NTT used its NTT R&D Forum conference and demonstration showcase this year in Tokyo to clarify the company’s vision around All Photonics Network (APN) technology.
APN is one of three cores that form its Innovative Optical & Wireless Network (IOWN) concept – the other two being its cognitive foundation approach (described as the optimal harmonisation of all ICT resources) and its work in digital twin computing.
Speaking to press and analysts (in Japanese) after the event’s main stage keynote, Shingo Kinoshita, senior VP & head of research and development planning for NTT explained some of the mechanics behind the technologies in motion here.
When asked what kind of technology is used to make up the optical engine, Kinoshita said that the chip in the centre of this technology is an ordinary LSI (large scale integration) chip. It contains an element that converts the optical fibre and the electrical wiring by light and electricity – and, importantly, it achieves very high speed and low power consumption.
NTT has taken its work in Large Language Model (LLM) technologies for generative-AI applications in a similar direction to other vendors pushing the so-called Small Language Model (or private AI) approach and built its own gen-AI model known as tsuzumi (a name for a traditional Japanese drum), which is currently in its extended development phase – just how tough was that to create?
“When we talk about AI collaboration, language model collaboration depends on whether the language is linked with its own elements as a natural language, or a vector. As of now, natural language is easier to interact with LLMs, so I think we’ll start from there. However, considering the speed and efficiency, there is a possibility that AI’s own protocol would be better, so we will explore that option,” said Kinoshita.
He thinks that, at this stage, it has been quite difficult to find out which LLM interacts with what, what kind of knowledge it has and what kind of characteristics it exhibits. He reminded the attending press that there are various protocols at play here, including ones that govern how AI can create a community and interact itself.
“I think various approaches are necessary in this area, including not only technology but also social science,” said Kinoshita-san.
When asked about NTT’s prospects for doubling the distance between datacentres to expand its work in this field, he said that the ability to double the distance between datacentres is pretty straightforward today. Because of the low latency characteristics of IOWN, this technology could be very useful in future, soon.
NTT & LLM technologies
Thinking about NTT;s work with LLM technologies, Kinoshita notes that – yes – it starts with the Japanese language. But he insists, the NTT Group is also very global, so the team would like to expand not only in Japanese but also in English, Chinese, Korean and across languages used throughout Europe.
Talking about the potential application of the lightweight tsuzumi LLM, Kinoshita suggested that in particular, smartphones and car navigation systems need to be lightweight and respond quickly – so this area could suit real world deployments.
Questioned on benchmarks, he was challenged over the fact that there are some things that can’t be measured by only one benchmark – so where does he stand on testing and rating work in this field.
“RakuDA is a benchmark that is often used. For GPT4, RakuDA asks test questions about history, geography and other topics that are fairly similar to Japanese textbooks. It performs a wider range of things, such as coding. So then, tsuzumi actually loses to GPT3 at the coding level for the programming part, we made Japanese a priority first, so we lost a little, but I think we can win by strengthening it,” explained Kinoshita candidly and openly.
Kinoshita’s presentation and Q&A session followed the more corporately packaged keynote delivered by Akira Shimada, NTT representative director, president and CEO. With the (chic) geeks in the room (probably, arguably) finding Kinoshita’s session more illuminating, the pair work as a suitable double act and (arguably) provide both the big picture and the deeper picture in what feels like a logical flow.
What next comes next from NTT?
The NTT developer network, open source foundation and expanded API selection pack?
We don’t know, but we hope so.