Will the traditional AI era eventually pass?

Author: Wufang

Source: Tiger Sniff

Image credit: Generated by Unbounded AI tools

Generative AI is generating a lot of momentum, and it continues to show more and more potential. For example, let AI automatically design chips, or let AI see a doctor.

In order to seize the leading position, the rushing companies not only do their best to release their own technical strength, but also make AI models bigger and bigger, so as to achieve the effect of powerful bricks flying and realize the emergence of intelligence.

The "China Artificial Intelligence Large Model Map Research Report" by the New Generation Artificial Intelligence Development Research Center of the Ministry of Science and Technology shows that so far, China has released 79 large models with a parameter scale of more than 1 billion, and the battle of large models has gradually entered the second half. More people began to pay attention to the combination of large models and real industries, and enterprises also showed the application scenarios of their products one after another.

But in all fairness, do applications that originally use traditional AI technology really need to use large models? What kind of large models are needed by the industry? At the Huawei Developer Conference 2023 (HDC.Cloud 2023), Huawei Cloud showcased its latest progress in "AI for Industries" and the Pangu model, and proposed to let AI reshape thousands of industries.

Large-scale industry models have become a new trend

No enterprise wants to be left behind in the upsurge of large-scale models. If you sing, I will appear on the stage. In recent months, there will be news about a large-scale model every few days. Under the market rotation, enterprises have already seized the opportunity and launched general-purpose large-scale models, and gradually turned to large-scale industrial models to show the practicability of their own products, including real machine demonstrations and program demonstrations.

The reason why the large model is well-known to the public is because to C products such as ChatGPT make people really appreciate the value of technology. In fact, the key to a large model is not how big it is, or how strong its general ability is, but whether it can solve the problems of existing applications. Only by reshaping thousands of industries and improving people's life quality based on the application itself can it be recognized accepted by people.

Throughout the world, due to challenges such as high computing power costs, information leakage, and policy supervision, the commercialization process of large-scale models to C is slower. Most companies choose to be pragmatic and oriented to B-end enterprise customers to meet the requirements of specific industry scenarios. . It can be said that it has become a general consensus to build a large-scale industry model.

Traditional small model AI can usually only perform specific tasks or specific problems. For the to B industry, large models can not only connect traditional system information together, but also realize more complex decision-making and planning.

But it is not easy to make a large-scale model of the industry. Many companies that want to apply AI have to give up in the end.

First of all, the business scenarios of enterprises are complex, and most of them need to be customized. It is necessary to process diverse tasks such as text, pictures, audio, video, and mechanisms. Contrary to this, most enterprises lack data samples.

Secondly, the large model is a game that burns money. Not only does it need to use a large computing power starting at a kilocalorie in the training phase, it also requires highly skilled professional AI application development talents.

Finally, data and knowledge are the core assets of enterprises, and it is necessary to ensure the privacy, security and compliance of enterprise data.

Professional things should be handed over to professional people, and the same is true for large models, instead of letting companies do basic work such as manufacturing screws, bearings, and gears.

Huawei Cloud, as an enterprise that released the Pangu basic large-scale model as early as 2021, is by no means a novice in the field of large-scale models. At that time, it aimed at the industrialization of AI. Today, HUAWEI CLOUD officially releases Pangu 3.0 and Ascend AI cloud service, becoming China's first full-stack independent AI model, adhering to the direction of AI for Industries, and deeply integrating AI with all walks of life.

As mentioned by Huawei executive director and Huawei Cloud CEO Zhang Pinganfa, the Pangu model does not compose poetry, but only does things. It focuses on industry scenarios and is committed to deepening government affairs, finance, manufacturing, coal mines, railways, pharmaceuticals, meteorology and other industries.

The change caused by replacing the small with the big

In fact, the big model has long been out of the concept stage, but is quietly changing everything in life.

Meteorology is inseparable from human beings, and it also brings a lot of damage to our development. For example, every year, about 80 typhoons are generated around the world, of which about 25 affect the Northwest Pacific Ocean and the South China Sea, and an average of 7 will land in my country. In 2022, the direct economic losses caused by typhoon disasters will be 5.42 billion yuan.

Traditional meteorological forecasts are mostly calculated by HPC high-performance computers. To predict the path of a typhoon in the next 10 days, it is necessary to spend several hours on a supercomputer with more than 3,000 nodes for simulation. With the increasing trend of computing power and the complexity of physical models, the bottleneck of traditional numerical forecasting has become prominent.

