Paving the Way for All Intelligence
[Shanghai, China, September 19, 2024] Good morning, everyone! Welcome to Huawei Connect 2024. I hope you enjoy your time here in Shanghai.
At Huawei Connect 2018, Huawei launched its AI strategy and full-stack, all-scenario AI portfolio, positioning AI as a general-purpose technology. At the same event in 2021, we introduced our Pangu models designed to help industries go intelligent. AI has advanced by leaps and bounds in the intervening years, capturing the interest of global investors, industries, and governments alike. Over the past six years, we have steadily pressed ahead with strategic AI development, culminating in our All Intelligence Strategy, which we introduced at last year's Huawei Connect.
Every industry and enterprise has its own unique story when it comes to intelligent transformation. I've heard of many successes, but also a fair amount of uncertainty. Today, I'd like to take this opportunity to share our observations, thoughts, strategies, and experience.
Eric Xu, Huawei's Deputy Chairman and Rotating Chairman, speaking at Huawei Connect 2024
AI is becoming the most impactful technology for industries
To begin with, let's look at how far different industries have come in commercial AI adoption. From a commercial application standpoint, no other technology has had such a profound impact in such a short amount of time. According to research from McKinsey and Stanford University, AI adoption in all the covered industries is high across three main functions: product and/or service development, marketing and sales, and service operations. Gartner's 2024 CEO Survey reveals that CEOs are very optimistic about AI.
Overall, ongoing advancements in AI are driving more in-depth intelligent transformation across industries, laying the groundwork for All Intelligence.
Envisioning the future of enterprises in the age of AI
As companies set out on this journey, they hope to quickly harness AI for value creation today while also honing their competitive edge for the future. At Huawei, we've been exploring how to help our customers succeed on both fronts. In this regard, it's crucial to first envision what intelligent enterprises will ultimately look like, and then use that vision as a guide for the strategies and actions that we employ today.
Huawei has years of experience with its own intelligent transformation, and we've been helping a wide range of industries go intelligent too. Based on this experience, we've envisioned the future of intelligent enterprises in the age of AI. I'd like to discuss this vision in more detail, including what intelligent enterprises will look like and what features they will have.
These enterprises will be characterized by what we call the "six A's": They will have an Adaptive User Experience, Auto-Evolving Products, Autonomous Operations, an Augmented Workforce, All-Connected Resources, and AI-Native Infrastructure.
The first four A's are the results of intelligent transformation.
The first "A" focuses on how future enterprises will serve their customers. We believe it's through an Adaptive User Experience. That is, intelligent enterprises will perceive and understand user behavior, requirements, interests, tastes, and environmental changes, and then adapt to provide services that best meet user needs. The ability of products to promptly respond to a massive range of personalized requirements needs to be designed from the start – it's not a matter of incremental tailoring. For example, AI-powered educational tablets will be able to automatically adjust curriculum and difficulty levels based on a student's age, learning progress, comprehension skills, and test results, giving each student a unique learning experience at different stages of development. The transition from a predefined user experience to an adaptative user experience is a significant advancement, and will be necessary in the future intelligent world.
The second "A" stands for Auto-Evolving Products. It relates to innate product functionality – the ability to auto-evolve. In the age of AI, products will be able to self-learn, iterate continuously, adapt to changes, self-optimize, and self-evolve. For example, self-driving vehicles will be able to learn by themselves – the more they drive, the better they drive. The shift from digital to intelligent products is a major advancement that will reshape the competitive landscape. Every enterprise needs to think about adding intelligent capabilities to its products.
The third "A" stands for Autonomous Operations. It means closed-loop autonomous operations for all business flows, from sensing and planning to decision making and execution. Take smart ports, for example. Their intelligent planning platforms can automatically generate operation plans, while autonomous container trucks handle horizontal transportation. Many companies have been working towards automation for years now, and truly autonomous operations will bring a massive boost to efficiency. Every enterprise should consider using AI to transform and enhance operations more broadly and deeply.
