Talking AI in South Korea and the rest of the globe: An interview with NAVER

An exclusive deep-dive with a global tech innovator and the nation's leading search engine

Hello everyone!

Today, we continue our Insider’s Corner series with an exciting new interview. This time, we’re sitting down with a spokesperson from NAVER — a search engine based in South Korea and one of the world’s leading tech companies. 

Founded 25 years ago, NAVER has evolved into a regional powerhouse, offering a broad array of web services and increasingly advanced AI technologies. NAVER’s Communications Department kindly facilitated our discussion with their ML scientists and product leads who will share insights on the state of AI in South Korea and beyond.

Image by Tortoise Media

According to the reputable Tortoise Global AI Index, South Korea is currently ranked 6th in the world, with impressive “Scale” and “Intensity” ratings representing the nation’s absolute and relative AI capacities, respectively. However, South Korea is only 18th when it comes to the “Commercial” metric.

The country is home not only to NAVER but also tech giants like Samsung, LG, and Hyundai (the parent company of Boston Dynamics). Given all that, what do you think is lacking?

NAVER, along with other Korean companies, has the capability to develop hyperscale AI independently. However, the AI business ecosystem, crucial for creating and utilizing new services based on foundation models, is not evolving as quickly as the novel technology itself. Despite being the first in Korea to develop an LLM, we can’t deny that the integration of AI into everyday life still faces challenges, including safety concerns and cost-efficiency.

Moreover, no country or company has yet perfected a robust business model for AI, with most still in the nascent stages of development. Efforts to expand the AI ecosystem commercially require national support, namely helping startups and research institutions leverage hyperscale AI to innovate and create new services.

South Korea boasts some of the world’s largest semiconductor producers, including SK Hynix. Meanwhile, NAVER is focusing primarily on the software side of AI. Is there any symbiosis between South Korean companies, a quid-pro-quo type of technology-sharing partnership, enabling everyone in the nation to stay ahead of the AI curve?

Actually, NAVER is actively engaged in research and development both in software and AI semiconductors. Our collaborations extend to major industry players like Intel. In South Korea, the whole AI ecosystem is a thriving synergistic network involving various companies. 

This network includes not only those producing or fine-tuning AI models but also a multitude of enterprises creating other innovative services, IT companies aiding in technology integration, as well as firms that develop the necessary data infrastructure for AI advancements. So, yes, through these interconnected relationships, each entity contributes to and benefits from the collective growth, keeping South Korea at the forefront of AI technology.

South Korea has one of the highest, if not the highest, smartphone penetration rates in the world, along with some of the fastest mobile internet speeds. How are AI and data companies, including NAVER, using this to their advantage?

It’s true, South Korea’s notable smartphone penetration rate and rapid mobile internet speed create an ideal environment for nurturing AI technologies. Moreover, Korean users are typically very eager to adopt new solutions, providing a great platform for AI testing through user feedback. Ever since we introduced HyperCLOVA X, a proprietary AI model developed by our engineers, we’ve been refining it iteratively based on user interactions.

Additionally, NAVER continues to leverage tech expertise to expand globally, customizing AI tools to meet regional needs. For instance, we recently partnered with Saudi Arabia’s Aramco Digital to develop an Arabic LLM, aiming to foster an AI ecosystem that caters specifically to the Middle East.

Because of NAVER, South Korea is one of a handful of countries in the world where Google isn’t the dominant search engine (e.g., China/Baidu). Was it a difficult feat for NAVER to achieve and how does it translate to AI superiority in South Korea?

Yes, NAVER has a unique position in this regard. This achievement stems from our early focus on enhancing Korean language content and services online. It began in the early 2000s when we initiated extensive collaborations to build a reliable repository of Korean web documents. Our expertise-sharing platform Knowledge iN, along with Blogs and Cafés, encouraged users to transition from being content consumers to being content providers, amassing a wealth of data, which in turn improved our web search services.

We were also one of the first companies to pioneer “comprehensive search,” which organizes search queries into distinct groups such as news, web documents, and shopping, among others. This approach, coupled with services like “Knowledge-base” that tailor search results to user interests in specific topics based on keywords, eventually led us to further enhance user experience with GenAI innovations. 

Can you tell us about how HyperCLOVA X is being assimilated across the company’s ecosystem? We’ve heard about your web navigation tool named Cue, NAVER Shopping reportedly offering GenAI support, NAVER Map providing real-time traffic info through phone cameras, and NAVER Pay becoming a GenAI-powered payment system. Can you elaborate on these integrations?

Being our flagship LLM, HyperCLOVA X is tailored to the Korean language and culture. Since its release, it’s been evaluated by Korean systems like KMMLU (Measuring Massive Multitask Language Understanding in Korea) and KorNAT (Korean National Alignment Test), showing high utility for Korean users. As of today, over 2,000 startups have created their own AI solutions by using HyperCLOVA X as a foundation model.