The large Pangu meteorological model only needs a single machine and a single card, and can complete a weather forecast within 10 seconds. It is the world's first AI model whose accuracy exceeds traditional forecasting methods. It can complete the path of a typhoon in the next 10 days in seconds. The accuracy of typhoon track prediction is the world's first, which is about 20% higher than that of the European Meteorological Agency.

The same change happened in the field of pharmaceuticals.

Antibiotics have saved countless lives, but since daptomycin was discovered in 1987, no new antibiotics have been discovered for nearly 40 years. Not only does drug resistance threaten everyone's health, it could also reduce GDP by at least $3.4 trillion a year by 2030 and push 24 million people into extreme poverty. The world urgently needs a new class of antibiotics to change the situation that patients have no drugs available when they face "super drug-resistant bacteria" infections.

Researching a new drug is not easy. For a long time, the development of new drugs cannot escape the curse of the "Double 10 Law", that is, the average cost exceeds 1 billion US dollars, and the research and development cycle is longer than 10 years. Not only that, but this is the best business format. In reality, it takes an average of 10 to 15 years for a new drug to be approved for marketing, costing more than 2.6 billion US dollars, and the clinical success rate is less than 10%.

The Huawei Cloud Pangu Drug Molecular Large Model uses a new deep learning network architecture to generate 100 million drug-like small molecule libraries with 100% novel structures. Compared with traditional methods, the accuracy of drugability prediction can be increased by 20%.

Professor Liu Bing of the First Affiliated Hospital of Xi'an Jiaotong University used the AI-assisted drug design service based on the large molecular model of Pangu drug on Huawei Cloud in the research and development of new drugs, and developed a super antibacterial drug Drug X (cinnamoylmycin), which is expected to become a nearly The first new target and new class of antibiotics in 40 years, shortened the R&D cycle of lead drugs from several years to one month, reduced R&D costs by 70%, and broke the "Double Ten Law" in the pharmaceutical industry.

It is not only patients who are changed.

Under the dark mine is the life of the workers. Workers who go down the well have been in a high-risk and high-pressure environment for a long time, not only trapped by the shadow of death, but also accompanied by lifelong injuries. But to the contrary, the current 300-meter underground still requires a large number of personnel to work on site, and they need more technical and humanistic care.

AI is a good hand in ensuring the safety of underground operations, and can add a good helper to uncertain manual processes. However, it is not as simple as imagined for AI to enter the coal mining industry. The downhole operating environment is harsh and requires high precision in image and video recognition. Moreover, the differences between mines are large, and the on-site operations are complicated, so the model cannot be easily reused. At the same time, the coal mining industry lacks high-quality artificial intelligence talents.

The Pangu Mine large model only needs to import a large amount of unlabeled mine scene data for pre-training, and then it can carry out unsupervised independent learning. A large model can cover the mining, excavation, machinery, transportation, transportation, washing and other business processes of the coal mine. More than 1,000 subdivided scenarios can realize full-time inspections, help staff find problems in time, avoid safety accidents caused by missed inspections, shorten downtime, and improve the work efficiency of underground inspection personnel. At present, it has been used in 8 mines across the country.

Countless large, medium and small cities are connected in series by railways. So far, my country's railways have a mileage of 155,000 kilometers and more than 1 million railway freight cars. With the improvement of my country's railway safety level, many failures have rarely occurred, and most people have never seen them, but once a failure occurs, it is usually a major failure. For example, if the bolster core plate came out, only one faulty sample was found nationwide.

Behind the safety is the hard work of countless people. Restricted by technological development, the currently widely used TFDS (Truck Operation Fault Dynamic Image Detection System) still uses manual methods for fault identification. Taking the 5T inspection workshop of a certain hub station as an example, nearly 800 trains and more than 40,000 vehicles are inspected on average every day, and the TFDS system takes more than 2.8 million pictures. Time to find the nuances in time to find out the faults in the train.

Through the Pangu large model, it was previously necessary to manually identify 4,000 pictures, but now it only needs to recheck more than 170 pictures, and the labor intensity of workers has dropped by 95.75%. In practical application, it can accurately identify 430+ types of various faults of 67 types of trucks, 100% identification of major abnormal faults, and the screening rate of non-faulty pictures is as high as 95%, exceeding customer expectations.

Such examples are too numerous to enumerate. In fact, every time I use One Netcom to do business and use smart products, it may be the credit of the big model. We have been directly or indirectly enjoying the dividends brought about by technological upgrades. .

How is the Pangu model different?

Today's large-scale model field is still hot. Domestic players staged a "hundred-model battle", among which there are many powerful Internet manufacturers. So what are the differentiating advantages of Pangu's large model?

