With the release of Ernie 4.0, Baidu claims that overall performance is “on par with ChatGPT.”

Baidu, a technological behemoth in China, unveiled version 4.0 of its widely embraced “Ernie” artificial intelligence (AI) large language model (LLM) chatbot. Per the English rendition offered by Baidu, embedded in an X post adorned with customary Google Translate subtitles, CEO Robin Li declared the updated model’s “comparable prowess to GPT-4 in overall performance.” Robin Li delineated the enhanced capabilities of the new Ernie model across four distinct dimensions: comprehension, generation, reasoning, and memory. Within the “comprehension” domain, Baidu asserts an overarching enhancement in human-computer interaction. The announcement thread cites, “ERNIE Bot can adeptly decipher ‘out-of-order statements, vague expressions, and implied meanings’ in text.” This achievement is noteworthy, given Ernie’s purported training in both Chinese languages and English—teaching models to grasp colloquial or conversational cues has traditionally posed challenges for LLM engineers. In a departure from its predecessors, Ernie 4.0 evidently boasts markedly improved capabilities in both the quality and speed of generating images, videos, and coherent copy. Another post in the thread attests, “With just one image and a handful of prompts, we crafted 1 video ad, 5 ad copies, and 1 poster in a mere 3 minutes. ERNIE Bot transforms a sole individual into a marketing powerhouse.” Under the “reasoning” banner, CEO Robin Li showcased the model’s advanced analytical problem-solving prowess by posing a convoluted question. Past LLM models have grappled with challenges necessitating any form of reasoning. In the demonstration, Ernie furnishes both textual and visual responses, along with a concise and demonstrably accurate answer. The concluding facet of the 4.0 update involves expanding Ernie’s so-called “memory.” To the best of our knowledge, there’s no scientific substantiation that LLMs or any AI system can genuinely “reason” or possess true “memory.” However, akin to these concepts is a model’s aptitude to process problems and recall prompts and outputs from preceding sessions. As per Robin Li, Ernie’s “memory” rivals that of ChatGPT. “Even after five rounds of conversation and the generation of thousands of words, ERNIE Bot can recollect content generated in previous instances.”