Home > Article > Technology peripherals > Robin Li spent more than 100 billion in 10 years! 5 million developers support China’s largest deep learning framework
What are you thinking about when you look up at the stars?
If you ask questions persistently, you will get responses you never imagined.
shen
Those who create a new universe from within a small space to beyond the world believe that they can never see the ceiling.
Determined people will not stop just because they arrive. They measure the unknown with the pace of creation.
Baidu CREATE 2022 conference showed us paintings jointly created by human designers and AIGC.
And just like the scenes described in these paintings, Baidu has never stopped innovating.
At the beginning of the conference, Robin Li put forward a thought-provoking point of view: the symbol of the fourth technological revolution is deep learning algorithms. Major innovations related to deep learning will have a major impact on our society, just like cars and the Internet.
##In real practice, there is no navigation map, only a compass. Baidu also came up with valuable innovations after figuring out the general direction and iterating step by step based on practice.
At the conference, Robin Li showed such a painting. This painting was generated on the Baidu AI painting platform using the keywords "crisis and hope".
This painting well represents the current situation facing artificial intelligence - experiencing ups and downs, but full of hope.
Yes, Robin Li still adheres to last year’s view that creators will usher in a golden decade of artificial intelligence.
The first year of AIGC is coming2022, AIGC is in full swing.
DALL·E2 has made text-generated pictures popular for a whole year. The subsequent Stable Diffussion and Midjourney have inspired countless people's artistic inspiration and even shocked many painters.
The Imagen models released by DALL·E2 and Google have also attracted many AI scholars to participate in research.
Although ChatGPT only made its debut at the end of the year, the magic given to it by "reinforcement learning" allowed it to once again set off a storm in AIGC during the national carnival.
In fact, at the beginning of 2022, Baidu Research Institute had already predicted the popularity of AIGC this year.
The ultra-large-scale pre-training model shows the trend of knowledge enhancement, cross-modal unified modeling, and the co-evolution of multiple learning methods, and is gradually becoming practical.
For example, AIGC (AI generated content, artificial intelligence created content), with the help of the cross-modal comprehensive technical capabilities of large models, can stimulate creativity, improve content diversity, reduce production costs, and Will achieve large-scale application.
Moreover, this is not the first time that Baidu Research Institute has made divine predictions. In 2020, the NLP model it predicted was fulfilled on GPT-3, and in 2021, the digital people it bet on exploded.
But this time, Baidu is very confident about the trends in the AIGC field.
Three talented creators, one click to realize your dream of directingAt the Create 2022 conference, Baidu deeply applied AIGC to almost every link. The creation of songs, scenes, and speech mind maps are all involving AI.
##
The paintings generated by Robin Li with the theme of "crisis and hope" are based on the cultural knowledge-enhanced cross-modality There is a large-scale model, and it is one of the three talented creators who will appear next.
At this Create conference, Wu Hua, chairman of Baidu Technical Committee, introduced us to three talented creators with extraordinary abilities.
They are the talented screenwriter - Wenxin ERNIE3.0Zeus, the talented illustrator - Wenxin ERNIE-ViLG 2.0, and the editing and animation master - VIMER-TCIR.
With these three talented creators, coupled with the virtual actors of your own design, you can also become a director and shoot your own ’s film and television masterpiece!
Wenxin ERNIE 3.0 Zeus language model will chat with you forever!
As the latest upgrade of the ERNIE 3.0 series of models, in addition to learning unlabeled data and knowledge graphs, ERNIE 3.0 Zeus also learns more than a hundred different forms through continuous learning Comprehensive control of task data is achieved.
After a "double-pronged approach" to general knowledge and specialized knowledge, the model's generalization ability has been significantly improved. Whether it is multi-language understanding or generation tasks, it can be easily handled, worthy of the name " "Generalist" belongs to the category.
Whether it is independent creation, free answer, propositional dialogue, emotional analysis, or more than 100 hierarchical prompts, ERNIE 3.0 Zeus can handle it smoothly.
Wenxin ERNIE-ViLG 2.0 image generation large model, the talented painter knows about it?
If Wenxin ERNIE 3.0 Zeus is the master of language, ERNIE-ViLG 2.0 is the master of painting. It can generate a beautiful painting based on a sentence or a paragraph of descriptive text.
Chinese tip, if you want any painting, just ask. There is nothing you can’t think of and there is nothing he can’t draw. Take a look at this ship Does the blue and white porcelain battleship look like a fine piece of art? I just don’t know if they will be shot to pieces if they really fight...
To achieve this goal, it is inseparable from the diffusion of enhanced knowledge behind the model itself Model.
Among the prompts in one sentence or several sentences, which are the core elements that need to be highlighted in the painting, and which are the decorative elements? Be knowledgeable.
In order to achieve as accurate painting as possible, during the learning process, ERNIE ViLG 2.0 introduced multi-source knowledge such as language and vision to guide the model to pay more attention Core semantic elements in text and images to achieve precise fine-grained semantic control.
In addition, ERNIE ViLG 2.0 can also select different network (such as noise reduction) modeling frameworks for different stages, effectively solving the problem of inconsistent requirements for model capabilities at different stages and reducing It eliminates the mutual interference of noise reduction tasks and improves the quality of image generation.
