to leave a comment.
On the 22nd, Bryan Catanzaro, Vice President of Applied Research at NVIDIA, emphasized the importance of the open-source ecosystem, stating, "It is crucial to support researchers in effectively utilizing artificial intelligence (AI) tools."
During a special lecture hosted by the Seoul National University AI Institute that afternoon, VP Catanzaro stated, "AI research is rapidly evolving through the combination of open models, data, and high-performance computing infrastructure." Approximately 300 people, including professors and students, attended the event.
In particular, VP Catanzaro unveiled key achievements of Nemotron, emphasizing the importance of NVIDIA's open-source ecosystem.
Nemotron is NVIDIA's open-source AI model, encompassing datasets, training methods, and software.
VP Catanzaro defined AI as a kit consisting of five layers: energy, chips, infrastructure, models, and apps.
He explained, "To build excellent AI, energy is fundamental, and on top of that, chips and data center infrastructure are built. Especially in the model layer, open AI model technology that allows companies to customize while maintaining their data sovereignty and unique platforms is essential."
VP Catanzaro diagnosed that the AI expansion paradigm is evolving in four dimensions.
According to VP Catanzaro, AI is expanding beyond the pre-training phase to include post-training, which learns interaction with humans; inference computation, which applies thought processes during inference; and agent systems, which use tools autonomously.
During the lecture, VP Catanzaro introduced in detail the technologies applied to the Nemotron-3 model.
The Hybrid-SSM-Transformer architecture applied to Nemotron-3 demonstrated higher accuracy and efficiency than existing Transformer models.
A Transformer model refers to a neural network that learns context and meaning by tracking relationships within sequential data, such as words in a sentence.
Techniques such as MoE (Mixture of Experts) to compress tokens into smaller spaces to reduce data communication costs, and MTP (Multi-Token Prediction) to speed up inference by predicting multiple tokens, were also introduced.
In addition, NVIDIA's strategy for the Korean market was shared during the lecture.
The Nemotron Persona Korea dataset contains approximately 7 million virtual persona information to enable the development of AI that accurately understands the regional context of Korea.
The dataset enhances the specificity of AI models by reflecting the characteristics of real Koreans without including personal information.
On this day, VP Catanzaro also offered advice to researchers facing difficulties in securing computing resources.
He added, "Focus on new ideas and theoretical foundations rather than industrial-scale production competition. AI still has many unsolved problems, and creative ideas from academia will be an opportunity to change the technological methods of industry."
He further emphasized, "NVIDIA plans to continue collaborating with companies worldwide through Nemotron and transparently disclose technological achievements."
Seoul National University plans to use this event as an opportunity to promote convergence research and expand AI talent development in cooperation with global AI leading companies, including NVIDIA.
Newsletter
Get key news delivered to your email every morning
to leave a comment.