Note: The article was first published on TechNode China written by Evan Huang and translated by Zinan Zhang.

In the dynamic AI era, venture capital is increasingly attuned to the transformative potential of this technology. As generative AI advances in creating text, images, and videos, a plethora of opportunities and challenges are emerging. This article explores the pivotal role of the Scaling Law, the emergence of super apps, and the promising future of AI-driven innovations. Highlighting insights from industry leaders, it underscores the potential for AI to revolutionize various sectors and entrepreneurial ventures, providing valuable directions for future venture capital investments.

The utility of the Scaling Law

The training and inference stages of large models demand substantial computational resources. The Scaling Law suggests that significant advancements in intelligence are achieved through consistent investment in vast amounts of data and powerful computing, provided the algorithmic architecture remains stable.

OpenAI, a strong proponent of the Scaling Law, has showcased the potential of generative AI across various fields by leveraging transformer architecture, extensive training data, and considerable computational resources.

Recently, Kevin Scott, Microsoft CTO, mentioned in an interview with Pat Grady and Bill Coughran of Sequoia Capital that they have yet to observe diminishing returns from scaling. He announced that the next generation of OpenAI models would soon be available, offering cheaper, more powerful solutions capable of tackling more complex problems. “This is the story with each generation of models as we scale up,” he remarked.

On May 18, Yang Zhilin, founder of Moonshot AI, discussed the computational aspects of the Scaling Law. He noted that initial improvements in model performance are driven by enhanced computational power and efficiency. However, further advancements require increased computational investment and ensuring that this investment effectively translates into intelligence. “This involves two issues: sustaining computational investment and maximizing the intelligence output of each computation unit,” he explained.

In an interview with TechNode, Wu Yunsheng, vice-president of Tencent Cloud, shared his perspective. “Currently, there are different viewpoints, including realistic and idealistic views. Some believe the Scaling Law has reached a plateau, where continued investment yields diminishing returns. Others argue it is still in a phase of rapid development.” He emphasized that the Scaling Law remains significant, citing rapid progress in multimodal research over the past year. “In this field, various capabilities improve significantly with added data or computing power. We will continue to explore and observe its development and changes across different scenarios and technologies,” he added.

The super app is on the way

As of March 28, 2024, there are 117 large models registered with the Cyberspace Administration of China, including Baidu’s ERNIE Bot, Alibaba’s Tongyi Qianwen, and the open-source ChatGLM. The rapid development of AI large models is becoming a key driver of innovation and breakthroughs in super applications. 

As these large model technologies mature and improve, they are gradually permeating various industries, sparking a range of entrepreneurial opportunities. From healthcare to fintech, from smart manufacturing to cultural creativity, the application potential of AI is limitless. 

Zhou Zhifeng, Managing Partner of Qiming Venture Partners, pointed out at the  World Artificial Intelligence Conference in Shanghai that compared to the timeline of application deployment during the internet wave, he predicts that the explosion of applications in the current AI wave will occur significantly earlier. Currently, generative AI is gaining substantial user favor in three “C fields” — Copilot, Creativity, and Companionship — showing a development trajectory similar to internet applications and transitioning from efficiency-enhancing applications to those aimed at providing enjoyment. He noted that the internet reduced the marginal cost of information distribution to almost zero, while the core of generative AI is to reduce the marginal cost of digital content creation to nearly zero, indicating that AI technology is bound to release enormous value.

When discussing the future of AI-driven super apps, Zhang Fan, COO of Zhipu AI, expressed optimism, arguing that although creating super apps is not easy, the AI era will see many unimaginable applications emerge. This process requires advancements in computing power, networks, hardware levels, and user habits, following the principle of gradual development from small-scale applications. Zhang emphasizes that by embracing and utilizing existing AI technologies to gradually transform current applications and products, the future will undoubtedly usher in super apps in the AI era.

Regarding the challenges of implementing generative AI applications, Zhou Zhifeng believes that reducing the cost of model usage necessary for the widespread adoption of generative AI, improving the effectiveness of large models, and enhancing user retention rates of generative AI applications are crucial. Since the growth period from zero to one for generative AI application companies is longer than in other fields, they need to overcome both TPF (Technology-Product Fit) and PMF (Product-Market Fit) challenges simultaneously. Therefore, the founding team needs greater patience, determination, and understanding of the technology, the product, and the world.

