Born to solve complex problems.

Text | Huashang Taolue, Xiong Jianhui

The direct contribution of AI software platform companies to human society may be less than 2%, with 98% promoting industrial and agricultural societies. - Ren Zhengfei

【A hundred boats compete, those who row hard come out on top】

In December 2020, Ren Zhengfei suddenly arrived in Jinzhong, Shanxi.

At the age of 76, he donned mining clothes and a safety helmet, went more than 500 meters underground, and personally inspected the country's first 5G coal mine.

There aren't many people who work so hard at such an advanced age, but Ren Zhengfei's idea is simple:

'I hope coal miners can wear suits and ties to work.'

A few years have passed, and Huawei Cloud has not only equipped many coal mines with 5G and created the 'Mine Hong' operating system but also developed the 'Pangu Mining Large Model.'

In 2023, a Huawei AI expert, like Ren Zhengfei, descended into the depths of a thousand-meter mine shaft of Shandong Energy.

This is a large state-owned mine with a high degree of mechanization and intelligence, but inspection work is still complex and fraught with risks. Whether it's mechanical failure, accidental falls by miners, or gas and flood accidents in the tunneling face, inspectors need to arrive immediately for emergency handling. Otherwise, it may lead to an escalation or loss of control of the situation.

For this reason, inspectors have to walk 10 kilometers on foot every day for the belt inspection of the main transportation system alone.

But with Pangu installed, things are different.

Inspectors no longer rush down the mine shaft all at once. Once there is an anomaly in the monitoring, the model automatically identifies and alarms.

Of course, the system inevitably still has false alarms. At this time, Huawei's AI experts always insist on going down the mine shaft together despite the inspectors' dissuasion. When they arrive at the scene, if there is no anomaly, the inspectors will click the mouse under the guidance of the AI expert to confirm that it is a false alarm and transmit the information back to Pangu.

As a result, Huawei's AI experts arranged for the large model to be retrained and redeployed, gradually achieving an accuracy rate of 98%.

In most cases, inspectors no longer have to risk their lives to confirm dangerous situations and issue early warnings.

And this is just the tip of the iceberg of the Pangu Mining Large Model.

In fact, the application of Pangu covers 9 specialties and 21 scenarios in coal mining, tunneling, main transportation, auxiliary transportation, lifting, safety supervision, impact prevention, washing and selection, and coking.

Furthermore, Shandong Energy uses the 'Pangu Prediction Large Model' to conduct intelligent analysis of raw coal quality data and process parameters for different coal seams, seasons, and ash contents, significantly increasing clean coal production. Shandong Energy Jining No. 2 Coal Mine alone produces an additional 8,000 tons of clean coal annually, resulting in a noticeable improvement in efficiency.

Obviously, Frost & Sullivan, an international authoritative analysis agency, understands the industry.

Recently, Frost & Sullivan released the 'China Industry Large Model Market Report, 2024,' which shows that in the field of industry large models, Huawei Cloud's Pangu Large Model stands out:

It holds the first market share in the government, industry, and finance sectors;

It is a leader in the medical, pharmaceutical, meteorological, and automotive fields.

Regarding this, Li Qing, Director of Research at Frost & Sullivan, stated:

'Relying on its profound technical accumulation, precise grasp of the industry, and rich experience, Huawei Cloud's Pangu Large Model holds an absolute advantage in multiple industries, providing a steady stream of power for the intelligence and digitization of various industries.'

Among these, the 'Pangu Industrial Large Model' deserves special mention for ranking first with a 38% market share.

Without exaggeration, the gold content of this 'first place' is unmatched by most large models.

This status is truly 'earned' by Huawei in the industrial field.

Just compare it with large models from Europe and the United States.

For example, JPMorgan Chase uses SpectrumGPT to quickly analyze financial data, Unilever uses generative AI to improve customer service efficiency, and AstraZeneca uses GenAI to design tumor target therapeutic antibodies.