The fourth "A" – Augmented Workforce – illustrates the future of employee experience and work methodology. It describes a workplace where every employee will have an intelligent assistant that understands their needs, helping them complete tasks more efficiently and with higher quality. For example, an assistant app can give field maintenance engineers quick access to mobile base station information such as fault location, root causes, and suggestions for handling the fault. AI exists to benefit humanity, and enhancing employee experience is key for any enterprise to get a head start in the age of AI.
The remaining two A's are the building blocks of AI.
All-Connected Resources is about connecting every part of an enterprise, from assets and employees to customers, partners, and ecosystems. All business objects, processes, and rules will be digitalized. Of course, the more information derived from connected resources, the better, but more importantly, the quality of information is key. Digitalization on a deeper and broader scale will generate the much-needed data for enterprise intelligence.
The sixth and last "A" is AI-Native Infrastructure. There are two aspects to AI-Native Infrastructure: "ICT for Intelligence" and "Intelligence for ICT". ICT infrastructure needs to be built out systematically to keep up with the needs of intelligent applications – that's ICT for Intelligence. And on the flip side, AI technology itself will be vital for managing the O&M of future ICT infrastructure, as well as assuring experience. This is what we call Intelligence for ICT.
With these six A's, we're trying to capture what a future intelligent enterprise will look like. We hope these can help provide some food for thought on how your company can use AI to position itself for success in the intelligent world to come.
Advancing Huawei's All Intelligence Strategy
At Huawei Connect 2023, we announced our All Intelligence Strategy. This strategy addresses a broad range of areas, and today I'd like to discuss our thoughts on seven specific initiatives.
I. Innovative architecture for sustainable computing solutions
Let's start with computing power. Intelligent transformation is a long-term process, and it's founded on computing power – not only now, but in the future. So moving forward, sustainable computing power will be the cornerstone of continuous advancements in AI.
Computing power largely depends on semiconductor manufacturing process nodes. However, the reality is that US restrictions on AI chips for China are unlikely to be lifted anytime soon, and the Chinese mainland will lag behind in semiconductor manufacturing process nodes for a relatively long time. This undermines our ability to make advanced chips. It's a challenge that we have to confront when developing computing solutions. In the Chinese mainland, sustainable computing power can only be achieved with chip manufacturing process nodes that are practically available.
These challenges are real, but they also present opportunities and new possibilities. They are what motivate us to keep innovating. AI is becoming the predominant source of demand for computing power – and this trend is driving structural changes in computing systems. We're now looking at how to increase computing power at the system level rather than the individual processor level.
These structural changes give us the opportunity to create a new computing architecture – a viable path for the independent and sustainable development of the computing industry.
We want to seize the opportunities presented by this AI revolution based on chip manufacturing process nodes that are practically available. And so, our strategy is to create a new computing architecture that is built on synergistic innovation across computing, data storage, and networks. We will also develop computing supernodes and clusters to sustainably meet long-term demand for computing power.
II. Huawei Cloud's upgraded stack for AI: Powering industries with AI
The technological breakthroughs we've seen in foundation models have greatly accelerated advancements in AI. It seems like every industry is talking about foundation models, trying to build their own AI computing infrastructure, or even train their own proprietary model these days. This is undoubtedly great news for computing solution providers like Huawei. But we don't believe this is viable in the long term. As always, we want our customers to achieve ongoing, lasting success, which is key to our own sustainable development. So today, I'd like to share some of our thoughts on this topic.
Not every company needs to build their own large-scale AI computing infrastructure. AI servers, especially AI computing clusters, are different from general-purpose x86 servers in that they need more advanced power supply and cooling systems. As foundation models become increasingly larger, demand for AI computing keeps growing, and AI servers need to be constantly upgraded. If companies want to keep up by upgrading their own data centers, they may end up under-utilizing their resources or falling behind demand.
