Baidu's Wang Haifeng: Ernie 4.0 Achieves 50x Improvement in Inference Performance
-
At the recent 20th China Computer Conference, Baidu's Chief Technology Officer Wang Haifeng stated that Baidu released the knowledge-enhanced large language model Ernie Bot on March 16, and its foundational model has now been upgraded to Ernie 4.0. The knowledge-enhanced large language model integrates learning from trillions of data points and hundreds of billions of knowledge elements to form a pre-trained large model. Building on this, it employs supervised fine-tuning, human feedback reinforcement learning, and prompting techniques, offering advantages in knowledge enhancement, retrieval enhancement, and dialogue enhancement.
Wang Haifeng noted that with a stronger platform, better data, and improved algorithms, Baidu has trained the larger and more effective Ernie 4.0 model. Through the collaborative optimization of PaddlePaddle and Ernie, the model achieves a weekly training efficiency of over 98%, a 3.6x improvement in training algorithm efficiency, and a 50x boost in inference performance. Ernie 4.0 shows significant enhancements in four key capabilities: understanding, generation, logic, and memory. The improvements in understanding and generation are comparable, while logic and memory show even greater gains—logic improving nearly three times as much as understanding, and memory more than doubling the improvement seen in understanding.
Public data reveals that within just over a month of its public release, Ernie Bot has attracted 45 million users, 54,000 developers, 4,300 scenarios, 825 applications, and over 500 plugins. The PaddlePaddle and Ernie ecosystem continues to thrive, now uniting 8 million developers, serving 220,000 enterprises and institutions, and enabling the creation of 800,000 models based on PaddlePaddle.