Even BMW, a manufacturer in the automotive industry, uses generative AI to optimize automotive manufacturing processes; while U.S. Steel relies on the large model MineMind to reduce truck maintenance costs.

These large models focus on the C-end and mainly serve as assistants; in reality, they are still trapped in the digital world, unable to break through the medium and become a force that changes the physical world.

In contrast, the Pangu Industrial Large Model has penetrated into roaring workshops, arduous railways, and high-risk mines, being implemented in over 30 industries and more than 400 scenarios, quietly reshaping various industries.

▲ Shandong Energy Xinglongzhuang Coal Mine Dispatching and Command Center

Like many miners at Shandong Energy, they have truly transformed from 'blue-collar workers' to 'white-collar workers,' starting to mine coal while enjoying air conditioning on the ground and becoming witnesses to the changes brought about by 'Pangu.'

Why is there such a big difference?

Because as the only country in the world with all industrial categories in the United Nations Industrial Classification, China has a complete industrial system and rich application scenarios.

Therefore, by following the development route of 'basic large model + industry small model,' AI can further enhance China's industrial level, leverage the advantages of China's complete manufacturing industry, rich scenarios, and vast amounts of data, and establish global competitiveness in AI.

Using 'Pangu,' which 'does not make poetry but only does things,' to deeply cultivate industrial scenarios and achieve leapfrogging in intelligent manufacturing is precisely what Huawei Cloud is doing and doing exceptionally well.

【Industrial AI must be meticulous】

At the end of 2022, when the world was stunned by the stunning performance of ChatGPT, Huawei Cloud had already determined the direction of AI development:

First, extending from small models to large models; second, deeply integrating AI with industries, i.e., AI for Industries.

From a theoretical perspective, Huawei Cloud must first take root in the industry for a long time and conduct in-depth and thorough foundational work under the surface to accumulate basic industry knowledge, data, and training results, transforming them into industry large models.

On this basis, the industry large models are then released to different enterprises for secondary development and fine-tuning according to their usage scenarios.

However, the difficulty of this task exceeds ordinary people's imagination.

In 2024, Baosteel and Huawei Cloud joined forces to embark on smart manufacturing, facing a series of almost insurmountable challenges.

Take steel rolling as an example.

During the steel rolling process, it is extremely difficult to control the size of the material. A steel billet 260mm thick is quickly rolled into a steel plate 1.2mm thick within 2 minutes due to various uncertain factors such as thermal expansion and contraction, extrusion deformation, and roll wear.

However, if the thickness tolerance exceeds 0.05mm and the width tolerance exceeds 5mm, it will be immediately judged as unqualified.

Therefore, the traditional rolling process requires an experienced master to complete 20 processes, adjust over 300 parameters of the production line, and make adjustments as needed, with no fixed rules.

Using traditional AI results in unstable performance and significant precision fluctuations, requiring continuous optimization, adjustment, and deployment, which takes at least a week.

Faced with this thorny problem, Huawei's AI engineers went deep into the workshop and observed the master's craftsmanship, but to no avail.

What to do?

The final inspiration came on the spot - if I can't handle the master, can't I handle the parameters?

Therefore, Huawei's AI engineers introduced the 'Pangu Prediction Large Model,' shortening the optimization time for the hot-rolled production line to 3-4 hours by predicting the optimal parameters.

▲ Baosteel Hot Rolling Production Line

As a result, this application was launched on Baosteel's 1880 hot rolling production line, improving precision by over 5% compared to traditional AI, increasing the steel plate yield rate by 0.5%, and generating an annual revenue of over 90 million yuan.

Another super challenging task lies in the blast furnace with an internal temperature of 2,300°C and a volume exceeding 5,000 cubic meters.

