From the Create Conference, we can see Baidu in the golden decade of AI

Editor | Yu Kuai

As a top technology conference, Baidu AI Developer Conference is a little different this time:

Cancel the offline venue and broadcast the entire event online.

At first glance, it seems nothing new. After all, live streaming has become part of ordinary people’s lives and can be achieved with just a mobile phone and an app.

But what if we want to make participants feel as if they are in the center of a venue with tens of thousands of people, or what if a virtual image can be generated with just a picture or a sentence?

Obviously, this kind of "live broadcast" is fundamentally different from the live broadcast understood by the general public.

The layman watches the excitement, while the expert watches the details.

The public may be shocked by the on-site effects, but what developers see is the powerful spatial sound field reconstruction, AI acoustic modeling and other technologies behind the more realistic conference experience, and the leading vision, voice, natural language understanding and other capabilities behind the more realistic interactive scenes.

It actually realizes the reconstruction of the sound field of the speaker in the center of the stage of a super-large open-air complex open venue, the real-time background sound restoration of the voices of 100,000 real-life audiences, and the immersive communication space sound effects created by 3D sound technology that cover the surrounding users with a simulated effect.

It combines technologies such as augmented reality, knowledge graphs, speech recognition, natural language processing and image recognition, as well as massive amounts of personal knowledge data, to achieve multimodal technology integration.

As the "host" of this online conference, Baidu once again refreshed people's understanding of technology.

This uncompromising believer in technology may be able to give us a glimpse into what future technology will truly look like.

In recent years, AI has become more and more mediocre after being disenchanted. Is the era of AI over?

Perhaps the opposite is true - the era of AI is only now truly arriving.

Looking back at the first three industrial revolutions, their respective hallmarks are: the widespread use of steam engines, the widespread use of electricity and internal combustion engines, and the widespread use of information technology. As a result, machines replaced manual labor and social productivity was greatly improved.

Note the key word: replace. What is replace? To replace one thing with another.

To achieve this, "widespread and large-scale" application is essential.

Today, in the fourth industrial revolution, in order to promote machines to replace brain power and achieve true intelligence, it requires the application of the whole society and all industries.

In fact, ever since Internet giants began to “raise pigs”, things started to go “wrong” with AI.

AI has begun to blossom in areas such as transportation, government affairs, electricity, education, factories, medical care, finance, culture and tourism, urban management, steel, ports, water affairs, agriculture, border and coastal defense, etc.

The reason why artificial intelligence is no longer "sexy" and cannot bring more dopamine to people is that it is no longer the moon in the mirror or the flower in the water, but has penetrated into every aspect of human life.

According to the five-stage Gartner model, today's AI has gone through three stages: from nothing to something, the god-making stage, and the difficult implementation stage. It is currently on the way from the trough to the climbing stage.

The height of AI at the policy level can be seen from the fact that "new generation artificial intelligence" ranks first in the frontier scientific and technological fields of the 14th Five-Year Plan.

It is no exaggeration to say that in China, if the past 10 years were the rise of AI technology, the future will be the 10 years of AI implementation.

To enter the golden decade, we actually need an "AI toolbox". With it, thousands of creators can transform themselves into "brothers who overcome obstacles" and "sisters who ride the wind and waves".

Baidu wants to, and can, become the “box maker”.

On the title page of building the AI ​​foundation, the insurmountable hard indicators are clearly written: computing power (computing facilities), algorithms (training frameworks), and data (knowledge).

Look at knowledge first.

For AI to be implemented on a large scale, it must have computing power facilities and training frameworks as a foundation, and industry knowledge as content.

The industrial age is a competition of strength, and the intelligent age is a competition of knowledge.

The core of the industrial age is power and the large-scale application of machinery and energy. The core of the Internet age is information exchange and large-scale information flow. In the intelligent age, the core is knowledge.

For example, to identify the weather, you need to understand the different types of clouds, their generation principles, and their evolution processes, and you also need to simulate the test environment and test sets. For another example, to identify pigs, you need to understand the entire growth cycle of pigs and the specific conditions of each cycle.

Traditional large models only learn from massive amounts of text and can only memorize the rules of massive amounts of text. These models cannot fundamentally understand how the world works and can only solve some problems on the surface of the text.

In order to overcome this hurdle and better understand the complex scenarios in the real world, Baidu began to explore the potential of knowledge-enhanced pre-training models early on.

A large model with knowledge content will have more prior insights into reality, and will naturally learn new things faster, which means it has stronger small sample capabilities.

In 2019, the first generation of knowledge-enhanced pre-training model ERNIE1.0 was released. In July this year, a knowledge-enhanced pre-training model with tens of billions of parameters, ERNIE 3.0, was trained. This promotes information sharing between structured knowledge and unstructured text, thereby greatly improving the model's ability to memorize and reason about knowledge.

