At the beginning of 2025, BAAI proposes ten artificial intelligence technology and application trends to analyze the trajectory of technological evolution, gain insight into future development directions, predict key driving forces, and guide human society toward a more intelligent, better, and interconnected future.
Trend 1 The Future of Science: AI4S Drives a Paradigm Shift in Scientific Research
AI4S (AI for Science) under the leadership of large models has become a key force driving a paradigm shift in scientific research. In 2024, the proportion of researchers using AI increased rapidly, and AI's transformative effect on scientific research methods and processes began to emerge.
Whether improving scientific research efficiency or advancing scientific research from focusing on optimizing specific tasks toward more complex, more dynamic, and more interdisciplinary problems, AI4S empowered by large models has shown effects completely different from the past. In 2025, multimodal large models will be further integrated into scientific research, empowering complex structure mining of multidimensional data, assisting comprehensive understanding and global analysis of scientific research problems, and opening up new directions for research in basic and applied sciences such as biomedicine, meteorology, materials discovery, life simulation, and energy.
Trend 2 "The First Year of Embodied Intelligence": The Co-evolution of Embodied Brains and Bodies
In 2025,embodied intelligencewill continue the narrative thread of expanding from bodies to embodied brains, and we can expect more in three aspects. In terms of industry landscape, nearly a hundred embodied startups may face a reshuffle, and the number of vendors will begin to converge; in terms of technical routes, end-to-end models will continue to iterate, and attempts at cerebellum large models may see breakthroughs; in terms of commercial monetization, we will also certainly see more embodied intelligence applications in industrial scenarios, and some humanoid robots will enter mass production.
Trend 3 "The Next Token Prediction": Unified Multimodal Large Models Achieve More Efficient AI
The essence of artificial intelligence lies in simulating the information processes of human thought. Humans' interaction with and processing of information always presents multimodal and cross-modal input and output states. Current language large models and concatenated multimodal large models have inherent limitations in simulating human thought processes. The native multimodal technical route, which connects multimodal data from the very beginning of training and achieves end-to-end input and output, offers new possibilities for multimodal development. Based on this, aligning data from modalities such as vision, audio, and 3D during the training stage to achieve multimodal unification and building native multimodal large models has become an important direction for the evolution of multimodal large models.
Trend 4 Scaling Law Expansion: RL + LLMs, Model Generalization Shifts from Pre-training to Post-training and Inference
Based onScaling Lawdrivingfoundation modelsthe cost-effectiveness of training paradigms that improve performance continues to decline, and scaling laws for post-training and specific scenarios are constantly being explored. Reinforcement learning, as a key technology for discovering scaling laws in the post-training and inference stages, will also receive more applications and innovative uses.
Trend 5 World Models Accelerate Their Release and Are Expected to Become the Next Stage of Multimodal Large Models
World models that place greater emphasis on "causal" reasoning give AI a higher level of cognition and more logically consistentreasoningand decision-making capabilities. This capability can not only promote deep applications of AI in frontier fields such as autonomous driving, robot control, and intelligent manufacturing, but is also expected to break through traditional task boundaries and explore new possibilities for human-computer interaction.
Trend 6 Synthetic Data Will Become an Important Catalyst for Large Model Iteration and Application Deployment
High-quality data will become a development obstacle to further scaling up large models.Synthetic dataIt has become the first choice for foundation model vendors to supplement data. Synthetic data can reduce the cost of manual governance and annotation, alleviate dependence on real data, and no longer involve data privacy issues; it improves data diversity and helps enhance the model's ability to handle long texts and complex problems. In addition, synthetic data can alleviate issues such as general data being monopolized by big tech companies and the high cost of acquiring proprietary data, promoting the application and deployment of large models.
Trend 7 Inference optimization iterations accelerate, becoming a necessary condition for the deployment of AI Native applications
Large model hardware carriers are penetrating from the cloud to end-side hardware such as mobile phones and PCs. On these resource-constrained devices (AI computing power, memory, etc.), the deployment and application of large models will face significant inference-side overhead constraints, posing huge challenges to deployment resources, user experience, and economic costs. Algorithm acceleration and hardware optimization technologies continue to iterate, and the two-wheel drive accelerates the deployment of AI Native applications.
Trend 8 Reshaping product application forms, Agentic AI becomes an important model for product deployment
In 2025, more general and more autonomous agents will reshape product application forms, further penetrate work and life scenarios, and become an important application form for the deployment of large model products. FromChatbot, Copilot to AI Agent, Agentic AI, since 2023 the industry's understanding of AI application forms has deepened. At the end of 2023, the AI Agent application orchestration framework converged, and the concept of Agentic AI was proposed. From Agent, which emphasizes the product concept more, toAgentic AI, which emphasizes the degree of application intelligence more, in 2025 we will see more multi-agent systems with higher intelligence and deeper understanding of business processes deployed on the application side.
Trend 9 AI application popularity is gradually rising, and it is still unknown who will win the Super App
Over the past year, the processing capabilities of generative models in images and video have been greatly improved. Combined with the cost reduction brought by inference optimization, and the continuous development of technologies such as Agent/RAG frameworks and application orchestration tools, the foundation has been laid for the deployment of AI super applications. Large model applications have upgraded from feature points to penetrating into AI-native application construction and the ecological reshaping of AI OS. Although it is not yet settled who will win the Super APP, from dimensions such as user scale, interaction frequency, and dwell time, the popularity of AI applications continues to climb, and it has reached the eve of an application explosion.
Trend 10 Equal emphasis on model capability improvement and risk prevention, and the AI security governance system continues to improve
As a complex system, the Scaling of large models brings emergence, but the unique properties of complex systems, such as unpredictable emergent results and cyclic feedback, also pose challenges to the security protection mechanisms of traditional engineering. The continuous progress of foundation models in autonomous decision-making brings potential risks of losing control. How to introduce new technical regulatory methods, and how to balance industry development and risk control in manual supervision? This is an issue worthy of continuous discussion for all parties involved in AI.
The dawn of technology will shine on the path of human progress. These technologies will promote each other amid fierce competition and collaboration, jointly composing a majestic chapter of symbiosis and shared prosperity between humans and intelligent systems. As witnesses, we will see technology inject surging momentum into human civilization and push the boundaries of human capability further and higher. Standing on the shoulders of technology, the body can be a hundred feet tall, and the hands can pluck the stars.