Embodied intelligence, infinite imagination

There were 45 intelligent robots, including 25 humanoid robots, showcased at WAIC this year. A video of a humanoid robot walking on the Great Wall was repeatedly played at the event. The humanoid robot L2 in the video has successfully conquered the steep slopes of the famous structure, achieving steady walking on it.

At the recent Huawei Developer Conference 2024, Zhang Ping’an, Executive Director and CEO of Huawei Cloud, unveiled the Pangu Model 5.0. During the introduction of the Pangu model for embodied AI, he showcased the broad potential of the KUAVO humanoid robot, equipped with the Pangu model, in both industrial and household scenarios, attracting widespread attention.

Chen Jianyu, an assistant professor at Tsinghua University and founder of the humanoid robot company Robot Era, believes that humanoid robots will be the ultimate form of general-purpose robots. This is not only because the pure humanoid form with two legs and two arms is more compatible with existing environments, but also because it’s easier to transfer training data from the human world. Technically, an end-to-end integration of the brain and cerebellum will be a crucial research direction in the future. Using human language as the interface between the brain and cerebellum is limited, and it is better to borrow from the end-to-end joint training process of autonomous driving, where physical layer data is directly fed back to the text and image models, significantly enhancing overall model performance.

Last week, Tencent, in collaboration with Shanghai Jiao Tong University, released the Top Ten Trends of Large Models 2024: Entering the Era of ‘Machine External Brain’ report, which pointed out that the combination of robot technology and large models provides a “body” for the machine’s external brain. In the future, humanoid robots will not only be able to perform physical tasks but also interact with humans more naturally and intuitively, endowing physical products with intelligent “brains”.

The report states that the development of humanoid robots relies on two major technical pillars: motion control and task training. The application of large models has greatly improved the robots’ learning efficiency and ability to execute complex tasks. The integration of these technologies not only drives technological innovation in humanoid robots but also opens possibilities for their widespread deployment in practical applications. This also heralds a future of human-machine symbiosis, where humanoid robots will play increasingly important roles in various industries, from household services to high-risk industrial operations, showcasing their efficiency and safety. Through continuous technological innovation and application expansion, humanoid robots will play a key role in improving the quality of life and work efficiency, further integrating into human daily life as indispensable assistants and the ultimate carriers of artificial intelligence.

Conclusion

In conclusion, the era of AI is not just a technological revolution but a transformative force that is redefining the landscape of innovation and investment. As we look to the future, the challenges of implementing generative AI applications remain significant.      The need to reduce costs, improve effectiveness, and enhance user retention rates is crucial for the widespread adoption of these technologies. However, the potential rewards are immense, offering a glimpse into a world where AI is not just a tool but an integral part of our daily lives, from household services to high-risk industrial operations.

In summary, the dynamic AI era presents a wealth of opportunities for venture capital and entrepreneurial ventures. As we continue to explore and invest in AI-driven innovations, the future holds huge promise for transforming industries, enhancing human-machine interactions, and ultimately, improving the quality of life for all.

Autotech startup Nullmax said on Tuesday that its latest generation of autonomous driving hardware and software package, allowing cars to navigate complex urban environments autonomously with features such as lane changing, will cost users as little as “several thousand RMB.” 

Why it matters: Shanghai and Fremont-based Nullmax is among the few players in the self-driving vehicle space claiming that cars will be able to function by themselves in urban scenarios without maps and lidar. Instead, the company said artificial intelligence models can be used to enable cars to navigate from points A to B.

• This makes its system even more cost-competitive on the Chinese market compared with Tesla’s Full Self-Driving (FSD) software, chief executive Xu Lei told reporters in Shanghai. Tesla has reportedly partnered with Baidu to leverage the latter’s lane-level navigation and standard definition mapping services, as part of its push to localize its most advanced driver-assistance software (ADAS) in the country.

Details: Xu told a press conference that his company is advocating a “pure vision” and “end-to-end” approach, as Tesla has been doing and many are following its lead, which involves deep neural networks, using cameras only to perform autonomous driving functions (our translation).