Whether it is realistic style, Chinese style, national trend, or Chinese painting style, ERNIE-ViLG 2.0 can generate to-the-top, different styles and vivid images based on short Chinese prompts. Realistic images.
For example, take the following gorgeous and elegant "Feast in Heaven":
Based on ERNIE-ViLG 2.0's literary style and unique style have delivered a comprehensive work that can be said to be comprehensive. The overall painting style is bright and colorful, without losing the ancient sentiment.
Now, we only need to enter a few keywords on Baidu's "Wenxin·Yige" platform, and we can get unique styles in minutes paintings.
#In addition to language and images, the generation and editing of video content is also where Wenxin Big Model shows its talents.
In terms of visual content generation, the large video generation model can automatically generate content based on a description text or an image provided by the user. Generate high-definition, smooth video efficiently.
In terms of visual editing, the VIMER-TCIR multi-task large model can be used for super-resolution, denoising, deblurring, and decompression Joint pre-training for multi-tasks, and repair and editing of a variety of different situations at the same time.
Currently, VIMER-TCIR has been implemented in scenes such as the restoration of old movies, and has greatly improved operating efficiency. A single machine can repair 285,000 frames of video every day, solving most of the problems of old movies. Movie screen repair issues.
The wave of AIGC has arrived. In the future, it is foreseeable that AI painting, AI video creation, etc. will soon become It's as easy as taking a photo with your phone.
With the continuous breakthroughs in technology, AIGC will very likely subvert the existing content production model and create content at one-tenth of the cost and at a production speed a hundred times a thousand times. Content with unique value and independent perspective.
In order to achieve such cool effects on large models, Baidu is not stingy in research and development. .
A total of more than 100 billion yuan has been invested in the past ten years, of which core R&D investment has accounted for more than 20% of core revenue for eight consecutive quarters.
According to statistics, Baidu’s R&D investment intensity in 2020 was 18.22%, ranking first among the top 500 private enterprises. In 2021, it was 20.03%, ranking second among the top 500 private enterprises.
By the way, such a "generous" investment has also given Baidu a leading edge in the underlying technology of artificial intelligence. .
After all, if the chip is stuck, so is the basic software.
As early as 2016, Baidu began to develop a deep learning framework called Fei Paddle, which is called an "artificial intelligence operating system".
Currently, 5.35 million developers have been gathered, 670,000 models have been created, and a prosperous deep learning ecosystem has been built.
Large models based on flying paddles can also effectively integrate multi-modal capabilities such as natural language processing and computer vision, and combine them with multiple industry business scenarios for optimization; and developers can also Building AI applications like building blocks greatly lowers the threshold for AI application.
We have already mentioned at the beginning of the article that Robin Li believes that deep learning-related Major innovations, including autonomous driving, intelligent dispatching systems in hydropower and other fields, will have a significant social impact.
Where does innovation itself come from? In Robin Li's view, innovation is driven by feedback.
Baidu has a lot of practical experience in "feedback-driven innovation" in its business development. For example, the reason why Baidu Kunlun chip has leading performance among AI chips is precisely because it has been optimized for Baidu's search services for ten years.
Baidu’s search service responds to billions of real user needs every day, performs 1 trillion times of deep semantic reasoning and matching every day, and can provide the most authentic and timely feedback. , thus forcing the optimization of large models, deep learning frameworks and chips.
Now, Baidu is one of the few artificial intelligence companies in the world that has a full-stack layout (chip layer, framework layer, model layer and application layer).
From the high-end chip Kunlun, to the Flying Paddle deep learning framework, to the Wenxin pre-trained large model, there are key self-developed technologies at all levels , there is a lot of feedback between each layer, and end-to-end optimization is achieved by continuously obtaining feedback.
The technical architecture of each layer is more general as you go down, and more specialized as you go up.
The more specialized artificial intelligence is, the more it can penetrate into industries and empower the development of the real economy.
At last year's Create conference, Robin Li predicted: "As the threshold for technology application continues to lower, creators will usher in a golden 10 years of artificial intelligence." Today, he still Think so.
In 2020, when Robin Li just started his business, he faced the bursting of the Internet bubble, and the world’s market value of 8 trillion evaporated. Subsequently, the Internet entered the In the golden decade, artificial intelligence will also experience the same ups and downs.
Baidu will continue to cultivate AI talents for society and the industry, invest more resources, and work with developers to make its greatest efforts for the development of AI in China.
At the end of the Create conference, the virtual band members appeared again.
## Vocalist/Guitar: Xi Jiajia, Drummer: Du Xiaoxiao, Bass: Ye Youyou, Keyboard: Lin Kaikai
Xi Jiajia said that he was so happy that his painting could be displayed at the opening of the conference!
And Lin Kaikai had become addicted to being a producer. He happily boasted that he was quite talented in arranging~
叶Youyou said that her design actually caught everyone's aesthetic point, which made her quite satisfied. So, which aspect of the design was her responsibility?
Du Xiaoxiao guessed correctly: it was the "Zhiyi" and "Qianliu" links.
And Du Xiaoxiao said that she had already written the press release draft.
In the joint brainstorming of the four members, the title of the manuscript came out - "Shocking!" This is a conference between man and machine."
The above is the detailed content of Robin Li spent more than 100 billion in 10 years! 5 million developers support China’s largest deep learning framework. For more information, please follow other related articles on the PHP Chinese website!