Other Stories

Full Video

S. Korea's tech giants plan to lead the world in AI by developing large language models Updated: 2021-06-23 13:25:10 KST

Many of us will have had the experience of dealing with a computer automated voice when calling customer service.
It can be frustrating if the AI doesn't understand what you're saying.
On the other hand, AI is advanced enough to write simple news articles about sports and finance, even if they do need a human touch in the end.
But there's a new generation of AI in development in the form of "language models" from South Korea's tech giants as well as Microsoft, Google and Facebook.
Experts say this will be the next big breakthrough in artificial intelligence.
Today, I'm joined in the studio by our Bae Eun-ji, who's been looking into this technology and is here to get us all up to speed.
Eun-ji, great to have you here today.

Good afternoon.

So, Eunji, all this sounds quite complicated.
Break it down a bit, if you would what is an AI language model?

Put simply an AI language model is a system that can analyze and use datasets and it has a strong resemblance to the human brain.
A human brain has 100-trillion synapses, which are the structures that enable neurons to carry messages through the brain.
( 1) In an AI platform, the thing that is similar to the synapses are called "Parameters."
( 2) They are variables used to pass information in the computer program.
So more parameters means better AI.
One of the largest language models ever trained is GPT-3 ( 3) a model with one-hundred-75 billion parameters.
GPT-3 was created in May 2020 by Open AI a research firm co-founded by Elon Musk.
It's not open to the public just yet but it can write essays, summarize long texts, and even write tweets just like humans.
Also it can convert our words into computer code.
So using this, anyone can easily make an app or create website layouts even people like myself who don't know anything about programming.

That sounds like something that would revolutionize a lot of the creative aspects of business, and all kinds of media, really.
How about Korean companies? Are there any that have created their own language models?

Well, yes.
Last month South Korea's biggest web service operator, Naver, unveiled its language model called 'Hyper-Clova' trained in the Korean language.
It can write e-mails and answer questions.
If you ask a question and tell it to explain the answer in a simple way, it will do just that.
The software engineer in charge of developing this says it was trained on a massive amount of data almost as much as GPT-3.

"Our model was trained on datasets of 560 billion tokens which is similar to the amount that GPT-3 was trained on. This is about 50 years' worth of news articles."

The platform isn't available to the public yet but Naver plans to provide its services to small and medium-sized companies, creators and startups.
Right now it's used in the Naver search engine to provide a wider range of search suggestions.
It helps users when they don't know how to spell something, or when they don't know the right words to use.
In the future, Naver hopes this service will help provide some level of companionship for lonely seniors, like as a conversation partner.

That sounds like a big breakthrough.
And other South Korean companies are also working on their own models?

That's right.
LG announced last month that it's going to invest more than one-hundred-million dollars over the next three years.
It aims to develop a program that contains six-hundred billion parameters three times larger than GPT-3 by the end of this year.

"For example, all the scientific discoveries in chemistry that humans made in the past 250 years are on paper, not digitalized. So this makes it hard for an AI to use this information. With a large AI model, we hope to discover new chemical materials. This will accelerate the development of electric vehicles and display panels."

Telecom operator KT has also recently established an AI research institute with Korea Advanced Institute of Science and Technology, or KAIST, to develop its own model.

From what you have explained so far, it seems it can do a lot of different tasks that have to do with language.
Can it be used in any other way apart from reading or writing?

Well Devin it can also be applied to robotics.
An AI lab that I went to at Seoul National University was working on a humanoid robot designed to interact with humans.
It aims to have the robot learn house chores and do other simple tasks like serving customers in restaurants while having actual conversations with people.

"The last robot to be introduced is me, Pepper. I am a humanoid robot made by SoftBank, and I use artificial intelligence."

The director, a computer science professor, said that a more sophisticated AI program will allow this robot to do even more complex tasks.

"The development of a large AI model can make this robot smarter. This robot right here has its brain connected to the computer. If we can use the AI tool, the robot can do a lot more tasks."

It does sound like AI will be doing a lot of work that humans currently do.
Thank you Eun-ji.
KOGL : Korea Open Government License
본 저작물 중 본문에 해당하는 뉴스 스크립트(텍스트)는 공공누리 제1유형-출처표시 조건에 따라 이용할 수 있습니다.