• The car could complete the highway on-ramp to off-ramp maneuver and drive through a construction zone in Chinese urban areas with affordable hardware of 7-11 cameras and a computing platform that can handle roughly 100 trillion operations per second (or TOPS), according to Nullmax.
• Xu expects some car models equipped with the technology to come to market in 2025, without saying more. Nullmax, with roughly 300 scientists and engineers, has been developing more safety-based ADAS functions for domestic automakers SAIC, and Chery, among others, TechNode has learned.
• Cars powered by Nullmax’s tech have traveled a combined 10 million kilometers (6.2 million miles). Xu added more actual driving data is required to provide cars with point-to-point navigation on city streets, while the company is training the ADAS system via simulation with AI-generated data.

Context: Chinese EV startups led by NIO, Xpeng Motors, and Li Auto have been ramping up efforts to transition from “rule-based” designs to an “end-to-end” autonomous driving method. Meanwhile, traditional car manufacturers are tapping into the power of AI by working with tech giants such as Huawei and NVIDIA, as well as startup unicorns like Horizon Robotics and Momenta.

• “What we see in the future is that the (AI-defined) AV stack will become an end-to-end model and will be trained in the cloud with massive data. More importantly, it will be validated in the cloud with simulation capabilities as well,” said Wu Xinzhou, NVIDIA’s vice president of automotive at its annual developer conference GTC in March.
• The US chipmaker in late 2021 released a system called “Omniverse Replicator” to facilitate the training of autonomous vehicles in the virtual world, Reuters reported. Chinese major EV makers, including BYD and NIO, are building their ADAS upon its DRIVE Orin computers, each of which offers 254 TOPS of performance.
• Nullmax was co-founded in 2016 by Xu Lei and Justin Song, both of whom spent time at Tesla. Xu worked on Tesla’s Autopilot computer vision team in 2015 and 2016, after leaving a senior engineering role at Qualcomm, while Song was responsible for engineering the Tesla Autopilot and car infotainment system from 2012 to 2015, according to their LinkedIn profiles.

READ MORE: Former Tesla engineer shares thoughts on end-to-end autonomous driving at WAIC 2024

Editor’s note: ‘Landing AI’ is a series of special reports focusing on the field of Artificial Intelligence curated by TechNode. By investigating the development of AI landing in China and the behind-the-scenes stories of the industry, we’re going to dive deeper into everything that’s possible under the new wave of AI.

On July 12, the Bilibili World 2024 event opened at the National Exhibition and Convention Center, attracting ACG (Anime, Comic, Game) fans from across China who traveled to Shanghai to become part of the Chinese streaming platform’s annual celebration. 

The three-day event this year featured over 700 exhibitors and 800 ACG-content creators. The Lenovo booth showcased two new AI products: the YOGA Air 14c AI PC and the Legion Y9000P AI PC.

Why it matters: Bilibili World 2024 draws in ACG fans to celebrate shared interests, which meant the AI-powered Lenovo PCs garnered significant attention.

Details: The Lenovo YOGA Air 14c AI PC is designed for work, while the Lenovo Legion Y9000P AI PC is principally marketed to gamers. 

• “In the current wave of AI development, computers are gradually evolving into personal assistant roles. As a leading enterprise in the PC industry, Lenovo will continue to maintain its leadership in this new era,” said Li Weichang, Vice President and General Manager of Lenovo China’s Consumer PC and Tablet Business, at the booth during the event.
• Li said an advanced AI PC should have five features: an embedded personal intelligent agent for natural interaction, a built-in knowledge base, combined CPU+GPU+NPU AI processing power, an open AI application ecosystem, and protection for personal privacy and data security.
• The Lenovo YOGA Air 14c AI laptop is equipped with an Intel Core Ultra 7 155H processor, featuring an advanced 6+8+2 core thread configuration, with a maximum frequency of up to 4.8GHz. The AI model offers up to 32GB RAM and 1TB storage.
• The YOGA Air 14c AI PC integrates an intelligent agent, Lenovo Xiaotian, powered by the firm’s self-developed Tianxi large language model. Lenovo Xiaotian is able to learn user habits and preferences at work and while engaged in entertainment. The AI assistant is then able to offer practical functionalities including AI-powered PowerPoint production, document summarizing, voice cloning, and painting.

