SK Telecom, A.X K3 design review and pre-learning experiment... Optimized GPU utilization

Even during the Chuseok holiday, SK Telecom [017670] is speeding up the development of its next-generation proprietary artificial intelligence (AI) model A.

According to the industry on the 25th, SKT's elite team recently reviewed the architecture of A.X K3 and began experimenting with pre-learning recipes that combine data composition and learning methods.

The plan is to focus on optimizing the learning environment and efficiently utilizing GPU resources and data based on the experience and assets accumulated during the previous model development process.

In particular, SKT is working to efficiently operate 1,000 government-provided NVIDIA B200 GPUs and its own infrastructure to be used in the third phase of development.

We are continuing technical cooperation with NVIDIA through weekly remote meetings. SKT explained that since the development of A.

We are also specifying ways to utilize the data accumulated during the development of A.X K2 and joint research with Seoul National University, KAIST, and consortium partner organizations. The proposed research project plans to pursue joint research in connection with the development of A.X K3 based on the expertise of each participating organization in the near future.

Meanwhile, SK Telecom is continuing verification and advancement work on the 'Everyone's AI' beta service scheduled to be released next month during the Chuseok holiday.

Aiming to become an AI assistant that performs everything from information search to follow-up actions such as applications and reservations, we are developing the service so that it can be used by phone or text. We also operate an 'AI grader' that evaluates whether the service accurately understands the intent of the question and performs the requested task, and repeats verification and improvement.