing training data with knowledge graph, these knowledge are internalized into large language models; Knowledge external use is to directly use knowledge graphs in knowledge reasoning and prompt construction.
After the "baptism" of huge knowledge graph and massive data, large models can be docked to thousands of industries as long as a small amount of fine tuning and a small number of scene adaptation in the inference deployment stage, which will help them greatly reduce the application threshold.
"Wen Xin plus flying oars, flying to the galaxy"
AT WAVE SUMMIT, A GRAND EVENT FOR DEEP LEARNING DEVELOPERS, WANG HAIFENG'S VISION OF GENERAL ARTIFICIAL INTELLIGENCE IS A RESONANCE SOUGHT BY TENS OF THOUSANDS OF DEVELOPERS.
Wang Haifeng said that on the flyblade industry-grade deep learning open source platform developed by Baidu, 8 million developers have been gathered, and 220,000 enterprises have used the flyblade platform to build 800,000 models.
In the early days, developers favored internationally renowned deep learning frameworks such as TensorFlow and Pytorch, but with the gradual improvement of the domestic platform of Flypad, development kits, tool components, basic model libraries, etc. are gradually complete, especially in the past few years, based on Flypad, Baidu has focused on creating a large model family of Wenxin, and more and more developers have turned to embrace Flypaddle. In 2019, there were only 1.9 million developers on the Flypaddle platform, and in four years, this number has doubled in a row.
"With 8 million developers, 220,000 businesses, and 800,000 models, these numbers not only witness growth, but also lay a solid foundation for the future." Wang Haifeng said that Baidu hopes to continue to build and co-create with all developers and enterprise partners to jointly promote artificial intelligence to empower thousands of industries and benefit millions of households.
Wang Haifeng disclosed the latest figures of the flying paddle platform. Image source: Baidu
The word "flying oar" is taken from the Song Dynasty literary magnate Zhu Xi's "That is, there are two flying oars, and they go down to Guangjin" in "Ji Shi Huai Zhiyanfu Zhongzong's Second Brother". It means that with the help of "flying oars", the Chinese intelligent cause will go faster and further.
The artificial intelligence learning training community gathered by developers of Flypaddle, originally called AI Studio, is now given a Chinese name "Galaxy Community" by Baidu. Millions of developers have learned, trained, improved their AI capabilities and implemented them in the Galaxy community.
AT THE WAVE SUMMIT DEEP LEARNING DEVELOPER CONFERENCE, WANG HAIFENG CHANGED THE POEM "HEARING ABOUT DOUBLE FLYING OARS, FLYING DOWN GUANGJIN" TO "WEN XIN PLUS FLYING OARS, FLYING TO THE GALAXY": "I HOPE WE ARE WITH ALL DEVELOPERS." With the blessing of flying oars and Wen Xin, we will build a galaxy community and go to the sea of stars with general artificial intelligence. ”