AI hardware products have fast iteration cycles, with an average one- to two-year gap between generations. Large models are rapidly evolving. Limited by the computing power of each generation, companies need to combine multiple generations of hardware to train a large model, but this greatly complicates resource scheduling. Older generations end up handicapping the performance of the newer generations, which in turn undermines the training of large models.
Companies also face O&M challenges. AI technology is evolving rapidly these days, so multiple generations of AI products tend to co-exist in a single DC. O&M is getting difficult and requires in-depth expertise, which poses a major challenge for companies that only have IT maintenance capabilities.
These challenges will persist for some time. Beyond building their own AI computing infrastructure, enterprises might want to consider alternatives to get the AI computing power they need in the way that works best for them.
Not every company needs to train a proprietary foundation model. Data is key to foundation model training. Yet it's difficult and costly to obtain enough quality data for pre-training a foundation model with roughly 10 trillion tokens. The number of model parameters keeps growing, which makes iteration and optimization extremely difficult. Iterative model training usually takes months or even years. This can greatly delay the benefits of AI in core business areas. And beyond that, talent is hard to come by. Foundation model technologies are evolving every day, and there is a distinct lack of technical experts with practical experience in this area.
Not every application needs a large model. Huawei's Pangu models have been used in many industries, and our experience suggests that a 1-billion-parameter model is enough for scientific computing and prediction scenarios, such as rain forecasts, drug molecule optimization, and technical parameter predictions. Models like this are also widely used for devices like personal computers and smartphones. Models with over 10 billion parameters are sufficient for domain-specific tasks, like natural language processing (NLP), computer vision (CV), and multi-modal tasks. Examples include knowledge base question answering, coding, banking agent assistants, and safety risk detection. Complex NLP or multi-modal tasks can be handled with 100-billion-parameter models.
In our view, companies need to choose the right models for the right scenarios, and use a mix of models to address their pain points and create value.
So far, I've shared our thoughts on AI development. And we believe that if a company doesn't have the ability or resources to build their own AI computing infrastructure or train their own foundation model, then cloud services are a more feasible, sustainable option. Huawei Cloud has upgraded its entire stack for AI to address these challenges. Our goal is to give every company access to on-demand AI computing power, and enable more efficient model training and inference.
First, Huawei Cloud is continuously building up its Ascend Cloud Service to give companies easy access to massive AI computing power. With Ascend Cloud Service, companies don't need to build or upgrade their own data centers, or operate or maintain AI computing infrastructure. With end-to-end synergy across computing, data storage, and networking, Ascend Cloud Service has been used to successfully train a model with over 100 billion parameters, uninterrupted for 40 days straight.
Second, Huawei Cloud has upgraded the ModelArts services to provide out-of-the-box access to mainstream foundation models, including Pangu models, open-source models, and third-party models. This means companies don't have to prepare tons of data or go through multiple iterations for foundation model training themselves. Huawei Cloud also provides complete toolchains for model tuning, deployment, and testing, lowering technical barriers to model fine-tuning and incremental training.
Third, Huawei Cloud is going all out to build up Pangu Models 5.0. This series includes models with over 1 billion, 10 billion, 100 billion, and even more parameters. We've developed this range of models so all companies can find the best fit for their unique needs and business scenarios. We've also established a model developer community with more than 100 models, offering companies a broader range of choices.
To sum up, cloud services are the best option for many companies that are looking to incorporate AI into their business. By providing Ascend Cloud Service and AI model services, we aim to give every company real-time access to on-demand AI computing power, and enable more efficient model training and inference.
Huawei Cloud's systematic security capabilities: Ensuring the security of large model training & inference
It's important to note that AI model training and inference on the cloud has brought about new security challenges. In response, Huawei Cloud has greatly enhanced its security capabilities.
