Nvidia's 'Nemotron Day' Kicks Off... "Building an AI Agent Ecosystem" (Comprehensive)

Nvidia's 'Nemotron Day' Kicks Off... "Building an AI Agent Ecosystem" (Comprehensive)

Full-cycle technology demonstration from synthetic data and reinforcement learning to GPUs

Emphasis on balancing enterprise AI security and utilization with open shell-based control technology

Nvidia, together with the National IT Industry Promotion Agency (NIPA), will host 'Nemotron Developer Day' from April 21st to 24th.

This event was organized to share the research achievements of Nemotron, Nvidia's latest open-source based AI model.

At the event held at Front One in Mapo, Seoul, researchers directly introduced synthetic data generation strategies, reinforcement learning know-how for model optimization, and allowed attendees to experience full-stack AI technology covering everything from Graphics Processing Units (GPUs) to learning techniques.

First, Nemotron is Nvidia's open-source AI model, encompassing datasets, learning techniques, and software.

Nemotron is available in various sizes, from Nemotron-3 Nano with 3 billion parameters to the high-performance model Nemotron-3 Super with 120 billion parameters.

Bryan Catanzaro, Vice President at Nvidia, explained, "Nemotron adopts a hybrid architecture to improve memory utilization and enhances inference efficiency with multi-token prediction technology."

Nvidia also introduced the 'Nemotron Persona Korea' dataset, which can correct distorted Korean datasets in global AI models.

Kim Hyun-woo, a researcher at Nvidia, pointed out, "Existing global large language models (LLMs) showed a distorted perception, for example, by stating that 90% of Koreans are nursing care workers or 77% are farmers."

To resolve these errors, Nvidia created a persona probability graph model consisting of 42 nodes and 75 edges, utilizing 62 national statistics and public data from sources such as Statistics Korea, the Supreme Court, and the National Health Insurance Service.

This model includes 1.7 billion tokens, comprising 1 million records and 7 million personas.

Researcher Kim explained, "Utilizing this dataset for AI training helps AI models understand Korean culture and context more accurately and is a great help in generating logical reasoning problems."

This dataset is publicly available on Hugging Face for anyone to use for commercial purposes.

Anthropic's OpenClo provides high freedom to AI models, increasing productivity, but also raises concerns about potential security risks.

Here, Nvidia's Open Shell acts as a kind of protective shield that restricts the activity range of AI agents within specific policies.

For example, an AI model can generate a summary of a received email, but it is restricted from replying to that email arbitrarily.

In this context, Nvidia Nemoclo acts as a platform that supports the safe use of OpenClo in an enterprise environment by utilizing Open Shell.

Lee Jin-soo, Director General of AI Policy Planning, stated, "This Developer Day is part of the global technology sharing and talent development cooperation (MOU) that the government promised with Nvidia last October," adding, "It will be an important opportunity for our AI developers to acquire global technological trends and internalize the capabilities to practically utilize them."