On December 8, Pengcheng-Baidu Wenxin 100 billion big model was released, and Baidu's "Wenxin" series of industrial-level knowledge-enhanced big models were unveiled. On top of the big model are EasyDL-big model zero-threshold AI development platform and BML-big model full-featured AI development platform, as well as big model development tools, big model lightweight tools and big model deployment tools.

Not only parameters, the Wenxin model comes with a core buff: knowledge enhancement.

The knowledge-enhanced large model can continuously learn from large-scale knowledge and massive unstructured data at the same time, just like standing on the shoulders of giants. The training efficiency and understanding accuracy are greatly improved, and it has better interpretability.

In addition to integrating knowledge and data for learning, Baidu Wenxin also uses knowledge-enhanced cross-language learning and knowledge-enhanced cross-modal learning to learn unified semantic representation and understanding capabilities from multiple languages ​​and multiple modal data, which has strong versatility. It can effectively improve results, reduce application costs, lower application barriers, and meet application needs in real scenarios.

Since its birth in 2019, Baidu Wenxin has made a number of technological breakthroughs in language understanding, text generation, cross-modal semantic understanding and other fields, and has won dozens of world championships in public authoritative semantic evaluations.

Based on this, Baidu Wenxin Model has been widely used in finance, medical care, insurance, securities, office, Internet, logistics and other industries.

Let’s talk about the training framework.

Baidu CTO Wang Haifeng once said that the deep learning framework is the "operating system of the intelligent era."

As the skeleton of the underlying language and algorithm model, the deep learning framework saves developers the cost of building the foundation from 0 to 1, thereby improving development efficiency.

Developers can choose models in the framework, assemble or train their own models, import data and obtain models, and finally deploy them according to the characteristics of their own industry and scenario needs, just like building blocks.

If chips are computing power platforms, then deep learning frameworks are programming productivity platforms. Both have become the foundation of AI infrastructure.

Previously, mainstream open source deep learning frameworks have been dominated by foreign technology giants, which makes domestic open source frameworks an inevitable thing in the increasingly tense Sino-US relations.

In 2016, Baidu officially open-sourced its deep learning framework amid the dominance of foreign giants, making PaddlePaddle China's first independently developed, open-source, and fully functional industrial-grade deep learning platform.

In the past five years, PaddlePaddle has undergone multiple iterations and upgrades in terms of development convenience, ultra-large-scale deep learning model training technology, high-performance inference engine deployed on multiple terminals and platforms, and industrial-grade open source model library. As of now, PaddlePaddle has gathered 4.06 million developers, developed 476,000 AI models, and served more than 157,000 enterprises and institutions, covering dozens of fields including industry, agriculture, medical care, urban management, transportation, and finance.

Today, PaddlePaddle has become the world's second largest open source deep learning framework and China's largest learning framework.

At the stage of full implementation of AI, the Wenxin big model has raised the wall of the AI ​​base, and PaddlePaddle has greatly lowered the threshold of the algorithm. Together with the Baidu Kunlun chip, it has become a solid arm of Baidu Brain's AI infrastructure.

As the infrastructure of the intelligent era, Baidu Smart Cloud has been ranked as one of the "four clouds of Chinese technology companies" in the first half of this year, and has led the market with a growth rate of 73%. In the AI ​​Cloud segment, it has ranked first for four consecutive years.

This solid AI base, like a satellite launch base, is continuously launching industrial application satellites.

The most eye-catching one is smart transportation.

Intelligent transportation is inseparable from the intelligence of cars, roads, and the intelligence of car-road collaboration. From cars to roads to people, Baidu has been making steady progress over the years.

"Apollo", which was born in 2017, is about to be upgraded to version 7.0. In this process, Apollo has grown into the world's most active open autonomous driving platform, with more than 210 global ecological partners, 80,000 global developers, and 700,000 lines of open source code.

The latest achievement of the autonomous driving travel service platform "LuoBoKuaiPao" announced in August is that it provided 115,000 services in a quarter, making LuoBoKuaiPao the world's largest autonomous driving travel service provider.

This gives Baidu the confidence to declare its ambition: by 2025, the service will expand to 65 cities and to 100 cities by 2030.

The task of marketing Baidu's driverless technology and smart cockpit technology over the years has been handed over to Jidu Auto Robotics, a company newly established this year.

This future car robot, which Baidu claims "has L4 level autonomous driving capabilities, accurate recognition of human-vehicle interaction and voice semantics, and can self-learn and iterate", is expected to be available to the public within 2-3 years.

After having so many stories about transportation, Baidu has finally presented the freshest industry stories.

During this year's Two Sessions, the digital economy once again became the focus of national attention, and the country proposed a goal of increasing the added value of core digital economy industries to 10% of GDP by 2023.

