Baidu has launched the latest iteration of its ChatGPT-like AI product Ernie Bot and claims it now rivals OpenAI’s in important metrics.
According to a statement from Baidu, the company behind China’s leading search engine, Ernie 3.5 was tested against two commonly used benchmarks for AI models. One tests the model’s ability to perform human problem-solving through public tests like the gaokao, the American SAT, law school admission tests, and civil service entrance examination tests. The other benchmark tests the model’s Chinese language ability. Baidu reportedly stated that Ernie 3.5 outperformed ChatGPT “in comprehensive ability scores” and “in several Chinese capabilities”.
Baidu unveiled the first version of Ernie in March, which was quickly followed by similar products from Tencent and Alibaba. The global response to Ernie’s launch was subdued, as it was introduced with a demo of it performing various tasks and was only made available to a limited number of users. With the public unable to unleash its curiosity onto the new technology as they had done with ChatGPT, investors were disappointed, and Baidu’s stocks tumbled as a result.
Developers of Chat-GPT rivals in China are up against a major obstacle in the form of state censorship. The closed ecosystem of China’s Great Firewall means the data available to train Ernie and other AI models is severely limited. AI makes predictions based on patterns it recognizes from datasets, much of which are from publicly available text data found on the internet.
The chief architect of the Great Firewall, Fang Binxing, has argued that generative AI could make this kind of government-led information control even more powerful and potentially damaging. “People’s perspectives can be manipulated as they seek all kinds of answers from AI,” Fang said. So far, the Chinese government has been keen to keep AI models under wraps from the public, so it remains to be seen whether this technology will be harnessed as a widespread propaganda tool.