First of all, HUAWEI CLOUD has hundreds of projects in the field of AI, and adheres to AI for Industries. Combining its accumulated experience in the industry for more than 30 years and the continuous cultivation of more than 10 industry corps, Huawei Cloud has accumulated rich know-how from industry customers and partners. Integrating with the large model, so that the large model has industry knowledge and experience.

Secondly, in addition to learning a lot of general knowledge, the Pangu model has also learned public data from more than 10 industries, covering finance, government affairs, meteorology, medical care, health, Internet, education, automobiles, retail, etc.

More importantly, the Pangu large model has achieved independent innovation from the underlying chip to the whole process platform. You know, in the AI upsurge, GPU has become a hot commodity, but under the multiple influences of geopolitical friction and supply shortages, deploying high-performance computing cards will only be difficult. Therefore, independent innovation has become a general consensus in the industry.

Reviewing the history of Pangea large-scale models is a process of continuously meeting the needs of the industry. In March 2020, Tian Qi joined Huawei Cloud and began to build a team; in April 2021, the Pangu large model was officially released, including the NLP large model and CV large model; in September 2021, Huawei Cloud released the scientific computing large model and A large model of drug molecules; in June 2022, Huawei Cloud released a large model of Pangu Mine; in November 2022, Huawei Cloud released a large model of weather.

Back in time, HUAWEI CLOUD officially released the Pangu Large Model 3.0, and released large models of government affairs, finance, and manufacturing at the same time. The large model was extremely popular a while ago. Why did Huawei Cloud choose this time to announce the progress of the Pangu large model?

In fact, when facing new technologies and trends in the industry, HUAWEI CLOUD prioritizes the needs of the industry, and only when the technology is mature enough will it introduce new technologies to the market. From the perspective of Pangu Large Model 3.0, this time Huawei Cloud clarified the industrial positioning of its products, integrated the previous large models, reorganized the structure, and expanded the network through the new large model. , to cover every industry.

Hu Houkun, the rotating chairman of Huawei, also emphasized at the 2023 World Artificial Intelligence Conference that the key to the development of artificial intelligence is to "go deeper and deeper" to empower industrial upgrading. At the current stage, Huawei has two focus points on the development of artificial intelligence: First, build a strong computing base to support the development of China's artificial intelligence industry. Second, from general-purpose large-scale models to industry-wide large-scale models, let artificial intelligence serve thousands of industries and scientific research well.

Pangu Large Model 3.0 adopts a layered design, including 5+N+X three-tier structure: 5 large L0-level basic large-scale models provide various general skills, N L1-level industry large-scale models help enterprises build their own large-scale models, and massive L2-level Scenario models focus on specific application scenarios or specific businesses, and provide customers with out-of-the-box model services.

Computing power is the food for large models. During the training process, each series of large models of Pangu almost needs hundreds or even thousands of Ascend chips for training. Ascend AI cloud service is the base of Pangu large models, providing it with a solid Base.

HUAWEI CLOUD's single-cluster Ascend AI cloud service with a computing power of 2000P Flops was launched simultaneously in Ulanqab and Gui'an. The HUAWEI CLOUD data center using the Tiancheng liquid-cooled platform can guarantee a 30-day long-term stability rate of 90% for kilocalorie training. Point recovery time does not exceed 10 minutes.

"In order to help global customers, partners, and developers train and use large models, we are committed to creating another pole of the world's AI for global customers and providing new choices for all AI developers," Zhang Ping'an said.

For many enterprises, data security compliance is the primary consideration. In addition to the public cloud deployment model, the Pangea large model can further provide a large model cloud zone and establish a cloud-specific resource pool for large model training and reasoning to ensure data security compliance. For more stringent data localization requirements, hybrid cloud deployment is also provided to help customers train large models on their own private HCS.

For a product, ease of use is the key. HUAWEI CLOUD provides an easy-to-use and reliable large-scale model toolkit, Kaitian aPaaS that gathers a large number of multi-industry scene APIs, and an exclusive community for large-scale models that includes rich and high-quality courses and technical certifications to help developers develop quickly.

It is true that the technology itself is revolutionary, but for Pangu’s large model to enter thousands of industries, it still needs to be given time to take root in the industry.

As Andrew Ng, one of the four kings of AI, said, "It's hard to imagine a big industry that won't be changed by artificial intelligence. Big industries include healthcare, education, transportation, retail, communications, and agriculture. Artificial intelligence will be in these industries This trend is very obvious.” In the future, every industry may be changed by the large models of various industries.