<p>Author | Liu Baodan</p>
<p>Editor | Huang Yu</p>
<p>After two years of rapid development, AI large models are about to enter a critical third year. The Zhiyuan Institute has made predictions about the direction of AI large models in 2025.</p>
<p>On January 8, the Beijing Zhiyuan Artificial Intelligence Research Institute released the &#34;Top 10 AI Technology Trends for 2025,&#34; predicting key directions such as Scaling Law, foundational models, embodied intelligence, super applications, and AI safety, from infrastructure to product applications.</p>
<p>&#34;Currently, we are at a new turning point in the development of artificial intelligence, and the emergence of capabilities in large models is accelerating the arrival of the era of general artificial intelligence,&#34; said Wang Zhongyuan, president of the Zhiyuan Institute. Native unified multimodal, embodied intelligence, and AI for Science will further deepen artificial intelligence's perception, understanding, and reasoning of the world, connecting the digital world with the physical world and driving innovative breakthroughs in scientific research.</p>
<h2>Trend 1: The Future of Science: AI4S Driving Paradigm Shift in Scientific Research</h2>
<p>AI4S (AI for Science), led by large models, has become a key force in driving paradigm shifts in scientific research. In 2024, the proportion of researchers using AI will rapidly increase, and the transformative effects of AI on scientific research methods and processes will begin to emerge. By 2025, multimodal large models will further integrate into scientific research, empowering the complex structure mining of multidimensional data, assisting in the comprehensive understanding and global analysis of research questions, and opening new directions for research in fundamental and applied sciences such as biomedicine, meteorology, materials discovery, life simulation, and energy.</p>
<h2>Trend 2: &#34;The Year of Embodied Intelligence&#34;: Co-evolution of Embodied Brain and Ontology</h2>
<p>In 2025, embodied intelligence will continue to expand from ontology to the narrative mainline of embodied brains, and we can expect more in three aspects. In terms of industry landscape, nearly a hundred embodied startups may face reshuffling, and the number of manufacturers will begin to converge; in terms of technological routes, end-to-end models will continue to iterate, and attempts at small brain large models may achieve breakthroughs; in terms of commercial realization, we will also see more applications of embodied intelligence in industrial scenarios, with some humanoid robots entering mass production.</p>
<h2>Trend 3: &#34;The Next Token Prediction&#34;: Unified Multimodal Large Models Achieving More Efficient AI</h2>
<p>The essence of artificial intelligence lies in simulating the information processes of human thinking. Current language large models and spliced multimodal large models have inherent limitations in simulating human thinking processes. A native multimodal technical route that connects multimodal data from the beginning of training, achieving end-to-end input and output, presents new possibilities for multimodal development. Based on this, aligning data from visual, audio, 3D, and other modalities during the training phase to achieve multimodal unification and constructing native multimodal large models becomes an important direction for the evolution of multimodal large models.</p>
<h2>Trend 4: Scaling Law Expansion: RL + LLMs, Model Generalization from Pre-training to Post-training and Inference Transfer</h2>
<p>The training mode driven by Scaling Law to enhance the performance of foundational models has seen a continuous decline in &#34;cost-effectiveness.&#34; Post-training and specific scenario Scaling Laws are being explored. Reinforcement learning, as a key technology for discovering Scaling Laws in post-training and inference stages, will also see more applications and innovative uses</p>
<h2>Trend Five: The Accelerated Release of World Models, Expected to Become the Next Phase of Multimodal Large Models</h2>
<p>World models that place greater emphasis on &#34;causal&#34; reasoning endow AI with higher levels of cognition and more logical reasoning and decision-making capabilities. This ability not only promotes the deep application of AI in cutting-edge fields such as autonomous driving, robotic control, and intelligent manufacturing, but also has the potential to break through traditional task boundaries and explore new possibilities for human-computer interaction.</p>
<h2>Trend Six: Synthetic Data Will Become an Important Catalyst for the Iteration and Application of Large Models</h2>
<p>High-quality data will become a barrier to the further scaling up of large models. Synthetic data has become the preferred choice for foundational model vendors to supplement data. Synthetic data can reduce the costs of manual governance and labeling, alleviate dependence on real data, and no longer involve data privacy issues; it enhances 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 acquisition costs of proprietary data, promoting the application of large models.</p>
<h2>Trend Seven: Accelerated Iteration of Reasoning Optimization Becomes a Necessary Condition for the Landing of AI Native Applications</h2>
<p>The hardware carriers of large models are penetrating from the cloud to edge hardware such as mobile phones and PCs. On these resource-constrained devices (AI computing power, memory, etc.), the application of large models will face significant limitations in reasoning costs, posing huge challenges to deployment resources, user experience, and economic costs. Continuous iteration of algorithm acceleration and hardware optimization technologies will drive the rapid landing of AI Native applications.</p>
<h2>Trend Eight: Reshaping Product Application Forms, Agentic AI Becomes an Important Model for Product Landing</h2>
<p>By 2025, more general and autonomous agents will reshape product application forms, further integrating into work and life scenarios, becoming an important application form for the landing of large model products. From Chatbot, Copilot to AI Agent, Agentic AI, the industry's understanding of AI application forms has deepened since 2023. From agents that emphasize product concepts to Agentic AI that emphasizes the level of application intelligence, we will see more highly intelligent multi-agent systems with a deeper understanding of business processes landing in applications by 2025.</p>
<h2>Trend Nine: The Heat of AI Applications is Rising, It Remains to Be Seen Who Will Dominate the Super App</h2>
<p>In the past year, the processing capabilities of generative models in images and videos have significantly improved, combined with cost reductions brought about by reasoning optimization, and the continuous development of technologies such as Agent/RAG frameworks and application orchestration tools, laying the foundation for the landing of AI super applications. Although it is still uncertain who will dominate the Super App, from the perspectives of user scale, interaction frequency, and dwell time, the heat of AI applications continues to rise, reaching the dawn before an application explosion.</p>
<h2>Trend Ten: Balancing Model Capability Enhancement and Risk Prevention, AI Security Governance System Continues to Improve</h2>
<p>As a complex system, the scaling of large models has led to emergence, but the unique properties of complex systems, such as unpredictable emergent results and feedback loops, also pose challenges to traditional engineering safety protection mechanisms. The continuous progress of foundational models in autonomous decision-making brings potential risks of loss of control. How to introduce new technological supervision methods and how to balance industry development and risk control in human supervision?
This is a topic worth continuous discussion for all parties involved in AI.</p>
<p>It is worth mentioning that China's self-developed technologies and products have also become representative cases of AI development trends. For example, in the multimodal field, the Zhiyuan Research Institute released the fully self-developed native multimodal world model Emu3 based on autoregressive technology, achieving unified understanding and generation of video, images, and text.</p>
<p>In the field of model applications, Doubao's monthly active user count reached 71.16 million by December 2024, making it the number one AI-native application in China and the second globally. In the service-oriented intelligent agent track, Ant Group's series of AI butler products, such as Zhixiaobao and Maxiaocai, have reshaped the form of AI products</p>

80020

00020

09880

- The Beijing Zhiyuan Artificial Intelligence Research Institute predicts key trends for AI models in 2025, focusing on areas like Scaling Law, foundational models, embodied intelligence, and AI safety.  
- AI for Science (AI4S) is expected to transform scientific research methodologies, while embodied intelligence will evolve through collaboration between physical and cognitive systems.  
- The report emphasizes the importance of synthetic data for model development and the need for improved AI safety governance as model capabilities increase.

A Comprehensive Understanding of the Top 10 AI Technology Trends in 2025