The last four sessions in the AI/ML track at Upgrade 2024 included a discussion of a unique application of AI in Australia, NTT’s own large language model (LLM) Tsuzumi, a strategic forecast for GenAI, and the relationship between AI and an NTT edge computing platform.
In the fourth session of this track, NTT SMART World Solutions VP Bennet Indart described how NTT DATA is working with Australian charity Climate Force to help regenerate a rainforest ecosystem on a former banana plantation. “They’re using all kinds of really interesting techniques,” Indart said. “They’re geo-tagging every tree that they plant. They are looking at drone footage, they’re using lidar, soil sensors, water sensors.” And in a three-year project that began in January 2024, they started using the NTT Smart Data platform to make more sense of the information they’re harvesting. Indart said the platform first acts as a “giant vacuum” for the data. “Then we use AI. We use some common data architectures that our team has put together to help to not only normalize the data but also make those correlations.”
The ultimate goal is to create a digital twin and accelerate growth and regrowth in the 527-acre tract. The outcome will not only support the rainforest, with its biodiversity, outsized contributions to oxygen production, sources of potential medical treatments, and other benefits. The lessons learned from this exercise could also serve as blueprints for other “upgrades” in agriculture, city planning, and more.
Alongside the meteoric rise in LLMs like ChatGPT is corresponding need for tremendous computing power, even dedicated data centers. NTT’s proprietary and lightweight LLM, Tsuzumi, offers a countervailing trend. “We have implemented the concept of AI constellation for sustainable society, shifting away from the heavy reliance on large monolithic models that require significant computing resources,” NTT Human Informatics Laboratories Senior Distinguished Researcher Kyosuke Nishida said.
The constellation approach involves using NTT’s optically driven IOWN network to tap into the collective intelligence of smaller and specialized AI. Compared to the 175 billion parameters of ChatGPT 3.5, for instance, Tsuzumi and Tsuzumi ultra-light use 7 billion and 0.6 billion parameters, respectively. Both versions can operate on a single GPU, and the smaller version can run on CPUs. Lower bit quantization of parameters also enhances speed and reduces memory load. Tsuzumi’s efficient tokenization, the process of breaking down data for easier pattern detection, is another reason for its speed. Tsuzumi is proficient in Japanese and English, and it outperforms other Japanese models and GPT 3.5 on several metrics, including its specificity and detail of answers. NTT has enabled Tsuzumi to understand visual content and has an even more ambitious long-range goal. “We aim to develop a software robot that can work with humans as a collaborator on the computer, and a robot that grows with people as a life partner,” Nishida said.
If AI seems transformational today, just look ahead. NTT DATA Director of the Global GenAI Strategy Team Sandra Trullemans said GenAI will evolve from a narrow to a general model, from requiring training for a certain task today to training itself by just observing the world, which could happen sometime over the next ten years. (Beyond that is “super AI.”) Navigating the transition to a general model will require attending to the core NTT values of security, trust, and sustainability, as well as adapting to the perceived loss of control and accountability that accompanies more powerful iterations of AI.
Trullemans said NTT’s multi-layered approach addresses these concerns. It includes an infrastructure layer, which contains the LLMs, standard Microsoft software, open AI databases, etc., where most proof of concepts and use cases are built today. Above that is the NTT enterprise platform that can supplement what may be lacking in an enterprise’s own infrastructure. “For our clients, they just need to care about their application, and we take care of all the security, trust and sustainability,” she said. The platform’s extensibility derives from several factors, including NTT R&D AI Security, the ABE framework originating from NTT Research, Tsuzumi for its lightweight and sustainable attributes, and an ecosystem of smart agent partner solutions. In addition to providing the right technology fit for enterprise clients, NTT DATA also addresses their organizational transformation needs and specific use cases that they are looking to solve, whether helping them to gain margin, or optimize processes in a manufacturing or production line or across different companies, or elsewhere.
In the final session of this track, NTT DATA Network Innovations CEO Ichiro Fukuda first asked what edge computing (his topic) had to do with AI. He answered by pointing to the NTT Research MEI Lab concept video of the autonomous closed-loop intervention system (ASIS) shown in the opening general session, saying it was a good example of the connection between the two. But making the edge work is difficult. “There’s a diversity of environment hardware,” he said. “How do you manage the procurement and logistics? How do you install the configured device? How do you secure the network? How do you update the patches?” The NTT SPEKTRA edge platform was designed to overcome these edge digitization hurdles, enabling edge service delivery while also linking to NTT’s global and private connectivity.
Fukuda provided a few use cases of what the platform can do. It enables the distribution of sensors and the collection of data from various environments. It allows for the integration of sensors and protocols, making it flexible for IoT and edge computing applications, such as collecting network performance data or monitoring environmental conditions in data centers. As with the ASIS prototype, the platform can also support the deployment of AI models at the edge. It allows for the training and building of AI models in the cloud, which can then be shipped to the edge for local inference. Executing AI algorithms directly at the edge reduces the need for sending data to the cloud for processing.
The edge platform also supports the inventory management case study solution described in this session by NTT SVP Enterprise IoT Products and Services Devin Yaung and “Launch by NTT DATA” Managing Director William Radcliffe. Success entailed creating a minimum viable product (MPV) within five weeks, achieving a solution that had eluded this client for years. But it took both the right technical assets and a unified NTT team. “If it wasn’t for the edge, none of this would be possible,” Radcliffe said. “More importantly, if it wasn’t for us coming together collectively, we wouldn’t be here today.”