With the advent of the digital economy, the digital transformation of industries is a good business. At the invisible starting line, countless companies are crossing layers of iron curtains, fighting in all directions, or fighting in valleys and ridges in a small space.

Baidu has plunged headfirst into the wave of digital transformation of the industry. With the technical strength of cloud computing and AI, Baidu Smart Cloud has taken advantage of its "cloud and intelligence integration" and has expanded into industries such as industry, energy, smart cities, agriculture, and urban governance.

Entering the water affairs industry, Baidu Smart Cloud has participated in the creation of the "Water Affairs Brain" in Quanzhou, Fujian, which intelligently upgrades urban water affairs processes. Through precise pressure regulation trained by AI technology, smart aeration, smart dosing, smart drainage and other models, water affairs operations are made more efficient.

Turning into the industrial sector, Baidu's AI and cloud computing have entered the Kaibo workshop in Changshu, Suzhou. AI helps quality inspection workers "listen to sounds" and "recognize images", improving efficiency and quality, helping companies accumulate and sediment knowledge, and promoting China's manufacturing industry to achieve the transition from "manufacturing" to "smart manufacturing".

In Beijing Haidian, the "City of Science", everything from illegal parking to missing manhole covers, from water stains on the road to random garbage dumping, can be intelligently identified, and the system monitors and notifies staff to handle them in a timely manner;

In Xinjiang, Baidu Smart Cloud has joined hands with Grid Power to build an "AI middle platform" to provide strong support for the export of electricity from Xinjiang and reduce the risks of inspections in harsh environments.

Moving into a high-risk industry, Baidu achieved safe production by reducing the company's operation and maintenance costs at a petrochemical plant in the chemical industry by increasing the frequency of inspections and the ability to detect hidden dangers in advance without increasing manpower input.

Baidu has also entered agriculture, not only in cities. In Shouguang, Shandong, the hometown of vegetables in China, Shouguang Brain, built by Baidu Smart Cloud, has put AI clothes on vegetable greenhouses. Two staff members plus an APP can manage dozens of vegetable greenhouses. As long as the villagers are willing, they can chat with Xiaodu anytime and anywhere to learn agricultural knowledge such as agronomy, pest and disease prevention.

These are just a small example of the intelligentization of thousands of industries.

Baidu's cloud computing and AI are like water and electricity for the digital transformation and upgrading of the industry. With the help of China's general trend of intelligent industrialization, they are quietly flowing in the fields of industry, finance, transportation, energy, cities, medical care, autonomous driving, etc.

Behind all this, Baidu Brain has benefited from years of technological accumulation and industrial practice, as well as the combined power of the PaddlePaddle platform, Wenxin big model, Kunlun chip, etc. Baidu Brain, which has evolved into an AI production platform, will use the weapon of "integration innovation and lowering barriers" to break through the shells of more traditional industries.

In 2010, when smartphones were just becoming popular and AI had not yet revealed its true potential, Baidu began to lay out AI technology in China and established a "Natural Language Processing Department."

Over the past decade, Baidu has maintained its consistent commitment and incredible determination to AI, regardless of the frenzy of the AI ​​market or the ridicule of AI's fall from glory to rock bottom.

Baidu founder, chairman and CEO Robin Li once said: "When we have 1 yuan, we will invest in technology; when we have 100 million, we will invest in technology; when we have 10 billion, we will still invest in technology."

In the past 11 years, Baidu's R&D investment in AI technology has exceeded 100 billion yuan. Becoming a national team in artificial intelligence, the constantly upgraded Baidu Brain, the expanding AI ecosystem, the growing number of industrial intelligence fields, and the upcoming mass production of smart cars all speak of the dedication of this AI evangelist.

Perhaps it is because technology has a natural appeal to Baidu, or perhaps it is the sense of social responsibility engraved in its culture, Baidu does not know how to stop and is always on the road to cutting-edge technology.

The continuous epidemic and the constantly mutating virus make vaccine development and production extremely critical. Among them, mRNA vaccine is a new type of vaccine that can be produced quickly and on a large scale, but it is extremely difficult to design a stable mRNA vaccine sequence, which limits the large-scale application of mRNA. With Baidu's heavy investment, it launched the industry's first mRNA vaccine sequence design algorithm, which can find a stable vaccine sequence in just ten minutes.

Baidu is also planning to develop quantum computing technologies that can be used in fields such as chemistry, finance, and materials, and is creating a domestic quantum computing cloud platform that provides one-stop services from applications to quantum processors. Their motivation is to lower the threshold for learning and applying quantum computing.

Baidu has implicitly shouldered the historical responsibility of "creating responsible and sustainable AI". Adhering to optimism and long-termism, they discuss concepts and are committed to applying them in practice. They are not greedy for fame and fortune, nor are they afraid to devote themselves to the difficult exploration of future technologies.