huawei's rotating chairman xu zhijun's latest speech: not all applications need to pursue a "big" model
2024-09-19
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
on september 19, huawei's official website posted the speech of huawei's vice chairman and rotating chairman xu zhijun at the 2024 huawei connect conference. xu zhijun stated that ai has become the technology with the greatest impact on the industry. huawei proposed a comprehensive intelligent strategy last year, and the core of the strategy is to seize the opportunities of artificial intelligence transformation.
xu zhijun said, "the advancement of the chips we can manufacture will be restricted. this is a challenge we must face in creating computing solutions."
in his opinion, because artificial intelligence is becoming the dominant computing power demand, computing systems are undergoing structural changes, requiring system computing power rather than just the computing power of a single processor. these structural changes provide opportunities to create an independent and sustainable computing industry development path through architectural innovation.
but he also said that the technological breakthrough of big models has greatly accelerated the process of intelligence. for a period of time, almost all walks of life have mentioned big models, and have built ai computing power and trained big models. however, from the perspective of long-term development, not every company needs to build large-scale ai computing power.
"we all know that ai servers, especially ai computing clusters, are different from general-purpose x86 servers. they have extremely high requirements for data center computer room environment, such as power supply and heat dissipation. as large models become larger and larger, ai computing power will also move towards a larger scale and change at a fast pace. ai servers are rapidly upgraded and replaced, and data center computer rooms are faced with the dilemma of either wasting or not being able to meet demand."
in addition, xu zhijun believes that not all applications need to pursue "big" models. a billion-parameter model can meet the needs of business scenarios such as scientific computing and prediction and decision-making, while a 10-billion-parameter model can meet the needs of a large number of specific field scenarios such as nlp, cv, and multimodality, such as knowledge question and answer, code generation, agent assistant, and security detection. complex tasks for nlp and multimodality can be completed with a 100-billion-parameter model.
the following is the full text of the speech (abridged version)
huawei vice chairman and rotating chairman eric xu: embracing the era of comprehensive intelligence
welcome to the huawei connect 2024. i hope you have a pleasant journey in shanghai. at the huawei connect 2018, i released huawei's artificial intelligence development strategy and full-stack, full-scenario ai solutions, and positioned ai as a general-purpose technology. in 2021, at the huawei connect 2021, i talked about how the pangu big model enables intelligence in all walks of life. from 2018 to the present, the development of ai has been changing with each passing day, and has attracted great attention from the global investment community, industry, and government. since 2018, huawei has steadily promoted its ai development strategy, and at last year's huawei connect 2024, it further clarified the company's comprehensive intelligence strategy. for intelligence, every industry and every enterprise has its own exploration. i have heard that many achievements have been made, but i have also noticed that there are still many confusions. today, i would like to take this opportunity to share our observations, thoughts, strategies, and practices.
ai has become the technology with the greatest impact on the industry
first, let's look at the commercial progress of ai. from the perspective of business applications, no technological advancement has ever had such a great impact in such a short period of time as ai. research by mckinsey and stanford university shows that ai applications in various industries are currently mainly concentrated in three aspects: product development, marketing, and business operations. secondly, from the perspective of corporate executives, gartner's survey results show that ceos have a very positive view of ai. so in summary, the continuous advancement of ai technology is driving the continuous deepening of intelligence in all walks of life and is moving towards comprehensive intelligence.
standing at the beginning of the era of comprehensive intelligence, each of our companies not only hopes to use ai to create value as soon as possible today, but also hopes to achieve leadership in the future intelligent competition. this is also a question we have been thinking about. i would like to take this opportunity to share our vision for companies in the intelligent era, that is, what companies that look forward to the intelligent era will look like and what characteristics they will have.
