Titanium Media AGI has learned that on the morning of January 8, the Beijing Academy of Artificial Intelligence (hereinafter referred to as "BAAI") released the "Top Ten Artificial Intelligence Technologies and Application Trends" report, covering new AI technology trends such as "embodied intelligence", world model, and synthetic data, so as to analyze the trajectory of technological evolution.
Wang Zhongyuan, the dean of the Beijing Academy of Artificial Intelligence, said that we are currently at a new inflection point in the development of AI. The emergence of capabilities in large - scale models is accelerating the arrival of the era of artificial general intelligence (AGI). Native unified multimodality, embodied intelligence, and AI for Science will further deepen AI's perception, understanding, and reasoning of the world, connect the digital world with the physical world, and drive innovative breakthroughs in scientific research. As a new - type research and development institution focusing on the AU field, the Beijing Academy of Artificial Intelligence hopes that, at this special moment, taking the ten trends as a starting point, it can point out the development direction for the AI technology field and move forward hand in hand.
Lin Yonghua, the vice - dean and chief engineer of the Beijing Academy of Artificial Intelligence, said at the meeting that everyone is looking forward to AI surpassing human intelligence, achieving artificial general intelligence (AGI), moving from the digital world to the physical world, and even helping us explore unknown fields (worlds) in the future. However, with the gradual advancement of the AGI goal, there may be various paths and methods (a blooming of diverse approaches) in the process of achieving this goal. As for which path can lead to the end - point and how far we still need to go to truly achieve AGI, there is currently no conclusive answer to these questions.

Specifically, among the top ten AI technology trends in 2025 announced by the Beijing Academy of Artificial Intelligence, the first trend is that AI for Science (AI4S) drives the transformation of the scientific research paradigm. According to statistics, in 2024, the proportion of researchers using AI increased rapidly. Nearly half of the researchers believe that AI will have a positive impact on their fields of work, while in the United States and India, the proportions are only 28% and 41% respectively. This means that the transformative effect of AI on scientific research methods and processes is beginning to emerge.
With the Nobel Prizes in Physics and Chemistry being awarded in the AI field, it has promoted the continuous integration of scientific research and AI technology, evolving from focusing on optimizing specific tasks to addressing more complex, dynamic, and interdisciplinary problems. In 2025, multimodal large - scale models will be further integrated into scientific research. They will enable the mining of complex structures in multi - dimensional data, assist in the comprehensive understanding and global analysis of scientific research problems, and open up new directions for basic and applied scientific research in fields such as biomedicine, meteorology, materials discovery, life simulation, and energy.
Trend 2: "The Year of Embodied Intelligence", the co - evolution of the embodied "cerebrum and cerebellum" and the body. In 2025, the narrative of "embodied intelligence" will continue to expand from the body to the embodied brain, and we can expect more in three aspects. In terms of the industry landscape, nearly 100 domestic embodied intelligence start - ups may face a reshuffle, and the number of manufacturers will begin to decline. Regarding the technological route, end - to - end models will continue to iterate, and attempts at "cerebellum large models" may achieve breakthroughs. In terms of commercial realization, we will surely see more applications of embodied intelligence in industrial scenarios, and some humanoid robots will enter mass production.
Trend 3: "The Next Token Prediction": Unified Multimodal Large - scale Models Enable More Efficient AI. The essence of artificial intelligence lies in simulating the information - processing process of human thinking. Human interaction with and processing of information always exhibit a multi - modal and cross - modal input - output state. Current language large - scale models and patchwork multi - modal large - scale models have inherent limitations in simulating the human thinking process. The native multi - modal technology route that breaks through multi - modal data from the beginning of training and enables end - to - end input and output offers new possibilities for the development of multi - modality. Based on this, aligning data from modalities such as vision, audio, and 3D during the training phase to achieve multi - modal unity and constructing native multi - modal large - scale models has become an important direction for the evolution of multi - modal large - scale models.
Trend 4: Expansion of Scaling Law: RL (Reinforcement Learning) + LLMs. Model generalization shifts from pre - training to post - training and inference transfer. The "cost - effectiveness" of the training mode that boosts the performance of basic models based on the Scaling Law continues to decline, and Scaling laws for post - training and specific scenarios are constantly being explored. As a key technology for uncovering Scaling Laws in the post - training and inference stages, reinforcement learning will also see more applications and innovative uses.
Trend 5: The world model is expected to be released at an accelerated pace and is likely to become the next stage of multimodal large - scale models. It is reported that the world model places more emphasis on the role of "causal" reasoning, endowing AI with a higher level of cognition and more logical reasoning and decision - making abilities. This ability can not only promote the in - depth application of AI in cutting - edge fields such as autonomous driving, robot control, and intelligent manufacturing, but also holds the promise of breaking through traditional task boundaries and exploring new possibilities for human - machine interaction.
Trend 6: Synthetic data will become an important catalyst for the iteration of large - scale models and the implementation of applications. High - quality data will become an obstacle to the further Scaling up of large - scale models. Synthetic data has become the first choice for basic model manufacturers to supplement data. Synthetic data can reduce the costs of manual governance and annotation, ease the dependence on real data, and eliminate data privacy concerns. It can also enhance data diversity, which helps improve the model's ability to handle long texts and complex problems. In addition, synthetic data can mitigate issues such as the monopolization of general data by large companies and the high cost of obtaining proprietary data, thus promoting the practical application of large - scale models.
