AI Large Models Land in Multiple Industries: 'Embodiment' Emerges as a New Trend
-
At the end of 2022, the successive debut of large models such as ChatGPT, Midjourney, and Stable Diffusion sparked a wave of development in artificial intelligence large models. Domestic companies like Baidu, SenseTime, 360, Unisound, and iFlytek have also released their own achievements, promoting the application of AI technology across various industries.
AI Large Models Show Great Potential
AI large models have demonstrated immense potential and rapid development trends in fields such as language understanding, image recognition, and natural language processing. They can help enterprises and institutions accelerate innovation, improve efficiency, and reduce costs. For example, in natural language processing, large models can be applied to machine translation, text summarization, and intelligent customer service. In image recognition, large models can provide more accurate image recognition and analysis capabilities.
A report jointly released by Sootoo and Large Model Home shows that, according to forecasts, the global AI large model market size will reach $21 billion by 2023 and $109.5 billion by 2028. As a key player in the global AI large model market, China is expected to see its large model industry market size reach RMB 14.7 billion by 2023 and RMB 117.9 billion by 2028.
'Novel-to-Movie' Becomes Possible
What changes will the development of AI large models bring? First, it will trigger a revolution in AI computing power. The increase in large model parameters leads to an exponential growth in computational demands during training. To train large-scale models quickly, powerful computing capabilities are required to support efficient distributed training and parallel computing. With advancements in hardware, such as high-performance computers and the widespread adoption of distributed computing platforms, these will become crucial methods for supporting larger-scale model training and iteration.
Second, it will spark a revolution in AI multimodality and multi-scenario applications. With the development of sensor technology and the rise of the Internet of Things, vast amounts of multimodal data—such as images, videos, and sounds—are being widely collected and applied. Large models can leverage this multimodal data for cross-modal learning, enhancing their performance in various perception tasks. This means that generating a movie from a novel or converting a movie into a novel will become possible.
Moreover, large models possess strong generalization capabilities, enabling knowledge transfer and sharing across different modalities and scenarios, thereby expanding their applications to diverse fields and contexts.
Third, by integrating multimodal data and intelligent algorithms, large models can empower multiple industries, improving efficiency and quality, driving the fusion of data and physical entities, and transforming industry landscapes.
For example, in the legal field, large models can serve as intelligent contract generators, automatically producing legally compliant and reasonable contract texts based on user needs and specifications. In entertainment, they can act as intelligent scriptwriters, generating engaging and appealing storylines tailored to user preferences. In manufacturing, they can function as intelligent quality controllers, automatically detecting and correcting product quality issues based on production data and standards. In education, they can serve as intelligent learning platforms, recommending and organizing learning resources according to knowledge graphs and learning paths.
Finally, large models will revolutionize user habits. Statistics show that a high percentage of people are already aware of large models, and a significant proportion have used them. Particularly in first-tier cities, awareness and the desire to use large models are even more widespread and intense.
"Embodied AI" may become a new technological direction for the development of large models. It is worth noting that embodied AI and humanoid robots are not the same thing. The most significant characteristic of embodied intelligence is its ability to autonomously perceive the physical world from a protagonist's perspective, learn through human-like cognitive pathways, and provide behavioral feedback as expected by humans, rather than passively waiting for data input. The development of this technological direction will further enhance the intelligence of large models, expand their application scenarios, and deepen interaction with humans.
Taking large models with natural language interaction as an example, they not only possess conversational abilities but can also perceive and understand objects and events in the environment, as well as respond appropriately based on human intentions and needs. Through human-machine dialogue, large models can issue commands to robots, understand robot feedback, decompose tasks into actions, assist robots in processing multimodal data such as images and sounds, guide robots to correct errors, and improve their learning capabilities.