we believe that enterprises in the era of intelligence should have "six a" characteristics. the first four a's represent the effects of intelligence, including:
the first a is about how companies need to serve their customers in the future. we think it is adaptive user experience, which means that intelligent companies should be able to perceive and understand user behavior, needs, interests, tastes, and environmental changes, and actively adjust to provide services that best meet user needs. products that can meet a large number of personalized and unique needs in a timely and simultaneous manner need to be specially designed from the beginning, not just tailored. for example: ai learning machines automatically adjust teaching content and difficulty based on students' age, learning progress, comprehension ability, and test feedback, so that each student can get a learning experience that suits them at different times. providing customers with a preset, certain experience to an adaptive experience is a leap, and every company needs to provide a customer experience that adapts to the intelligent era.
the second a answers what kind of products companies will need to create. we believe that they are auto-evolving products, which means that products in the intelligent era will have the ability to learn autonomously, continuously iterate, adapt to changes, and be able to self-optimize and self-evolve. for example, self-driving cars will become easier to drive. the transition from product digitization to product intelligence is a leap that will greatly change competition. every company needs to think about integrating intelligent capabilities into its products.
the third a answers the future of daily business operations, namely autonomous operation, which means achieving highly autonomous business operations, from perception, planning, decision-making to execution, and end-to-end autonomous closed loop. for example, ports can automatically generate operation plans through intelligent planning platforms, and automatically complete horizontal container transportation through self-driving container trucks. enterprise operation automation has been pursued by many companies for many years. the autonomy of operations is a leap forward in improving operational efficiency. every company needs to think about using ai to empower and change business operations in a wider and deeper range.
the fourth a answers the future of employee work experience and work style, namely augmented workforce, which means that every employee should have an intelligent assistant that "understands me" to complete every task efficiently and with high quality. for example, the on-site maintenance personnel of the operator's base station can quickly obtain information such as the fault location, root cause of the fault, and handling suggestions through the maintenance assistant app. the meaning of ai's existence is to make ai benefit mankind, and to give employees a better work experience is the key foundation for the competitiveness of every enterprise in the era of intelligence.
the next two a's represent the foundation of intelligence. the fifth a, all-connected resources, means to achieve full interconnection of the company's assets, employees, customers, partners, and ecosystems, and to digitize all business objects, processes, and rules. it is necessary not only to increase the quantity of information, but also to improve the quality of information, so that the company has the data and information foundation necessary for intelligence, that is, in-depth and comprehensive digitization.
the sixth a stands for ai-native infrastructure. it means that, on the one hand, ict infrastructure must be constructed in a systematic manner to adapt to the needs of intelligent applications, namely ict for intelligence; on the other hand, the operation and maintenance management and experience assurance of the infrastructure itself must be fully intelligent, namely intelligence for ict.
the characteristics of these 6 a’s are our initial thoughts and summary based on our own practice and understanding. we hope that they will be helpful for everyone to think about how to make good use of ai. they are provided for your reference. we hope that every company can become a winner in the era of intelligence.
seize the opportunities of ai transformation
huawei proposed a comprehensive intelligent strategy at the hc conference in 2023. the comprehensive intelligent strategy covers a wide range of areas.
first, let's talk about computing power. intelligence will inevitably be a long process, and computing power is the key foundation of intelligence, both in the past and in the future. therefore, the sustainability of intelligence depends first on the sustainability of computing power. computing power depends on semiconductor technology, but we must face a reality, that is, the us sanctions on china in the field of ai chips will not be lifted for a long time, and china's semiconductor manufacturing technology will lag behind for a long time due to us sanctions.this means that the advancement of the chips we can manufacture will be limited. this is the challenge we must face in creating computing solutions.
based in china, only computing power built on actual available chip manufacturing processes is sustainable in the long run. otherwise, it is unsustainable. huawei sees challenges, opportunities and possibilities, which further inspires our passion for innovation.artificial intelligence is becoming the dominant computing power demand, prompting a structural change in computing systems. what is needed is system computing power, not just the computing power of a single processor.these structural changes provide us with opportunities to create an independent and sustainable computing industry development path through architectural innovation.
