“It is not inferior in any aspect to GPT-4,” Baidu’s CEO Robin Li boldly claimed, stating that the latest model was “significantly improved” compared to its original Ernie Bot, at the recent launch of Ernie 4.0.
Robin Li demonstrated Ernie 4.0 at the company’s annual Baidu World conference in Beijing. He said the model has achieved comprehension, reasoning, memory and generation, which uses algorithms to produce and create new content. Li said that Ernie 4.0 was able to understand complex questions and instructions and apply reasoning and logic to generate answers to questions.
While many speculated that OpenAI’s challenger might emerge from the United States, it seems the competition has arisen from China, Baidu’s Ernie 4. Ernie, not to be confused with the friendly Sesame Street character, has been making waves in the AI community, and this latest iteration promises to be a game-changer.
ERNIE 4.0’s performance has improved by nearly 30% since its beta testing in September and is currently available in an invitation-only beta.
How Far Has Ernie Actually Come
What sets Ernie 4 apart from its predecessors is its planned integration into almost all of Baidu’s products. This includes prominent services like search engines, maps, and cloud computing. Such widespread integration is a significant move in the AI world, underlining Baidu’s confidence in Ernie 4’s abilities.
ERNIE 4.0 and the user-facing text chatbot, ERNIE Bot, were designed for understanding, generating, reasoning, and memorising. Live demonstrations showcased ERNIE Bot’s ability to understand complex human requests and generate various types of content, such as text, images, and videos. Clearly, Baidu realises that the multimodal race is on.
In a live demonstration, Li prompted Ernie 4.0 to generate advertising materials including advertising posters and a marketing video. He also asked Ernie 4.0 to come up with a martial arts novel complete with characters with various personalities.
The architecture of Ernie 4 is a topic of interest. The model relies on self-supervised pre-training and innovative techniques for data processing. It involves a unique approach to constructing datasets and employs techniques such as natural language processing. The model is said to harness massive datasets, including the impressive A3 Corpus, comprising around 10 billion parameters.
Notably, Ernie 4 emphasises green and sustainable AI. It adopts an online distillation technique, which seeks to transfer knowledge signals from a “teacher model” to multiple “student models” of varying sizes. This approach results in significant power savings compared to traditional distillation methods. Ernie 4 even introduces a “teacher assistant model” to bridge the knowledge gap between deep and shallow models.
Moreover, the model boasts an end-to-end adaptive distributed training framework, designed for versatile and adaptive usage in industrial applications and production environments. This framework considers resource allocation, model partition, task placement, and distributed execution, making it an attractive choice for various real-world scenarios.
Lukewarm at Best
Baidu introduced its first Ernie Bot in March. However, there is a notable challenge when it comes to Ernie 4. Unlike its competitors, it is not yet publicly available, which has left tech enthusiasts and developers eager for a glimpse of its capabilities. Access to source code or demos has been limited, leaving much to the imagination.
However, analysts expressed disappointment with the launch of Ernie 4.0, highlighting a lack of major improvements compared to its previous version. Baidu’s Hong Kong shares also fell by 1.32% during morning trading, underperforming the broader Hang Seng Index’s 0.7% rise.
According to Lu Yanxia, an analyst at IDC, while substantial enhancements may become evident with hands-on use, specific upgrades in Ernie 4.0 are not immediately apparent.
Potential Disruption
China has recently sought to regulate the generative AI industry, requiring companies to carry out security reviews and obtain approvals before publicly launching their products. Companies that provide such AI services must also comply with government requests for technology and data.
Regardless, Baidu had also recently raised $140 million, however, analysts’ and market reaction to the announcement indicates that the Chinese giant has a long way to go.
Baidu also announced the integration of AI technology into its suite of products, including Baidu Search, Baidu GBI, Infoflow, Baidu Wenku, Baidu Maps, Baidu Drive, and its business intelligence offerings for enterprise customers—to enhance user experiences and productivity.
However, one thing is clear, through its integration into a variety of Baidu’s offerings it will look to challenge various entities. For instance, many believe Ernie 4.0 is set to disrupt marketing, and will challenge traditional creative platforms like Adobe and Canva—-who will have to step up their game.
The post Baidu’s Ernie 4 Aims to Outshine GPT-4 appeared first on Analytics India Magazine.