Our security approach is to ensure defense by design for the most severe attacks. Based on zero trust, we've built seven layers of defense, covering physical security, identity verification, networks, applications, hosts, data, and O&M. We've also developed a unified security operations center. This security posture has helped Huawei Cloud successfully defend against 1.2 billion attacks each day, ensuring zero service interruption, zero data loss, and regulatory compliance.
In terms of security mechanisms, Huawei Cloud's Graded Security Cloud provides customers with a secure digital space, with support for both physical and logical isolation. Cloud platform operations are both transparent and auditable so customers can use cloud services with confidence.
In terms of security technology, Huawei Cloud offers an end-to-end, full-stack data security protection solution to manage the entire data lifecycle at the hardware, software, and app layers. This guarantees comprehensive data security throughout data movement, model training, and inference, and ensures the end-to-end security compliance of training data and generated content.
In terms of intellectual property (IP) infringement, Huawei will defend our customers against any third-party IP infringement claims related to content generated by Huawei Cloud large model services, and will compensate customers for the losses, costs, or expenses incurred as the result of any final judgment or settlement. [1]
III. Harmony Intelligence: Providing an intelligent experience across all scenarios
Devices are a crucial part of an AI-empowered future. Huawei was the first company to bring AI to smartphones. Back in 2017, we launched the HUAWEI Mate 10 smartphone with AI chips. It was the first phone with AI capabilities for imaging, translation, and other functions. With this device, we kicked off the age of Mobile AI.
Now, we find ourselves in the age of foundation models. With architecture that maximizes synergy between devices, chips, and cloud, we have deeply integrated AI technology into HarmonyOS. The result is Harmony Intelligence, with AI at its core. It is powered by intelligence at all layers, from the kernel to system apps. It allows for more open ecosystem-wide collaboration, as well as more trustworthy privacy protection and security protection.
With Harmony Intelligence, our smart assistant Celia is evolving into an AI agent, capable of more intuitive multi-modal interaction and more converged sensing. Celia will be able to accurately perceive user intent as well as the digital and physical worlds, and offer personalized content and intelligent services for all different types of scenarios.
Together with HarmonyOS ecosystem partners, we will take the intelligence of products further to better meet consumer needs across all scenarios, from office and learning to lifestyle and entertainment. And we are implementing tiered exposure of AI model capabilities and controls to support third-party apps and foster a thriving HarmonyOS-native app ecosystem.
Overall experience, not computing power, is central to on-device AI
Equipping devices with AI capabilities has become the new norm. These days we see all sorts of AI phones and AI PCs. But how do we define what a "smart device" is in the age of AI? Opinions on this vary throughout the industry.
Our position is that user experience should always come first. Smart device users are more interested in overall experience than in obscure technical specs, such as chip process nodes, Tera-FLOPS, and the number of model parameters. So we advocate that overall experience – not computing power – should be central to on-device AI.
With this as our starting point, Huawei has teamed up with the Institute for AI Industry Research at Tsinghua University to propose standards for defining the intelligence levels of AI-powered devices, from L1 to L5. Our goal is to give consumers a more distinct feel for the capabilities of different AI-powered devices, and to promote industry consensus on the evolution of AI-powered device capabilities. The hope is that consensus can help drive coordinated development across the industry.
Specifically, we've adopted an experience-centric approach for quantifying intelligent user experience. We hope to deliver a better experience by reaching higher levels of intelligence in AI-powered devices. And we look forward to working with all industry players to further refine these standards and propel coordinated development in on-device AI.
IV. Autonomous Driving Network (ADN): Reshaping network experience and O&M
In the network industry, Huawei was the first to advocate using AI for telecom networks, and we proposed the concept of an Autonomous Driving Network (ADN) back in 2018. Now, we are integrating the Telecom Foundation Model and digital twins into ADN. And we are working with the TM Forum, China Mobile, and other partners to support level-4 high autonomous networks in high-value scenarios. We aim to gradually reach level-4, and eventually level-5 full autonomous networks.
