Nine years ago， with AlphaGo’s victory over Lee Sedol in a Go match， few could have predicted that this triumph would usher artificial intelligence into our lives with such a powerful presence.
Initially， people were surprised to find that AI wasn’t some esoteric， complex concept. It could manifest as a readily accessible chatbot or a tool capable of generating images and videos from a single prompt.
However， it wasn’t long before AI’s reach began to extend further.
New product launches for phones and computers now seem incomplete without an AI focus， as if lacking a soul. When you go to the hospital for an X-ray， it’s possible the initial diagnosis won’t come from a doctor. Meanwhile， the myriad of AI features in various apps can be quite dazzling. The development of embodied AI， reminiscent of science fiction films where humans coexist with robots， has taken a peculiar turn.
Beyond performing as a backup dancer on national television， one might encounter robots or robotic dogs trying to attract business while shopping， with various marketing events eagerly featuring robots to liven up the atmosphere.
Furthermore， robots are exhibiting an increasing array of capabilities. Just a few years ago， they moved like an octogenarian， but now they can run， jump， perform backflips， and even box – some have even organized humanoid robot sports events. The entire AI industry is progressing at an astonishing pace.
To put it plainly， if you’re still talking about “Internet Plus，” it’s time for some serious catching up. We have firmly stepped into the “Artificial Intelligence Plus” era.
Every industry is contemplating， or has already begun， to leverage AI to reshape and disrupt existing models， a race to avoid being left behind by the times. This advancement is， of course， inseparable from enterprises pushing the boundaries of technological innovation. However， upon closer inspection， it’s evident that behind every seemingly brilliant AI solution， there are cloud service providers shouldering significant operational demands.
The mention of cloud service providers might induce a yawn in some， but let’s delve into something more engaging. Consider embodied AI: do you truly believe that robots capable of running and jumping evolved solely on their own? That’s a rather simplistic view.
Frankly， the hurdles that embodied AI needs to overcome are substantial. Challenges include data scarcity， software algorithm limitations， and insufficient generalization capabilities. In essence， their “brains” are not yet sophisticated enough – they might manage a flashy “three-step dance，” but for real-world tasks， they are still quite far from being truly competent.
Data scarcity alone is a significant impediment for these robots. The industry has made considerable efforts to address the issue of high-quality data scarcity. Real-world data is both expensive and limited. NVIDIA， for instance， developed the Omniverse platform， which allows robots to undergo rigorous training within a digital twin environment. There are also dedicated data collection facilities that utilize real robots for offline data acquisition.
Building datasets is also a standard practice. For example， the Beijing Institute of Humanoid Robotics is accelerating the creation of a high-quality dataset comprising millions of data points. However， generating a million data points doesn’t happen spontaneously; dataset creation involves production， collection， and annotation – tasks that are incredibly demanding when undertaken independently.
Specialized tasks require specialized teams. This is why the Beijing Institute of Humanoid Robotics partnered with Baidu AI Cloud， given their extensive experience in autonomous driving， which includes data collection and annotation.
More importantly， Baidu AI Cloud holds a dominant position in the AI public cloud market， so its capabilities are validated by market data. Yet， if it were merely about data handling， engaging a specialized data company might suffice. However， mastering data is merely the first step in the robots’ long journey. The subsequent stages involve resource-intensive model training， simulation testing， and model inference.
Therefore， collaborating with a cloud service provider like Baidu AI Cloud， which offers comprehensive full-stack AI development capabilities， proves to be more efficient.
In terms of computing power， Baidu AI Cloud’s flagship GPU， Baige， is a formidable force in both training and inference. For scenarios like robotics， which require repeated simulations and strategy optimization， single-card performance is insufficient. Baige excels in ensuring the stability and efficiency of massive thousand-card cluster collaborations.
With adequate computing power， there are no longer any hesitations in undertaking ambitious projects. To enhance the intelligence of robots， Baidu AI Cloud offers the Qianfan large model platform， empowering technical personnel to innovate freely.
For instance， the robots developed by the Beijing Institute of Humanoid Robotics， with the assistance of Baidu AI Cloud’s Qianfan large model platform， can comprehend human language (through precise natural language processing) and interpret the visual world (via multimodal information processing). It’s highly probable that Baidu AI Cloud’s contributions played a significant role in the success of their TianGong robot， winning championships in both the 100-meter dash and half-marathon.