This is not only a super-large 'black box' but also a 'Eight Trigrams Furnace' with extremely complex reactions. There are over 1,400 parameters in 8 major categories and 77 subcategories that affect the quality of molten iron; moreover, the substances inside the furnace exist in solid, liquid, and gas states, and each parameter may be interconnected and coupled with multiple other parameters.

Previously, changes inside the furnace were perceived through external sensors, but there was a reaction lag of 2-7 hours for different materials, which was barely useful.

As a result, it still relied on the master's experience, akin to mysticism.

In such an AI 'forbidden zone,' the Pangu Large Model managed to turn the furnace temperature, molten iron, silicon content, etc., inside the blast furnace into a 'Matrix' through 'furnace condition simulation,' transforming the 'black box' into a 'gray box' and then a 'white box,' achieving precise control.

This furnace condition optimization alone reduces costs by over 1 billion yuan annually for Baosteel.

In the field of industrial large models, these groundbreaking solutions have opened up a new path for the intellectualization of China's steel industry.

However, if we measure the difficulty of AI implementation by complexity, the steel industry is not even close to the ceiling. In the field of railway fault detection, the Chinese still face world-class challenges.

In a railway locomotive depot in northern China, the train inspection workshop.

Over 100 train inspectors are rapidly scanning images on the screen with their naked eyes. These images are instant snapshots captured by the detection system as the train passes by.

The job of train inspectors is to quickly identify problems and troubleshoot train faults. To this end, they must scan over 1,000 images in 10 minutes, working continuously for 12 hours.

This is an extremely tedious job that requires no slack.

There are tens of thousands of train components and hundreds of potential hazards, but none can be overlooked. For example, if a problem with the brakes is not detected and the train needs to make an emergency stop, there is a risk of derailment or overturning.

With a total railway mileage of 159,000 kilometers in China, relying solely on manual monitoring and troubleshooting is already overwhelming. However, achieving intelligent monitoring of such professional images is a super challenge.

Years ago, the railway department gathered a group of experts to develop a batch of AIs, but the results were costly and ineffective.

In 2023, Pangu arrived.

Soon, Huawei's AI engineers discovered a problem: early AIs struggled to determine 'what is abnormal,' ultimately failing in the face of ever-changing faults.

Therefore, Pangu adopted reverse thinking, teaching the large model to learn 'what is normal'; conversely, anything 'abnormal' is considered an 'anomaly.'

Once the mindset was opened, the possibilities were endless.

As a result, Pangu not only accurately identified over 430 types of faults but also detected significant hidden dangers such as angle cock faults, far exceeding the railway department's expectations.

Nowadays, in the field of high-speed rail, the Pangu Railway Large Model has been integrated with inspection robots, which can accurately identify 32,000 points on a single bullet train, covering 8 categories and over 350 complex faults, with an identification accuracy rate of 98%, effectively reducing the manual burden and significantly improving inspection efficiency.

In fact, every implementation of 'Pangu' in industrial scenarios is a process of overcoming insurmountable obstacles.

Whether it's 'parameter prediction,' 'digital simulation,' or 'reverse learning,' these require the deep integration of large models and machine vision, a thorough understanding of different industrial mechanisms, and adaptation in ever-changing scenarios, which is an extreme test of the basic capabilities of large models.

In this regard, ordinary language large models are merely 'giants in words but dwarfs in action,' incomparable to industrial large models.

For Huawei Cloud, only by gritting its teeth and overcoming insurmountable obstacles can 'Pangu' be integrated into real industrial scenarios, enabling AI to play a role in intelligent manufacturing upgrades and truly realizing a 'new industrial revolution.'

【The path to creation, those who bravely advance will prevail】

Today, the Pangu Large Model is driving Chinese manufacturing towards a smarter and more efficient direction.

At Cotte, Huawei Cloud has created Cotte AI Agents 2.0 based on the Huawei Cloud platform, Pangu Large Model, and 'Kunpeng + Ascend' computing power to meet its discrete manufacturing needs.

