The Chinese megacorporation claims its newest model rivals OpenAI’s popular model in generating text, images, and video.
Baidu, one of China’s largest technology companies, released version 4.0 of its popular “Ernie” artificial intelligence (AI) large language model (LLM) chatbot.
According to an English language translation provided by Baidu, embedded in an X post alongside the usual Google Translate subtitles, CEO Robin Li claimed the updated model “stands on par with GPT-4 in terms of overall performance.”
Baidu CEO Robin Li announced ERNIE 4.0 foundation model, now available for invited testing. ERNIE 4.0 elevates foundation model capabilities in understanding, generation, reasoning, and memory.
Read the thread to find out how. #BaiduWorld2023 pic.twitter.com/lSVwVl88QH
— Baidu Inc. (@Baidu_Inc) October 17, 2023
Robin Li described the updated capabilities of the new Ernie model through four distinct verticals: understanding, generation, reasoning, and memory.
Under the category of “understanding,” Baidu claims overall improvement in human-computer interaction. “ERNIE Bot can now accurately interpret ‘out-of-order statements, vague expressions, and implied meanings’ in text,” reads one post in the announcement thread.
This is particularly noteworthy, as Ernie has reportedly been trained in both Chinese languages and English — teaching models to understand colloquial or conversational prompts has, traditionally, been a challenging task for LLM engineers.
Related: Biden considers tightening AI chip controls to China via third parties
In contrast to previous versions, Ernie 4.0 also appears to have significantly improved capabilities in both quality and speed when it comes to generating images, video, and coherent copy. “With just one image and a few prompts,” reads another post in the thread, “we created 1 video ad, 5 ad copies, and 1 poster in just 3 minutes. ERNIE Bot transforms a single person into a marketing team.”
In the area of “reasoning,” CEO Robin Li demonstrated the models’ advanced analytical problem solving capabilities by posing a complex question. Similar LLM models of the past have struggled with problems which require any form of reasoning. In the demonstration, Ernie provides both text and imagery as well as a succinct and demonstrably correct answer.
Memory: Even after five rounds of conversation and writing thousands of words, ERNIE Bot can remember previously generated content, showcasing the remarkable memory capabilities of ERNIE 4.0 foundation model. pic.twitter.com/23ajzd5uxl
— Baidu Inc. (@Baidu_Inc) October 17, 2023
The final leg of the 4.0 update involved expanding Ernie’s so-called “memory.” To the best of our knowledge, there’s no scientific evidence that LLMs or any AI system can “reason” or have any form of actual “memory,” but analogous to those terms would be a model’s ability to process problems and recall prompts and outputs from previous sessions.
According to Robin Li, Ernie’s “memory” is about as good as ChatGPT’s. “Even after five rounds of conversation and writing thousands of words, ERNIE Bot can remember previously generated content.”
