"The speed of innovation is more important than innovation itself." Elon Musk once said so.

Recently, Tesla, helmed by Musk, officially unveiled the second-generation Optimus humanoid robot at the 2024 World Artificial Intelligence Conference, just nine months after the first-generation model was introduced.

The rapidly upgraded humanoid robot lived up to expectations and became a major attraction at this year's WAIC (World Artificial Intelligence Conference). Besides this, what other hot topics did this year's WAIC spark? What forward-looking guidance did it bring?

Open Source and Closed Source are Complementary Relationships

This year, renowned domestic large models such as Huawei Pangu, Baidu Wenxin, Ali Tongyi, and Tencent Hunyuan all participated in WAIC, with large models continuing to be a major highlight.

During the opening of the conference, when discussing the choice between closed-source and open-source large models, two leading companies found themselves at odds. Baidu's founder, chairman, and CEO Robin Li said that open-sourcing models cannot achieve the same synergy as a collective effort, and that commercial closed-source models are the most powerful.

Alibaba Cloud CTO Zhou Jingren reiterated Alibaba Cloud's choice for open-source and emphasized that two years ago, Alibaba decided to open-source its Tongyi large model. Today, Tongyi Qianwen has achieved true full-scale, full-modal open-source, bridging the gap between open-source and closed-source models.

In comparison, there is no definitive conclusion as to whether open-source or closed-source large models are better.

From a technical perspective, closed-source models, due to their non-disclosure of source code, offer higher security and service quality, and can generate profits through license sales or model-based services. However, license fees are high, external review and regulation are difficult, and upgrades and iterations are highly dependent on internal teams, limiting their speed.

Open-source, on the other hand, has lower technical barriers and costs, attracting developers and researchers nationwide to participate, leading to faster innovation and iteration speeds and compatibility with more application areas. However, the open sharing of technology also brings risks of infringement, with quality, stability, and security difficult to guarantee.

In contrast, the only path for closed-source models to achieve "success" is to evolve into "super apps" that create value, while open-source models leverage their unique compatibility to create powerful customer acquisition tools. Thus, behind these seemingly different directions lie respective commercial interests.

From an application perspective, open-source and closed-source models are not like the iOS or Android systems on mobile devices, where one must choose one. Especially in ToB scenarios, application ends require both technology sharing and application security, as well as meeting diverse needs.

In this regard, Baichuan Intelligence CEO Wang Xiaochuan said that he expects 80% of enterprises to use open-source large models in the future, as closed-source models cannot better adapt to products or are particularly costly, leaving closed-source models to serve the remaining 20%.

Thus, the two are not mutually exclusive but can be complementary in different products and application scenarios. Ultimately, the core issue is not whether to choose open-source or closed-source when it comes to creating value with large models, as a basic model without applications is worthless.

Three Directions for AI Implementation: Large Models, Robots, and Terminal Products

As a component of the AI family, large models have always been a hot topic. With the conclusion of WAIC 2024, the latest trends in AI application implementation have become clear.

(1) Large Models Accelerate Commercialization

Following the "Hundred Models War," the "intensive cultivation" of large models has continued. Large model technologies represented by ChatGPT and Sora have entered a new round of iteration. At this year's WAIC, industry giants such as Baidu, Alibaba, Tencent, and Huawei brought numerous new technologies and products, demonstrating the application potential of large models in various industries such as finance, healthcare, and government affairs.

To date, Baidu has the Qianfan large model platform and Wenxin Yiyan, Alibaba has Alibaba Cloud Bailian and Tongyi large models, Tencent has Tencent Cloud Hunyuan and Yuanbao large models, and ByteDance has Huoshan Ark and Doubao large models, among others.

In 2023, large models saw significant enhancements in long-text processing, digital capabilities, reasoning abilities, RAG (retrieval-augmented generation), GPTs, multi-modality, native applications, and open-source. Simultaneously, the four major challenges of training and deployment costs, industry adaptation capabilities, hallucination issues, and data security have been optimized and resolved.

Image source: Shudian Technology

On this basis, customized large models have brought profound and radical changes to To B and To C businesses, with rigid demand gradually emerging among enterprise users and end-users. Based on this, To C products can continuously collect user feedback and accumulate model application practices to feed back into To B businesses, thereby accelerating the commercialization of large models.

(2) Robots and AI are Highly Integrated

This year, the humanoid robot zone was also a major highlight of the conference. In the central hall of the World Expo Exhibition Hall, an array of 18 humanoid robots showcased their "talents" and interacted warmly with the audience, highlighting the synergy of robot technology and significant breakthroughs in the field of "heterogeneous swarm intelligence".

With the rapid development of humanoid robots and AI, the industry has entered a stage of deep integration, and AI large models + humanoid robots are sparking the next wave of technological enthusiasm.

From a technical perspective, there is potential for integration in seven areas: natural language interaction, knowledge bases and reasoning, multi-modal perception and decision-making, motion planning, task planning and execution, emotional interaction, and continuous learning.

