The race for artificial intelligence It is the new diamond of the economy of many countries. one to whom they are throwing money as if the world were going to end and that it is having serious implications on issues that affect citizens such as energyhe employment and with one last controversy: the exorbitant price of RAM. The great powers they want to be sovereign in this field, and South Korea has just light his first hyperscale artificial intelligence model.
His name could be some son of Elon Musk: AX K1.
In short. Developed by the giant SK Telecom, AX K1 is a model that has 519 billion total parameters, although during inference, which is the practical use case, it “only” activates about 33 billion. It’s still accurate (as accurate as an AI can be) but consumes far fewer resources. That 519B – A33B mode is based on the ‘architecture’mixture of experts‘ that selects in real time and dynamically the optimal parameter subsets for each task.
These parameters are like the neural connections that allow the model to “learn” during training, and the fact that South Korea already has a hyperscale model is a huge leap in the country’s position within the global picture of AI.
Master Model. The design of this model allows stable performance in tasks such as advanced reasoning, mathematics and multilingual comprehension, but there is also an interesting concept: it works as a “Master Model”. These models are the ones that transfer knowledge to smaller models. While the master knows everything, the lighter model is specialized in a specific task.
And, although the large model consumes an enormous amount of resources, the “student” that inherits complex capabilities without having to manage so many parameters can run on devices and environments with more limited resources. For example, the AX K1 with those 512B can “transfer its knowledge” to those below the 70B scale, much more specialized and cheaper.
“As Korea’s leading AI company, we will continue to push forward our efforts to deliver AI for everyone” – Tae Yoon Kim
“AI for everyone”. In less words: the master model allows the expansion of AI to be accelerated because the hyperscale is used for research, but the lower scale is used for more everyday products. And, precisely, that is what SK Telecom seeks: for its IOA to be the basis on which the country operates. In collaboration with different universities, associations and thanks to the memory manufacturer SK Hynix –one of the giants of the sector and part of SK Telecom-, the company hopes it will be the foundation of an “AI for all.”
This implies that they will deploy it in their services and, as it is open source, its API can be the basis of other models in university, business and even national ecosystems. In fact, there is already talk of very specific solutions, such as access to AI through text messages and even phone calls, but also multilingual search services and even a boost for AI in video games. And, of course, for humanoid robotics either for education.
The great advantage that the consortium that owns AX K1 has is that it is one of the largest groups in the world, with a presence in the semiconductor, telephone, transportation, construction, energy and video game industries. Therefore, you can easily scale this technology.
Third in contention. SK Telecom has confirmed that it plans to continue expanding its model with agent-based execution and those 519Bs allow Korea to become “one of the top three artificial intelligence nations in the world,” in the words of Tae Yoon Kimone of those responsible for the model.
The group’s intention is to help “consolidate South Korea as one of the world’s top three artificial intelligence nations,” a race that is taking place resources difficult to contextualize in both the United States and China and which is crushing markets like RAM for consumers.
Image | SK Telecom
In Xataka | The exorbitant deployment of data centers for AI has a new problem: salt caverns
GIPHY App Key not set. Please check settings