In terms of ADN for Telcos, we are committed to providing superior user experience with zero wait, zero interruption, and zero touch. We also aim to support simplified O&M with self-configuration, self-healing, and self-optimization.
At the same time, we've started applying ADN to enterprise networks, because O&M in this domain has its fair share of challenges too. First, it's difficult to guarantee a positive experience for employees in a fully wireless office with huge demand for cloud-based applications and video. Second, O&M workloads grow in scope and complexity as networks grow larger, whether it be office and production networks, data center networks, branch networks, or cloud networks. And this is further compounded by increased diversity of network elements.
Today, we are introducing ADN for Enterprises, which will guarantee zero service delays, zero network disruptions, zero-wait service provisioning, and zero security risks.
V. Autonomous driving solutions: Prioritizing safety and experience, and paving the way for a driverless future
Autonomous driving for vehicles was one of the key focuses of our initial AI investment, because the end goal of autonomous driving is fully unmanned driving – one of the most challenging AI applications out there. Huawei's ADS 3.0 allows for more accurate autonomous driving decisions, more efficient mobility, more human-like driving experience, and greater driving safety. The solution's Navigation Cruise Assist (NCA) enables one-tap automatic cruising – on both public and internal roads – to the parking lot at your destination, whether it's above or underground. We've also upgraded our all-directional collision avoidance system to support emergency braking in a wider speed range and all-directional obstacle avoidance.
These advances have really given consumers a feel for how much intelligent driving can improve their overall mobility safety and experience. Chinese consumers are already very familiar with the value it provides, so many of the cars people buy these days are equipped with an advanced intelligent driving system. In essence, intelligent driving capabilities have become a key consideration for Chinese consumers when they're looking to buy a new car.
Moving forward, we will leverage fusion sensing technology to evolve ADS and gradually build up to several key goals. We will enable autonomous driving on highways; safe, stable driving on urban and suburban roads; and versatile driving across different terrains in rural and mountainous areas. We will also provide auto valet parking with features like zero scratches and zero software crashes. Through proactive safety, we will deliver enhanced, all-directional collision avoidance. These are our future goals for key driving scenarios. Our ultimate goal is to enable unmanned driving, and we will keep working hard to make it a reality.
VI. Jointly building ecosystems and creating a unified developer platform for shared success
Ecosystem development has always been a key component of our overall strategy. We have and will continue to build up ecosystems together with our partners, and will keep expanding our unified developer platform to promote shared success. From 2017 to 2019, we began our work on multiple ecosystems, like those for Huawei Cloud, Ascend, Kunpeng, and HarmonyOS.
In 2024 and over the next five years, Huawei will invest even more into ecosystem development in an effort to guide and drive broader development in the computing and device industries. Our goal is to offer the world a second option for computing, and a third option for mobile OS.
VII. Advocating and practicing AI for good: Contributing to human, societal, and environmental well-being
AI has an unlimited set of applications, but they all come down to serving people. At Huawei, we advocate and practice AI for good.
We believe that AI should serve people by improving efficiency and quality of life. AI can enable the digital transformation of industries, reshaping production and paving the way towards an intelligent world. We aim to develop AI systems that are accessible to every person, home, and organization.
AI should be used for good – to create greater value for society. During the design, development, and use of AI technologies, it's crucial to carefully evaluate their potential and long-term impact on society, and take necessary measures to prevent harmful application.
AI should be used to protect the natural environment and promote sustainable development. It's important to actively leverage AI to study and address issues of global concern, such as the United Nations' Sustainable Development Goals.
The age of All Intelligence is here. It is unlocking new opportunities and new challenges for everyone and every enterprise. So let's work together to pave the way for All Intelligence, providing every person with an intelligent personal assistant, helping every company become an intelligent enterprise, and powering every vehicle with autonomous driving.
Thank you!
[1] Subject to the terms and conditions of individual contracts.