• The flagship Y9000P AI features Intel’s i9-14900HX processor and Nvidia’s RTX 4090 graphics card, supporting direct GPU connection technology with a maximum power consumption reaching up to 250W. In terms of display, the flagship model provides a 2,560 x 1,600 resolution 240Hz screen for a detailed gaming experience.
• The Legion Y9000P AI laptop is also equipped with Lenovo Xiaotian, which interacts with users in various scenarios including document summarizing, research question answering, and AI performance tuning.

Context: Held since 2017, Bilibili World is a community-focused annual large-scale in-person event organized by Bilibili, a leading Chinese video-sharing website. On June 29, the first round of the ACG event’s ticket sales saw 27,000 VIP tickets sell out within 30 seconds, with 100,000 general admission tickets gone within one minute, according to the company’s own app.

• Intel’s China District Technical Director, Gao Yu, attended the event and introduced the Intel Core Ultra processor, integrating three types of computational engines — CPU, GPU, and NPU (collectively known as XPU) — for various AI tasks. Yu hinted at an upcoming next-generation AI PC processor based on the Lunar Lake architecture, promising even more powerful AI functionalities.

Kuaishou, one of the main rivals to TikTok’s China sibling Douyin, showcased several fresh features for its text-to-video model Kling AI at the World Artificial Intelligence Conference (WAIC) in Shanghai last week, including the ability to generate videos up to 10 seconds. 

At WAIC, visitors queued to experience the Sora-like tool that is currently available by invitation only. Users sent simple prompts to generate videos, such as “a panda eating salmon” and “the Mona Lisa putting her glasses on with her hands”, with the resulting clips demonstrating Kling AI’s ability to render the inputs almost perfectly.

AI-generated videos have subsequently flooded the Chinese internet, with Kling AI being used to create clips featuring characters from historical films undertaking modern day tasks and spawning multiple memes.  

A video featuring Rong momo, a character from “My Fair Princess” who has become a well-known internet meme in China, feeding Princess Ziwei a chicken drumstick has gone viral on social platforms these days. The AI-generated video is based on the drama’s most famous scene in which Rong momo tortures Ziwei by repeatedly stabbing her with a needle.

Why it matters: Kuaishou will be hoping that its suite of self-developed large model series, including language model KwaiYii, image-focused Kolors, and video-centered Kling, will give it an edge as it continues to challenge to ByteDance’s Douyin and TikTok.

Details: More than 500,000 users have applied to help beta test Kling, senior vice president of Kuaishou Gai Kun revealed last weekend at a WAIC forum, with the number of videos generated reaching 7 million as of now. The Sora rival’s hype is such that English language posts teaching users outside of China how to apply for a Kling AI trial can be found on X, formerly known as Twitter.

• Kuaishou provided practical tips on the screen at the WAIC event, including advising users to use simple words and sentence structures and to avoid overly complex language. It also emphasized that its model is not sensitive to numbers, giving an example that if the prompt is “10 puppies on the beach,” the number might not be consistently maintained in its outputs.
• A member of staff from Kuaishou’s large language model team told TechNode that they were not at liberty to disclose the data used to train Kling AI, but indicated that it was open source.
• The TikTok rival meanwhile announced at WAIC that its Midjourney-like model Kolors would become open source, a move which Kuaishou said aims to contribute to a more prosperous ecosystem for the text-to-image generation community.
• Kuaishou’s investment in research and development has quadrupled over four years, with expenditure increasing from RMB 2.9 billion in 2019 to RMB 12.3 billion in 2023.

Context: Kuaishou, China’s second-largest short video company, launched its AI strategy in 2023, according to CEO Cheng Yixiao, who said that generative AI has a “very rich combination of business scenarios and huge value potential” for the content platform.

Editor’s note: ‘Landing AI’ is a series of special reports focusing on the field of Artificial Intelligence curated by TechNode. By investigating the development of AI landing in China and the behind-the-scenes stories of the industry, we’re going to dive deeper into everything that’s possible under the new wave of AI.