[Based on Nvidia’s Megatron-LM code, apparently, and 2–4 week-long training runs on 1120 GPUs (140 DGX nodes, ‘SuperPod’). Trained on Korean, but they plan to do English as well. For more details, see “[컨퍼런스 리뷰] NAVER AI NOW 2021 하이라이트”.] South Korea’s IT giant Naver unveiled a supersized artificial intelligence platform on Tuesday that can process massive amounts of data, saying it aimed to lead the era of “hyperscale” AI. Dubbed HyperCLOVA, the firm’s latest AI tool is not only the local industry’s most advanced but is also the first large-scale AI trained in the Korean language, Naver said.
…At Tuesday’s event, Naver demonstrated various uses for HyperCLOVA. The AI tool could chat with humans, come up with product introductions from a few keywords, summarize complicated documents, and even select and organize the data needed to train another AI tool.
HyperCLOVA’s AI language tool has 204 billion parameters, more than San Francisco-based Open AI’s GPT-3.
…Naver has already deployed HyperCLOVA for its portal site, adding a function early this month that allows it to automatically correct typos and recommend related search keywords. The firm plans to add foreign languages, videos and images into HyperCLOVA’s neural network and to advance its machine learning capabilities. It also plans to actively work with businesses and researchers to diversify the AI tool’s uses.
The race to develop hyperscale AI is heating up among local companies. Telecommunications provider SKT joined hands with Kakao in March and declared they were developing their own version of a hyperscale AI platform. LG Group announced last week that it would invest $1 million in hyperscale AI technology and come up with a prototype by the end of this year.