This synergistic collaboration is not an isolated case; it has become the standard practice in the embodied AI industry. Companies like Unitree Robotics， a prominent player， have also opted for Baidu AI Cloud’s AI services. Baidu AI Cloud clearly recognizes its role as an industry infrastructure provider and is continuously evolving.
At the recent Baidu Cloud Smart Conference， Baige and Qianfan unveiled new iterations. Baige 5.0 enhances training and inference efficiency. With the introduction of Kunlun Core super nodes， capable of running trillion-parameter models in mere minutes， we might even see synchronized figure skating added to next year’s robot sports competitions.
Furthermore， the upgraded Qianfan 4.0 features the RFT toolchain， which enables model fine-tuning with just a few hundred data samples to achieve results comparable to using tens of thousands of data points. This is a significant boon for the data-starved embodied AI sector.
Beyond robots navigating the physical world， let’s explore the extent to which AI is revolutionizing the gaming landscape. Even game NPCs are now striving for a more lifelike presence. For example， the AI town ‘Aivilization’ developed by the Hong Kong University of Science and Technology boasts 100，000 intelligent agents， each exhibiting human-like behavior.
Image Source: Youtube
The challenge with lifelike NPCs is that manual scripting is simply not feasible at scale. While large models can be employed， their effectiveness hinges on their intelligence and the ability to manage simultaneous interactions with numerous players without performance degradation.
NetEase Fuxi’s mobile game， ‘Justice League，’ is a prime example where AI NPCs engage in free-flowing conversations with players， akin to a sophisticated game interface for ChatGPT. Some players have become so engrossed in conversing with NPCs that they’ve practically turned ‘Justice League’ into a dating simulator game.
Image Source: Xiaohongshu user
In this endeavor， NetEase Fuxi collaborated with Baidu AI Cloud. Baidu AI Cloud’s Wenxin series of persona large models imbue AI NPCs with the capability to understand， reason， and embody specific personalities. Concurrently， the underlying computing power of Baige ensures the stability of the game under high concurrent access， thereby guaranteeing player experience.
Combined with the recent upgrades to Qianfan 4.0， Baidu AI Cloud has not only enriched its model library but also significantly streamlined Agent development， opening up a world of possibilities for the gaming industry.
Finally， let’s consider the AIGC applications we interact with daily， such as text-to-image and text-to-video generation. These are all highly compute-intensive processes.
Technically， ensuring the stability and consistency of generated results requires extensive model training， and hardware， system configuration， and software failures can even interrupt the training process. Therefore， for most users， it is more practical to delegate these demanding tasks to cloud service providers.
For instance， Shengshu Technology has leveraged Baidu AI Cloud’s Baige heterogeneous computing platform. While the mention of thousand-card clusters is common， the key highlight here is achieving an effective training duration exceeding 98.8%. Baige also offers a suite of operational tools to ensure the stability of the thousand-card cluster， preventing disruptions during critical training phases.
The recent upgrades to Baige 5.0 have further optimized network and inference systems， leading to faster communication， lower latency， and more intelligent scheduling strategies to enhance inference system throughput， pushing the efficiency of AI training and inference to new heights.
While we have only examined three industries here， in reality， AI is flourishing across various sectors， including smart terminals， intelligent vehicles， industry， finance， and education. The foundational groundwork for this AI proliferation is being laid by cloud service providers， including Baidu AI Cloud.
Some readers might wonder how to choose among the many cloud service providers available. A recent IDC report indicates that the domestic market is primarily dominated by four major players: Baidu， Alibaba， Tencent， and Huawei.
The performance of Baidu AI Cloud， a leading provider， highlights the intense competition and innovation within this space. Cloud services in the AI era have moved beyond a rudimentary approach of simply accumulating GPUs; the true engineering challenge lies in enabling the efficient collaboration of thousands of GPUs.
Baidu was the first to champion the “AI Cloud” concept， and with its “Cloud-Intelligence Integration” strategy， each product upgrade and ecosystem synergy effectively answers the question: What kind of cloud services are needed in the AI era?
Today， their vision is clearer: the intelligent economy era requires a cloud that embodies “Cloud-Intelligence Integration with an Intelligent-First approach.”
From underlying computing power to model training and full-stack AI application development， Baidu AI Cloud liberates developers from the complexities of hardware and model deployment， allowing them to focus on refining exceptional AI applications.
Just as the Industrial Revolution was propelled by steel， gas， and electricity， the AI revolution requires its own essential utilities. Our collective imagination for AI is likely to be shaped by the evolution of cloud service providers.
