On January 8, Beijing Academy of Artificial Intelligence (BAAI) unveiled its “Top 10 AI Technology Trends for 2025.” This highly anticipated report provides insights into emerging directions across AI infrastructure and applications, offering predictions in areas such as scaling laws, foundational models, embodied intelligence, super apps, and AI safety.

The trends were presented with rigorous supporting arguments, analyzed by experts from academia and industry, including representatives from Tsinghua University and BAAI research centers. Notably, the trends include groundbreaking Chinese innovations in AI technologies and products.

In the realm of multimodal AI, BAAI introduced Emu3, an entirely self-developed, autoregressive native multimodal world model that seamlessly integrates video, image, and text understanding and generation. On the application side, Doubao, a native AI application, reached 71.16 million monthly active users in December 2024, becoming China’s leading and the world’s second-largest AI-powered app. Meanwhile, Ant Group’s AI products like Zhixiaobao and Maxiaocai redefined the landscape of AI agents.

AI Safety Gains Ground

As AI’s capabilities grow, so do concerns over safety. In March 2024, BAAI hosted China’s first high-level international AI safety forum, culminating in the signing of the “Beijing AI Safety Consensus” with global AI leaders and scholars. In April, the United Nations Technology Summit endorsed two safety standards for large models, with Ant Group leading the development of the “Large Language Model Safety Testing Methodology.”

BAAI’s president, Wang Zhongyuan, emphasized, “We are at a turning point for AI development. The accelerated emergence of large models propels us into an era of general artificial intelligence (AGI). Unified multimodal, embodied intelligence, and AI for Science will deepen our understanding of the world and connect digital and physical realms. BAAI aims to guide the industry with these trends, shaping the future together.”

The 10 Key AI Trends for 2025:

Trend 1: AI for Science (AI4S) Revolutionizes Scientific Discovery

Underpinned by large models, AI4S is reshaping scientific research methodologies. By 2025, multimodal models will empower the analysis of complex data structures in fields like biomedicine, materials science, and climate modeling, redefining foundational and applied sciences.

Trend 2: The Era of Embodied Intelligence Begins

The industry will see a consolidation among nearly 100 startups focused on embodied intelligence. Technical breakthroughs in end-to-end modeling and the rise of “small brain” large models are expected. Industrial applications, including humanoid robot mass production, will showcase real-world embodiment at scale.

Trend 3: Unified Multimodal Models Push Efficiency Boundaries

AI evolves by mimicking human cognitive processes. Native multimodal models that integrate vision, audio, and 3D data during training will offer unprecedented efficiency and accuracy, moving beyond piecemeal approaches.

Trend 4: Scaling Law Extensions to Post-Training and Inference

While traditional scaling laws are less cost-effective, reinforcement learning and task-specific fine-tuning offer new pathways to maximize model performance, particularly in post-training and inference.

Trend 5: World Models Redefine Multimodal Evolution

With a focus on causal reasoning, world models provide AI with logical decision-making and advanced cognitive abilities. This evolution will unlock new possibilities in fields like autonomous driving, robotics, and intelligent manufacturing.

Trend 6: Synthetic Data Catalyzes Model Innovation

The scarcity of high-quality data is a bottleneck for scaling models. Synthetic data emerges as a viable solution, reducing costs, safeguarding privacy, and enhancing model robustness across long texts and complex tasks.

Trend 7: Inference Optimization for AI Native Applications

The shift from cloud to edge devices like smartphones and PCs creates challenges in deployment efficiency. Iterative advances in algorithm acceleration and hardware optimization will pave the way for broader adoption.

Trend 8: Agentic AI Shapes Product Applications

Agentic AI represents a leap forward from static AI assistants to intelligent agents capable of understanding complex workflows. These agents will transform both business and everyday life in 2025.

Trend 9: The Race for AI Super Apps Intensifies

Generative models have seen significant improvements in image and video processing. Coupled with cost-reducing inference optimizations, they form the backbone of the next wave of AI super applications.

Trend 10: Balancing Model Advancements and Safety Governance

The rise of autonomous decision-making in large models introduces risks of unintended outcomes. BAAI stresses the need for robust safety mechanisms, combining technical and regulatory approaches to ensure balanced growth.