It’s true that HyperCLOVA X has already been integrated into several NAVER services to elevate user experience. When it comes to GenAI-enabled web search, the idea is to offer comprehensive semantic processing rather than simply indexing web pages. This is precisely how our GenAI search service Cue: leverages HyperCLOVA X to process complex queries. More specifically, it employs multi-step reasoning to discern users’ intent, thereby improving search accuracy. This approach also deepens user understanding behind generated answers.

One of the most end-user-relevant solutions that relies on our LLM is CLOVA X, NAVER’s chatbot capable of performing universal tasks such as writing, coding, and logical reasoning. This tool enhances usability through integration with various external platforms, including SOCAR (a car-sharing service), Kurly (an online shopping platform), and TRIPLE (a travel service). We also offer SmartBlock — a GenAI feature that organizes documents into topic-based groups.

Naturally, HyperCLOVA X is being utilized actively to provide personalized content recommendations to users in the NAVER App. We’re also currently developing a new GenAI advertising service, CLOVA for Ad, in partnership with the global sports brand Nike. 

Some analysts argue that tech companies are actually using AI within the B2C sector as a sort of test tube, with the end goal of transitioning to the B2B sector, ultimately selling to other companies rather than individuals. Amazon is often cited as the most illustrative example of this, and increasingly Google is as well. What’s your assessment of this claim? 

Certainly, NAVER harnesses its extensive B2C track record to offer customized AI solutions to companies through cloud services, addressing business needs and providing a competitive edge. Having said that, HyperCLOVA X was developed primarily to enhance service experience for Korean users, reflecting linguistic and cultural nuances. This focus on localized solutions has been integral to NAVER’s strategy since the company’s inception, when we developed our first search engine. 

So, the idea that tech companies are using the B2C sector merely as a “test tube” for future B2B applications certainly does not align with our own approach. At NAVER, B2C services are intricately woven into the everyday lives of Korean users, highlighting our commitment to refining these services through AI rather than using them as a stepping stone for B2B expansion, which we treat as an independent endeavor.

On the opposite side of the spectrum, some industry insiders, including Sam Altman of OpenAI, insist that the future of AI is heading towards more domain-dependent AI tools based on small language models (SLMs). What are your thoughts on that?

The trend towards smaller, more efficient models is indeed notable in the hyperscale AI domain. In April, we launched DASH, a variant of our HyperCLOVA X model, which reduces size, boosts speed, and optimizes expenses, while maintaining solid performance. Still, despite this shift towards more condensed models, the principle of scaling laws remains a significant factor. By default, larger models, equipped with adequate data and compute resources, can develop more diverse and sophisticated abilities necessary for complex tasks beyond mere text translation or summarization.

To this end, we’re set on diversifying the AI market by offering a range of models. HCX-DASH caters to those that need basic functions, performance speed, and cost-efficiency. At the same time, our other models tackle more complex functionalities and include multimodal capabilities featuring image, audio, and video. This crucial model variety ensures that different users have the best tools for their individual requirements.

Web search is evolving alongside GenAI. As you already pointed out, AI-enabled chatbots offer end-users more targeted AI summaries rather than simple indexing of web results (e.g., Perplexity, You.com, Brave, etc).

While this offers a great deal in terms of convenience and speed, many critics argue that there are also significant risks. End-users may receive one-sided results or even fabricated answers (i.e., hallucinations). Additionally, it could substantially reduce website traffic for businesses. Do you consider these legitimate concerns? 

Yes, current AI models often lack diversity in user choices because they generate direct answers to queries, which might influence user perceptions. This issue ties closely to the risk of market dominance by AI models developed by a select few tech companies, particularly those trained primarily on data from English-speaking countries. 

To preserve cultural integrity and ensure diversity and accuracy, it’s essential to support an assortment of AI models that respect and represent the varied perspectives of different ethnic groups, national identities, and religions. We believe that HyperCLOVA X is contributing to this important objective. At NAVER, we recognize these concerns as legitimate and are taking active steps to ensure AI safety, including operating dedicated red teams and engaging in ongoing research to mitigate biases within our AI models. 

Regarding the impact of AI on website traffic, opinions vary, and it’s a topic that requires further investigation. In practice though, the convenience offered by GenAI appears to be offsetting any negative effects, leading to its increased adoption worldwide. We believe the key here is to develop the right “non-zero-sum” approach for GenAI integration, where both AI companies and content creators actively pursue win-win strategies.

What are your impressions and takeaways from the recently held AI Seoul Summit?

The most significant insight from the AI Seoul Summit was the expanded concept of AI safety. Previously focused on the more technical aspects like controlling hallucinations and resisting jailbreaking, this year’s summit emphasized the importance of having a responsible and diverse AI ecosystem that respects regional values and local cultural heritages.

Notably, this isn’t just about training AI models in a conscientious manner or fine-tuning them to avoid inherent objectivity issues and performance biases linked to underfitting and overfitting. It’s also about the need for an increased number of AI players and models within the global AI landscape as a way of ensuring a more sustainable and enduring AI environment.