Baidu on Thursday said the text-to-picture function of its ChatGPT-like tool ERNIE Bot is powered by its own cross-modal generation big model ERNIE-ViLG, amid recent speculation that the search giant’s platform might be translating prompts into English and then using other companies’ AI tools to generate results. An influencer claimed to have tested the bot by asking it to draw a picture of “yikeyidouzi,” an incomprehensible term in Chinese, which literally translates to “one can bean” in English; Baidu’s chatbot service generated a picture of a can of beans in response. That sparked a wave of users posting images onto social media that they claimed were erroneously generated by ERNIE Bot, such as an image of a turkey drawn from the prompt “Turkey” (the country). The Beijing-based company has since addressed some of the issues by fixing prompts that previously generated incorrect images, with a hot-air balloon and a young Turkish girl now appearing in response to a prompt of “Turkey.” As criticism mounted, Baidu issued a statement via Weibo on Thursday, asserting that ERNIE Bot is “completely Baidu’s own large language model,” trained using global internet public data. Baidu added that the tool will continue to develop and fine tune its AI-generated image capabilities, demonstrating its “self-research power.” The company also appealed for users to give ERNIE Bot time and to believe in its capabilities, not spread rumors. [TechNode reporting]
0 Commentaires