Trend 7: The acceleration of inference optimization iteration has become a necessary condition for the implementation of AI - Native applications. The hardware carriers of large - scale models are infiltrating from the cloud to end - side hardware such as mobile phones and PCs. On these devices with limited resources (such as AI computing power and memory), the implementation of large - scale models will face significant overhead limitations on the inference side, posing huge challenges to deployment resources, user experience, and economic costs. The continuous iteration of algorithm acceleration and hardware optimization technologies will drive the implementation of AI - Native applications in a two - pronged manner.
During the round - table discussion, Wen Zujie, the person in charge of large - model alignment at Ant Group, said that at the OpenAI launch event, a real - life version of "Her" was presented. You can interact with it in real - time. It uses large - scale models to observe your actions and the surrounding environment, enabling a more natural interaction. Multimodality doesn't just refer to video generation; it also includes text - image multimodality, OCR multimodality, and other capabilities. For example, Ant Group's "Explore" can not only use visual capabilities to take pictures and recognize objects, but also achieve multi - round conversational interactions based on multimodal capabilities. This provides a product perception that is more in line with real - life experiences, and these areas may hold great promise.
Trend 8: Reshape the form of product applications, with Agentic AI becoming an important model for product implementation. In 2025, more general - purpose and autonomous agents will reshape the form of product applications, further penetrating into work and life scenarios, and becoming an important application form for the implementation of large - model products. In 2025, we will witness the implementation of more multi - agent systems with higher intelligence levels and a deeper understanding of business processes on the application side.
Ni Xianhao, the person in charge of the industry research group of the Beijing Academy of Artificial Intelligence, said that from Chatbot to Copilot, and then to Agent and Agentic AI, the industry has a deeper and deeper understanding of the application forms of AI. Especially from Agent to Agentic AI, it marks a more practical shift from judging whether a product belongs to an Agent to exploring the intelligence level of the product. In the coming year, we may not necessarily see many distinct changes in application forms, nor will there be many completely different Agent application models.
Trend 9: The popularity of AI applications is gradually rising, and it's still uncertain which one will become the Super App. In the past year, the processing capabilities of generative AI models in the fields of images and videos have been significantly enhanced. Coupled with the cost reduction brought about by inference optimization, and the continuous development of technologies such as the Agent/RAG framework and application orchestration tools, the foundation has been laid for the implementation of AI super - applications. The applications of large - scale models have been upgraded from functional points, penetrating into the construction of AI - native applications and the ecological reshaping of AI operating systems (AI OS). Although it remains undetermined which application will ultimately become the Super APP, from the perspectives of user scale, interaction frequency, and duration of stay, the popularity of AI applications is on the rise, and we are on the verge of an application explosion.
Ni Xianhao said that there is currently a certain opportunity to develop a "super - application". Although the growth rate of mobile Internet users has reached its peak, among non - Internet users, we can still identify a user group of over 100 million. These users have spending power but, due to generational factors, cannot access the Internet easily. They face pressing needs such as non - cash payment, information acquisition, and online ticket - booking and registration. To address these issues, we don't necessarily have to pursue the utmost in the intelligence and autonomy of the agent. Based on a good basic model, combined with the adaptation of manufacturer interfaces (plugins, tools) corresponding to the different capabilities mentioned above, it is feasible in terms of both model and engineering capabilities to create an agent that can meet the needs of this user group.
Ni Xianhao believes that under this logic, how to connect with different manufacturers to complete interface adaptation and encapsulation has instead become an equally important issue. This kind of interface adaptation corresponds to the requirements for manufacturers' channel construction and operation capabilities. "Therefore, in the narrative of super - applications characterized by 'All in One', large companies may have a greater chance. The above - mentioned channel construction capabilities are relatively mature for large companies. However, for start - ups, these tasks need to be built from scratch, which is extremely difficult," said Ni Xianhao.
Trend 10: Equal emphasis on enhancing model capabilities and preventing risks, with the continuous improvement of the AI security governance system. As complex systems, the Scaling of large - scale models has led to emergence. However, the unique attributes of complex systems, such as unpredictable emergent results and circular feedback, also pose challenges to the security protection mechanisms of traditional engineering. The continuous progress of basic models in autonomous decision - making brings potential risks of losing control. How to introduce new technical supervision methods and how to balance industry development and risk control in manual supervision? These are issues worthy of continuous discussion for all parties involved in AI.
Wen Zujie said that AI security has a significant "adversarial" characteristic, which is a seesaw - like relationship. That is, when the means of attack are strengthened, the means of defense will also be enhanced accordingly. Therefore, in terms of the security capabilities of large - scale models, we need to continuously improve the strength of both the offensive and defensive ends. Adopting the approach of "using large - scale models to counter large - scale models" is a positive development trend. In addition, regarding security fencing technology, we must ensure the security of both input and output to reduce the risk of being attacked. By establishing a complete set of policy systems and security protection combinations, we can promote the secure application of AI large - scale models.
In fact, as a crucial engine of new - quality productive forces, AI not only represents the cutting - edge trend of science and technology but also serves as a key driving force for future economic development. At present, it has already yielded significant economic and social benefits.
The latest data from research firm IDC shows that as AI applications continue to penetrate deeper and become more practical, industry - specific large - scale models have achieved initial applications in multiple industry sectors such as finance, healthcare, education, retail, and energy. By 2025, global AI spending will reach $227 billion. It is estimated that by 2030, AI will contribute $19.9 trillion (approximately 145.9 trillion yuan) to the global economy, driving a 3.5% increase in global GDP. At present, nearly 98% of corporate leaders consider AI a priority for their organizations. Regarding future prospects, many industry experts have expressed their hope that in 2025, they can witness the birth of the next - generation large - scale model like GPT - 5, and see significant progress in the security and theoretical interpretability of large - scale models. "I don't know if this is too good to be true, but I hope there will be an AI with a learning efficiency similar to that of humans."