Dispute highlights rising uncertainty over “from-scratch” requirements in Korea’s flagship AI initiative
SK Telecom has pushed again towards claims questioning the originality of its large-scale AI mannequin, A.X K1, as controversy surrounding South Korea’s government-led nationwide AI mannequin challenge continues to widen. The controversy, which initially targeted on rival groups from Naver and Upstage, has now expanded to incorporate SK Telecom’s elite workforce after allegations surfaced that A.X K1 bears similarities to China-based AI mannequin DeepSeek.
The nationwide AI mannequin challenge, overseen by the Ministry of Science and ICT, is a core authorities initiative aimed toward growing a basis mannequin distinctive to South Korea. SK Telecom submitted A.X K1 on the challenge’s first official briefing late final month, positioning the mannequin as a key pillar of its broader AI technique.
SK Telecom’s response to similarity claims
On Jan. 8, SK Telecom mentioned A.X K1 was developed utilizing an impartial structure and skilled solely from scratch. The corporate emphasised that the mannequin accommodates 519 billion parameters skilled from randomly initialized weights, arguing that each its scale and improvement course of distinguish it from present overseas fashions.
SK Telecom mentioned allegations of similarity stem from confusion between coaching structure and inference-related components. Based on the corporate, parts cited as resembling DeepSeek’s V3—equivalent to multi-head latent consideration (MLA) and combination of specialists (MoE)—seem solely in inference code used to run open fashions effectively, not within the coaching code that determines originality.
“From-scratch improvement refers to how a mannequin is designed and skilled,” the corporate mentioned, including that inference code is extensively reused throughout the AI {industry} and doesn’t undermine impartial improvement.
Technical debate displays industry-wide practices
Business observers word that architectures equivalent to MLA and MoE are more and more frequent throughout giant language fashions, making surface-level similarities troublesome to keep away from. AI engineers typically reuse standardized inference code to make sure compatibility and efficiency, whereas sustaining distinct coaching pipelines and parameter initialization.
The dispute has due to this fact shifted from whether or not SK Telecom copied a selected mannequin as to if Korea’s nationwide AI challenge has clearly outlined how far architectural overlap is suitable. A number of engineers argue that originality must be judged totally on coaching methodology, information, and weight initialization, moderately than on post-training deployment code.
Public reassessment by {industry} figures
The controversy eased considerably after Lee Seung-hyun, vp of AI agency FortyTwoMaru, publicly revised earlier remarks questioning A.X K1’s originality. In a social media put up, Lee apologized to SK Telecom’s improvement workforce, stating that the mannequin was skilled solely from scratch.
Lee acknowledged that whereas A.X K1 drew inspiration from methods equivalent to MLA and twin normalization, these had been optimized by SK Telecom’s personal scaling and engineering strategies, producing outcomes that went past easy reuse.
A part of a broader sample throughout the challenge
SK Telecom’s case follows related disputes involving different individuals within the nationwide AI mannequin challenge. Upstage beforehand confronted allegations that its Photo voltaic Open mannequin reused components from Zhipu AI’s GLM-4.5-Air, which it denied, saying solely inference code types had been referenced. Naver Cloud additionally got here beneath scrutiny for utilizing an exterior encoder from Alibaba’s Qwen mannequin, however mentioned the selection didn’t violate from-scratch ideas.
Collectively, the incidents have highlighted a recurring concern: the absence of detailed authorities tips on how “from scratch” must be interpreted for giant, fashionable AI techniques constructed inside world open-source ecosystems.
Strategic stakes for SK Telecom
Regardless of the controversy, SK Telecom has continued to advance its long-term AI plans. The corporate not too long ago consolidated its AI operations into a brand new in-house unit, AI CIC, and has mentioned it plans to take a position about 5 trillion received over the subsequent 5 years to broaden AI providers, platforms, information facilities, and proprietary basis fashions.
SK Telecom CEO Ryu Younger-sang has mentioned the reorganization is meant to make sure the corporate doesn’t miss what he described because the “golden time for AI,” and to place SK Telecom as a central participant in South Korea’s AI ecosystem.
Broader implications for nationwide AI coverage
The dispute over A.X K1 underscores a bigger problem dealing with the nationwide AI mannequin challenge: balancing the purpose of technological sovereignty with the realities of contemporary AI improvement, the place shared architectures and open-source parts are extensively used. As the federal government strikes ahead with evaluations, {industry} observers say clearer requirements will probably be wanted to stop related disputes and supply certainty to taking part groups.
For now, SK Telecom’s protection of A.X K1 has turn into a focus in a broader debate over how South Korea defines independence in AI—one that’s prone to form the course of nationwide AI coverage within the years forward.
Elevate your perspective with NextTech Information, the place innovation meets perception.
Uncover the newest breakthroughs, get unique updates, and join with a worldwide community of future-focused thinkers.
Unlock tomorrow’s developments in the present day: learn extra, subscribe to our publication, and turn into a part of the NextTech neighborhood at NextTech-news.com
