Our goal is to enhance the intelligence of generative AI models through extensive research in large language models (LLMs) and code models. Our commitment includes building reliable models and services using technologies such as reinforcement learning, Retrieval-Augmented Generation (RAG), prompt engineering, and efficient model serving. We tackle fundamental challenges such as integrating learning and perception with high-level knowledge. This approach aims to enhance reasoning capabilities by incorporating the user’s context and domain knowledge. This research not only improves our product competitiveness but also boosts internal productivity .

We develop technologies that provide real-time multilingual translation in various apps such as translation tools, browsers, and Notes within smartphones. Additionally, typical use cases include converting the style of sentences when users write, recommending spelling/grammar, and suggesting appropriate response content.

Our development of voice and sound AI technology aims to achieve a human-like understanding of speech and sounds, facilitating natural conversations and expressions. We are also enhancing listening experiences through advanced analysis and understanding. Our core technologies is the E2E Speech LLM solution, which includes voice-related solutions such as voice recognition, synthesis, and processing, as well as sound recognition, source separation, quality improvement, audio search, and voice biomarker-based Brain health. These technologies enable Samsung devices such as smartphones, home appliances and wearables to offer tailored intelligent services in various settings.

In the case of videos in the smartphone gallery, there are cases where background noise is included, so we develop technologies that can reduce or amplify it. This can be used to search for target audio sources and separate them, like an audio eraser."

As more Samsung devices come equipped with cameras, there is a significant opportunity to transform these from smart to AI-driven devices with the support of vision AI. We aim to provide a comprehensive and on-device vision solutions for Samsung devices, ranging from low-level camera sensor processing to high-level visual intelligence for smartphone and TV.

Photos usually contain various text information, and the technology to recognize this text information has many applications. For example, in the case of a prescription photo, it contains information such as drug names, drugs, and dosages, so it can be registered in an app that automatically manages them. Another example is when signs/menus/newspapers/books contain text, it can perform useful functions such as copying & pasting these texts or searching and translating them. In addition, we develop generative AI technologies such as removing unnecessary objects from photos in the smartphone gallery, and generating wallpaper images that dynamically changes according to the weather and time.

We are developing a generative AI model named Samsung Gauss. This model is inspired by Carl Friedrich Gauss, a renowned mathematician known for formulating the theory of normal distribution, which is crucial in fields like machine learning and AI. Samsung Gauss embodies our aim to harnessing the power of AI to significantly improve consumer lives worldwide.

Samsung Gauss2 offers enhanced performance and efficiency by handling various data types simultaneously as a multimodal model that integrates language, code, and images. It is available in three distinct models designed for different purposes: Compact, Balanced, and Supreme.

• Compact Model: This small-sized model is specifically designed for efficient operation in limited computing environments, maximizing each device's unique capabilities for optimized on-device performance.
• Balanced Model: This model achieves a balance between stability and efficiency, providing consistent performance across a variety of tasks.
• Supreme Model: Aiming for top-tier performance, this model utilizes a new model architecture built upon the Balanced model, enabling significant reductions in computational requirements during training and inference while maintaining high levels of both performance and efficiency.

Samsung Gauss2 supports up to 14 languages as well as various programming languages, depending on the model. To ensure maximum efficiency for these supported languages, Samsung has developed and incorporated its own stabilization techniques for training large language models (LLMs) and designed a custom Tokenizer.

Looking ahead, Samsung aims to continually enhance the performance of its code.i service through Samsung Gauss2, improve the natural language question-and-answer functionality of the Samsung Gauss Portal, and support multimodal features such as understanding tables, charts, and creating images. Each of these enhancements is intended to boost productivity within the company.