Breakthroughs have also been made in practical applications. For example, at WAIC 2024, Tesla's second-generation Optimus robot applied deep learning to visual perception, achieving precise target recognition and grasping, as well as upright walking and route planning.

In the field of natural language interaction, humanoid robots such as Xiaomi CyberOne, equipped with large language models, have achieved highly natural voice interaction. In the field of emotional interaction, Hanson Robotics' Sophia robot achieves rich emotional expression through facial expression synthesis and voice synthesis.

Through the integration of these technologies, we can create humanoid robots with multi-dimensional capabilities such as perception, decision-making, planning, control, interaction, and learning, ultimately endowing them with true intelligence and humanization.

(3) The Emergence of "New Species" of AI Terminals

In addition to large models and humanoid robots, AI terminal products are gradually infiltrating daily life. Just as in the three industrial revolutions, where steam engines emerged in the steam age, light bulbs were invented in the electrical age, and computers were created in the information age, each revolutionary "new species" has been applied to terminal scenarios, and the products of the AI Great Voyage era will be no exception.

Yang Yuanqing showcasing new AI PC products

This year, multiple industries have seen the emergence of "new terminal species." For example, new AI PC products from Dell, Lenovo, and Huawei, the Thunderbird AR glasses X2 Lite with built-in large model voice assistants, the Spacetime Pot Translator X1 for AI translation, and Apple's Vision Pro mixed reality headset, which garnered much attention in the first half of the year.

Tracing back to its roots, the emergence of terminal products is mainly due to the coordinated development of AI models, AI applications, and AI hardware. From the perspective of the AI industry structure, the upstream of the industrial chain is the foundation layer, including computing power, etc., the midstream is the algorithm and model layer, and the downstream is the application layer. First, the improvement and support of "hardware" such as computing power and chips, followed by the "intensive cultivation" of large models and algorithms, gave rise to terminal "new species".

With the continuous evolution of the AI ecosystem and technology, future AI terminals will also see comprehensive innovation and upgrades in architecture design, interaction methods, content, application ecosystems, and more. Perhaps, we will shift from asking "What should AI terminals look like?" to wondering "What will AI terminals look like?"

2024, Ushering in the "AI+" Era

In 2024, with the rapid development and maturity of technologies such as big data processing, high-performance computing, and deep learning, artificial intelligence has been able to solve a large number of complex problems. Coupled with increasingly mature large model technologies, the accuracy, efficiency, versatility, and flexibility of artificial intelligence have been significantly improved.

At the application level, as data generated by modern society begins to grow exponentially, providing "nourishment" for the evolution of AI, AI is able to better learn and understand the real world. Additionally, individual users and enterprise customers are increasingly demanding personalized and efficient services.

Driven by the rapid development of underlying technologies and market demand, artificial intelligence (AI) has transitioned from the "+AI" model to the "AI+" model.

In comparison, during the earlier "+AI" stage, AI was merely a supplementary technology applied to existing business processes and products in traditional industries, aiming to improve efficiency and solve specific problems.

In today's "AI+" stage, AI is no longer just an adjunct to business processes but a crucial driver of innovation and development across all industries. Its core lies in data-driven and self-learning capabilities, utilizing neural network models to train on vast amounts of data, simulating the learning mechanisms of the human brain, and achieving performance beyond humans in areas such as speech recognition and image recognition.

The completion of this stage transition signifies a shift from AI being a mere technical add-on tool to a key element leading industry transformation, upgrading from empowering a single function to reshaping entire business models.

The reshaping of business models, specific to application scenarios, covers areas such as transportation, lifestyle services, industrial manufacturing, cultural transmission, healthcare, rural construction, and more.

In the AI+ transportation field, with the support of AI algorithms, vehicles can process massive sensor data to achieve high-precision perception of their surroundings. They can also predict the trajectories of other vehicles, providing decision-making basis for autonomous vehicles. Most importantly, AI algorithms continuously learn and optimize, enhancing the performance and safety of autonomous driving systems.

In the AI+ manufacturing field, thanks to breakthroughs in sub-technologies such as large models, machine learning, and computer vision, AI can be applied to all aspects of the manufacturing process. By mining data from each individual link, it can empower overall prediction, production, management, and decision-making, achieving fine-grained management and helping enterprises reduce costs and increase efficiency.

Data from the Ministry of Industry and Information Technology shows that after intelligent transformation, the R&D cycle in manufacturing has been shortened by about 20.7%, production efficiency has increased by about 34.8%, defective product rates have decreased by about 27.4%, and carbon emissions have been reduced by about 21.2%.

In the long run, the potential of "AI+" is already starting to emerge.

Conclusion

In 2024, the imagination for AI application implementation is exploding. Large models are accelerating commercialization, AI and humanoid robots are deeply integrated, and new terminal species are emerging in large numbers, with the historical gears of the Great Voyage era beginning to spin at high speed.

Not only the AI industry but the entire industrial chain is also expanding rapidly, with AI+ transportation, AI+ automobiles, AI+ culture, AI+ manufacturing, and more moving from theory to practice, entering production and daily life.

The "AI+" era has truly arrived.

Source: Songguo Finance