This endows Cotte with 'superpowers' such as flexible design, intelligent manufacturing, and flexible supply chains, disrupting the supply-demand relationship and reshaping corporate capabilities.

At Hailiang Copper, the 'Copper Foil Process Optimization Large Model' jointly created by Huawei Cloud has incorporated core data from 12 major processes, over 90 process and equipment parameters, over 30 quality inspection indicators, and 26 system platforms.

During this process, Huawei's algorithm and computing power PhDs went deep into the workshop to collaborate closely with Hailiang's materials science and mechanics PhDs, increasing the industry's highest yield rate from 90% to over 95%.

This 'small step' in yield rate has broken through the industry ceiling, reducing the R&D verification cycle from several days to just a few hours, directly generating tens of millions of economic benefits, and taking a 'giant leap' that has shaken the industry.

In the automotive field, the Pangu Automotive Large Model has compressed the R&D cycle of automakers such as FAW, Dongfeng, and Changan from six months to one week.

As a result, Huawei Cloud has closely cooperated with over 50 automakers to provide innovative services to over 300 automotive industry customers, becoming a leader in China's automotive large model field.

Behind this success lies not only Pangu's deep industry involvement and grounded approach but also its powerful foundational capabilities.

In June 2024, Pangu Large Model 5.0 was released, marking a significant transformation from 'not making poetry but only doing things' to 'solving difficult problems and tackling hard tasks.' Its confidence stems from significant upgrades in three dimensions: full series, multimodal, and strong thinking.

In terms of the full series, there are different parameter specifications such as billions, ten billions, hundreds of billions, and trillions, adapting to scenarios of different scales.

In terms of multimodal, it can understand information such as text, images, videos, radar, infrared, and remote sensing, bridging the gap between the digital and physical worlds.

In terms of strong thinking, the thinking chain technology is deeply integrated with strategic search, enabling intelligent entities to understand and predict environmental changes.

Some have commented that with the release of Pangu 5.0, AI has truly begun to touch the physical world.

This is undoubtedly good news for industrial large models.

With its dedicated focus on the industrial field, Huawei Cloud and the Pangu Large Model have won a series of industry accolades, including:

In the "IDC MarketScape: China Industrial AI Comprehensive Solutions 2024 Vendor Assessment," Huawei Cloud ranks in the Leader category, coming in first in terms of strategy, capability, and revenue.

In the "China Industrial Cloud IaaS+PaaS Market Share, 2023: Growing Differentiation," Huawei Cloud excels in the Chinese industrial cloud IaaS+PaaS market, ranking first in the overall market and three sub-markets.

The Pangu large model has not only become the industry's first large model product to receive an excellent (Level 5) rating in the China Academy of Information and Communications Technology's (CAICT) large model standard compliance verification but has also passed the CAICT's financial large model standard compliance verification with an outstanding (4+ level) rating. It also won the "Excellent Scientific and Technological Achievement Award" at the 2024 China International Big Data Industry Expo.

China's possession of the world's most complete and powerful industrial system is the result of generations of Chinese people's hard work, climbing up from the bottom to accumulate industrial advantages. It is the path blazed by our ancestors, undaunted by obstacles, to reach the summit, and it is also the strong confidence of today's Chinese people, who are not afraid of any sanctions or blockades.

Therefore, China's industry-specific large models must be firmly controlled by the Chinese themselves.

Moreover, they cannot be partial or incomplete; they must be comprehensive and full-stack.

Only Huawei has the ability to build a full-stack autonomous AI industrialization capability.

In this sense, Huawei's Pangu large model is following the same difficult but correct path as our ancestors, the "path of creation." It will surely help Chinese manufacturing take another leap forward.

--END--

Welcome to follow [HuaShangTaoLue] to meet influential figures and read tales of strategy and wisdom.

All rights reserved. Reproduction without permission is prohibited.

Some images are sourced from the internet.

If infringement is involved, please contact us for removal.