August 26, 2024

NTT's Large Language Model "tsuzumi":A High-performance and Low-energy-consumption Large Language Model with Expertise in Specific Fields

Since ChatGPT, OpenAI's generative artificial intelligence (AI) chatbot, was released in November 2022, large language models (LLMs) have rapidly drawn worldwide attention, and new LLMs have been announced one after another. While the "bright side" of such LLMs, namely, high performance, has been brought to attention, their "dark side," namely, increased energy consumption, has also been pointed out. On November 1, 2023, NTT's LLM called "tsuzumi" was announced. It was developed with a particular vision in mind, creating a future in which many and small LLMs having different characteristics are combined, rather than creating one huge and monolithic LLM, to achieve social well-being through AI-human collaboration. Thus, tsuzumi has made it possible to achieve the bright side and address the dark side of LLMs simultaneously, and its commercial service was launched on March 25, 2024. We interviewed Kyosuke Nishida, a senior distinguished researcher at NTT Human Informatics Laboratories, about how the tsuzumi project was launched, features of tsuzumi, trends in LLMs, and his thoughts on the challenges of competing in a highly competitive market.

August 26, 2024

NTT's LLM "tsuzumi"

In November 2023, NTT announced tsuzumi, a large language model (LLM) based on NTT laboratories' over-40 years of accumulated research in natural language processing. The tsuzumi LLM excels in Japanese-language processing ability, and its key features include being lightweight and providing multimodal support, which enables handling of non-text media. Its lightweight model reduces power consumption and makes on-premises use possible. Its multimodal feature enables comprehension of photos and diagrams, distinguishing tsuzumi from other publicly available LLMs. In this article, we give an overview of tsuzumi on the basis of these key features.

August 26, 2024

Machine Learning Using an On-chip Photonic Neural Network--Changing the Future through Photonic Technology

Artificial intelligence (AI) is now entering its fourth wave, and this time, it is penetrating everyone's daily life and bringing many surprises. However, training an advanced AI model requires a huge amount of resources such as time in units of several months and power equivalent to all of the power consumed by a medium-sized city. To solve this problem, studies are being conducted on AI-model computing using optical circuits that are fundamentally different in principle from past technology. In this article, we talked with NTT Distinguished Researcher Mitsumasa Nakajima about the potential of future photonic technologies and his research on "machine learning using an on-chip photonic neural network" that can greatly reduce the power consumed by machine learning through the use of photonics.