the core of our strategy is to fully seize the opportunities brought by the ai revolution, based on the actual available chip manufacturing processes, collaborative innovation in computing, storage and network technologies, create a computing architecture, and build a "super node + cluster" system computing power solution to continuously meet computing power needs in the long term.
the technological breakthrough of big models has greatly accelerated the process of intelligence. for a period of time, almost all walks of life have mentioned big models, and have built ai computing power and trained big models. this is undoubtedly a major benefit for computing power providers like huawei. however, from a long-term development perspective, we always believe that only the continued success of our customers can lead to the continued development of huawei.
first, not every enterprise needs to build large-scale ai computing power. we all know that ai servers, especially ai computing power clusters, are different from general-purpose x86 servers. they have extremely high requirements for data center computer room environment such as power supply and heat dissipation. as large models become larger and larger, ai computing power will also move towards a larger scale and change at a fast pace. ai servers are rapidly upgraded, and data center computer rooms are faced with the dilemma of either wasting or not being able to meet demand.
secondly, the industry now launches new ai hardware products every one to two years on average, with a fast iteration speed. compared with public clouds, enterprises are limited by the small scale of computing power. faced with rapidly changing large models, it is difficult for each generation of computing hardware to complete the work independently. instead, they hope to use multiple generations of products in combination for model training. this leads to high complexity in resource scheduling. in addition, due to the "short board" effect of historical generational products, the performance of the new generation of products is dragged down and the ability to train large models is affected.
finally, there are challenges brought by operation and maintenance. ai technology is still in its growth stage, with rapid technological changes, the coexistence of multiple generations of products, and high skill requirements, which makes operation and maintenance difficult. this is a major challenge for many companies that only have traditional it maintenance capabilities. as these challenges will continue to exist for some time,therefore, i believe that every company should think about how to obtain ai computing power that suits them, rather than just building their own ai computing power.
in addition, not every enterprise needs to train its own basic large model. the key to training a basic large model is data, and preparing enough high-quality data is a big challenge. the amount of pre-training data for basic large models has reached the level of 10 trillion tokens. for enterprises, this not only means high costs, but also whether they can obtain enough data is also a challenge.
secondly, model training is difficult. the number of parameters in the basic large model continues to increase, and model iteration and optimization are difficult. it usually takes months to years to complete model iteration training. each company should focus on its core business. training the basic large model by itself will affect ai's ability to empower core business as soon as possible.
finally, it is difficult to acquire talent. the relevant technologies involved in basic large models are updated every day, and there are few technical experts with practical experience. for enterprises, it is also a challenge to establish sufficient technical talent resources.
not all applications need to pursue "big" models. from huawei pangu's practice in the industry, the billion-parameter model can meet the needs of business scenarios such as scientific computing, prediction and decision-making, such as rainfall prediction, drug molecule optimization, and process parameter prediction. the billion-parameter model is also widely used on end-side devices such as pcs and mobile phones. the 10-billion-parameter model can meet the needs of a large number of specific field scenarios such as nlp, cv, and multimodality, such as knowledge question and answer, code generation, agent assistant, and security testing. complex tasks for nlp and multimodality can be completed with a 100-billion-parameter model.
therefore, we believe that what enterprises need is to choose the most appropriate model based on the needs of their own different business scenarios, solve problems and create value through a combination of multiple models.
terminal ai should not be centered on computing power
in the era of intelligence, terminals are an indispensable part. in the field of terminals, huawei was the first to introduce ai into smartphones. as early as 2017, huawei launched the mate10, which had a built-in ai chip and applied ai smart imaging, ai translation and other capabilities to mobile phones for the first time, ushering in the era of mobile ai. today, as ai enters the era of big models, we have deeply integrated ai technology with the hongmeng operating system based on the architecture of terminal, chip and cloud collaboration, and rebuilt the ai-centered hongmeng native intelligence, achieving comprehensive intelligence from the kernel to system applications, while achieving more open ecological collaboration and more reliable privacy and security protection.
