On November 19, Robin Lee, the founder of Baidu, said at the 2024 China 5G+Industrial Internet Conference that it is necessary to realize that AI is a new industrial revolution. Only by thinking in a higher dimension can we truly make use of the big model, enable thousands of industries and improve social productivity.
Data shows that as of early November, the daily average number of adjustments for Baidu Wenxin's large model reached 1.5 billion, an increase of 7.5 times compared to the 200 million disclosed in May. Robin Lee said that the large model has a high call volume and a fast growth rate, indicating that more and more applications are using the Wenxin large model. With the continuous improvement of Retrieval Enhancement (RAG) capability, the growth rate of basic model calls has also been very rapid in the past six months.
"When the big model was first released, it had hallucinations and often talked nonsense seriously. When the model had hallucinations and randomly generated various content, it was not available in most scenes." Robin Lee said that the biggest change of the big model in the past 24 months was the basic elimination of "hallucinations".
At present, the RAG at the textual level has been continuously improved, achieving the usability and trustworthiness of large models. However, in terms of multimodal technologies such as imaging, accuracy and controllability still need to be addressed in order to expand the application space of AI. To solve the problem of "illusion" in image generation, Baidu has developed iRAG, a retrieval enhanced text image technology that combines Baidu's search image resources with basic model capabilities to generate hyper realistic images.
"Now using Wenxin multimodal model to generate can remove the 'illusion' and the so-called 'AI flavor', and the generated pictures look more realistic and retain the accuracy." Robin Lee believes that the future multimodal retrieval enhancement will also have rapid development, so that the multimodal large model will enter a more practical stage.
With the capability of basic large-scale models ready, application driven industrial innovation is rapidly landing. Especially in the field of autonomous driving, large models have already had very good applications. Especially behind end-to-end, pure visual big models, L4 level autonomous driving and other technologies, big model technology support is needed.
It is reported that in May this year, Baidu first released the world's first L4 level end-to-end autonomous driving model Apollo ADFM, which can balance the safety and generalization of technology, achieving safety more than 10 times higher than that of human drivers. The Apollo 10.0 version, an autonomous driving open platform equipped with this large model, will soon be released to users worldwide. This upgrade will significantly enhance the safety, intelligence, and usability of the autonomous driving open platform.
"We can't compare big models and generative AI with PC Internet and mobile Internet." Robin Lee said that AI is a new industrial revolution. We should refer to the development process of steam engine revolution, electric power revolution and information revolution, and think about how to seek benefits and avoid disadvantages from this dimension. Only in this way can we truly utilize big models to empower various industries and improve social productivity.