Cho Kyung-hyun, a New York University (NYU) professor known as an "AI luminary," said it was "regrettable" that Naver Cloud was disqualified in the first-stage evaluation of the government's "independent AI foundation model project" because it used pretrained vision and audio encoders.
On the 16th, Cho wrote on his social media account, "For disclosure of interest, I am a member of Upstage's board and advise on the technical aspects of training Upstage's large language model (LLM)."
On the 15th, the government eliminated Naver Cloud and NC AI in the first round from among the five elite teams participating in the "national AI model project." The NC AI elite team fell short in benchmark, expert and user evaluations, while Naver Cloud failed to meet the originality requirement. As a result, the elite teams centered on Upstage, SK Telecom and LG AI Research have advanced to the second stage.
Regarding this, Cho said he was "surprised that the Naver Cloud model was disqualified because it was not evaluated as being built 'from scratch,'" adding, "As I see it, the 'I' in AI lies in seamlessly integrating sequences of observed information—such as tokens, images and audio—through very high-performing neural network models. In this sense, many of these sensory encoders are similar to token embedding layers (layers that convert tokens into vectors)."
Cho said there could be differing views on Naver Cloud's disqualification, but he opposed tightening the evaluation criteria in response. "In the end, a decision had to be made, and I respect the final outcome regardless of my disagreement," he said. "Some are calling for the criteria to be defined more carefully and more strictly because of Naver Cloud's disqualification, but I strongly oppose such proposals."
He added that such proposals are "the kind of thing people who are used to and comfortable with multiple-choice exam-based promotion and selection would make," calling them "claims by those with a bureaucratic mindset." He wrote, "I oppose the argument that it does not matter whether an exam actually measures a candidate's true ability, as long as it can produce a single score to rank candidates."