based on the native intelligence of harmonyos, huawei will upgrade "xiaoyi" into an intelligent entity, achieve more natural multimodal interaction and more comprehensive integrated perception, and work with harmonyos ecosystem partners to jointly build intelligent capabilities for future products; and achieve full openness from ai model capabilities to ai control layers, enable third-party applications, and prosper the native application ecosystem of harmonyos.
we have also noticed that it has become a general trend to introduce ai capabilities into various terminals, such as building ai phones and ai pcs. as a result, there are various opinions in the industry on how to define smart terminals in the ai era. we always believe that the consumer experience is the first priority. consumers find it difficult to understand what chip technology, computing power tflops, model parameter quantity... actually mean, but they pay more attention to their personal experience. therefore, we advocate that terminal ai should be experience-centered rather than computing-centered.
based on this concept, we and tsinghua university artificial intelligence industry research institute jointly proposed the ai terminal intelligence l1 to l5 classification standard, and we look forward to colleagues in the industry to improve and optimize the classification standard and jointly promote the orderly development of terminal ai.
we hope to achieve driverless driving by around 2030
the car autonomous driving solution is also an important area of ai that huawei invested in at the beginning, because the goal of autonomous driving is unmanned driving, which is one of the most challenging scenarios for ai application. the ads 3.0 version we launched can make autonomous driving decisions more accurate, traffic more efficient, the experience more human-like, and driving safer. it also realizes "one-click" arrival from parking space to parking space, and connects all scenarios from public roads to campus roads to underground parking spaces. and further upgrade the omnidirectional collision avoidance system to cover more speed ranges and achieve omnidirectional obstacle avoidance.
these advances have allowed consumers to truly feel the safety and experience improvements brought by intelligent driving. now, chinese consumers are very familiar with intelligent driving of cars. the proportion of new cars equipped with advanced intelligent driving versions is very high, and the intelligent driving capabilities of cars have become a key consideration for chinese consumers when buying new cars. next, we will continue to evolve autonomous driving solutions based on fusion perception, and gradually achieve: on highways, you can rest as soon as you get in the car, and sleep peacefully on long journeys; on urban and suburban roads, it is easy to drive everywhere, safe and stable like an experienced driver; in rural and mountainous roads: go up the mountains and go down the countryside, and drive with confidence in all terrains and all weather conditions. in parking scenarios: achieve leaving the car and going, zero scratches, and zero stuck; in terms of safety, we must achieve all-round and all-directional active safety, mainly to clear the main responsibility collision and reduce secondary responsibility. on the basis of achieving these key scenario goals, we hope to achieve driverless driving around 2030.
building a unified developer platform
developing an ecosystem has always been an important part of huawei's strategy. in 2024 and the next five years, huawei will make strong strategic investments in the development of the ecosystem. through the development of the ecosystem, it will guide, promote, and drive the development of the computing and terminal industries, provide a second choice for the world's computing field, and provide the world with a third mobile operating system.
the applications of ai will be endless, but in the final analysis, it is to serve people. we insist on advocating and practicing ai for good. we believe that: ai should serve people, improve people's work efficiency and quality of life; enable industry digitization through ai, change the industry's production methods, and become the core engine for various industries to enter the intelligent world; we must lower the threshold of ai technology so that everyone, every family, and every organization have equal opportunities to obtain and use ai technology.
ai should be used for benign purposes such as creating greater welfare for society; in the design, development and use of ai, we will carefully evaluate the long-term and potential impact of ai technology on society to avoid the abuse of ai technology. in addition, ai should be used for ecological environmental protection and sustainable development, and actively use ai to study and solve global issues of concern, such as the united nations sustainable development goals.
the era of comprehensive intelligence has arrived, bringing new opportunities and new challenges to everyone and every company. let us work together to promote comprehensive intelligence, so that everyone has their own exclusive smart assistant, every company becomes an intelligent company, and every car can be driverless.
(this article comes from china business network)
