can the tv series be generated by inputting "dream of red mansions"? the director of beijing general artificial intelligence research institute said so
2024-09-20
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
input a dream of red mansions, output a tv series? how far can generative models go? zhu songchun, director of the beijing institute of general artificial intelligence, answered this question.
on september 20, the 2024 beijing cultural forum was held in beijing. zhu songchun attended the parallel forum "cultural trends: integration of emerging business forms and technologies" and delivered a keynote speech entitled "artificial intelligence: giving machines a 'heart' and giving humanities 'reason'".
zhu songchun first introduced the limitations of current generative artificial intelligence. in the multiple wensheng video cases he demonstrated on site, there were a lot of physical common sense errors and subject inconsistency problems, which were mainly manifested in the spatial scale errors, occlusion relationship errors, hand structure errors and background content errors of the generated videos.
he summarized the limitations of generative statistical models driven by big data as follows: the current generative models of text, images, and videos are a reintegration of existing data, which makes it difficult to understand and communicate the creator's true intentions, lack cognitive architecture, and lack understanding of physical and social common sense and consistent internal expression. excellent works of art should be based on the creator's inner heart, but ai lacks a human value system.
based on this, he believes that the fields of technology and art should work together to create new art forms with synaesthesia, concentricity, and resonance, to create universal intelligent entities, and ai models with common cognitive structures, internal expressions, and value systems.
