MWC24 reveals a kaleidoscope of AI applications

AI was so much in evidence everywhere at this year’s MWC that you could have been forgiven for thinking it had become the AI World Congress. As unsettling as it might sound, AI was omnipresent!

But beneath the pervasiveness of ‘AI Everywhere,’ were multiple examples of its use diversifying into granular sub-categories of networking, service operations, and customer experience. For example, AI-based network solutions on show seemed to fall into somewhat separate but related threads around fault detection, prediction and resolution, network optimization, and network planning/upgrading.

In that sense, MWC served as both as a demonstration of the hype-cycle and something of a reality-check. Amidst the over-excitement about AI being a universal panacea for the industry were real-life deployments, working service models, practical solutions, and paths to continued future evolution.

On-device AI - and hybrid architectures

One prominent theme amongst silicon and device vendors was on-device AI, touted as a means for enhancing performance and security whilst reducing the costs of AI processing—driven by the increasing size of AI models, the cost of running these models on cloud resources, and the increasing use of AI across devices and applications.

In particular, dedicated on-device capabilities for generative AI are becoming a major focus of the smartphone market. For example, MediaTek announced that its system-on-a-chip offerings are now optimized to support Google's new Gemini Nano large language model and Meta Llama 2 applications, without cloud connectivity. This points to a greater share of AI tasks being managed on-device or through hybrid models rather than depending entirely on the cloud.

Another common topic at MWC, certainly amongst silicon vendors and hyperscalers, was hybrid AI architecture. Just as traditional computing evolved from mainframes and thin clients to today’s mix of cloud and edge devices, AI processing must be distributed between the cloud and devices for AI to scale and become more powerful, efficient and highly optimized. Hybrid AI architecture distributes and coordinates AI workloads across the cloud and device, rather than processing in the cloud alone, thereby dispersing part of the burden to the edge.

Foundation models, such as general-purpose, large language models (LLMs) like GPT-4 and PaLM 2, have achieved unprecedented levels of language understanding and generation capabilities. Most of these models are quite large and are growing rapidly. Although the size of state-of-the-art models continues to grow rapidly, another trend from MWC is toward much smaller, focused models that still provide high-quality outputs. Running these small language models in the cloud will help reduce the cloud resources required.

Creating alliances

Radio access networks (RANs) are complex, leading vendors and operators to seek automated tools to improve performance and efficiency. To date, collaboration has been through mostly ad hoc, bilateral or small groups but MWC saw the launch of the AI-RAN Alliance, a forum created to unify disparate efforts. These include accelerating AI utilization in the RAN, improving spectral efficiency, integrating operational processes to optimize infrastructure usage, and generating AI-driven revenue opportunities by deploying AI services at the network edge. Founding members include Amazon Web Services (AWS), Arm, DeepSig, Ericsson, Microsoft, Nokia, Nvidia, Samsung, SoftBank and T-Mobile. AI is as inevitable in the radio network as everywhere else, but whilst it can help manage complexity, it can increase cost and energy consumption. The AI-RAN Alliance has a task on its hands to find the right balance of applicability, affordability and sustainability.

MWC—always a popular launch-pad for new alliances—also saw the launch of the Global Telco AI Alliance (GTAA). The joint venture of Deutsche Telekom, e& Group, Singtel, SoftBank and SK Telecom plans to develop multilingual LLMs based on gen-AI models tailored to the needs of CSPs, helping them improve customer interactions via digital assistants and chatbots by making them more attuned to the telecom domain and better at gauging user intent.

The LLMs are currently being optimized, with training of the AI models already underway. Data from customer service chats were processed and anonymized for data protection compliance and is used to fine tune the model for telco-specific questions. (e.g. how do I reset my WiFi router?) The training is targeted to ensure the LLM understands the unique language and CSP customer scenarios with the aim of supporting personalized and more efficient CX. This pooling of resources between a group of non-competing CSPs is an interesting, if curious development, suggesting that ‘only a telco could know what a telco AI chatbot should know’ and begs the questions: will others join the GTAA, or will other similar CSP-based alliances form to co-develop AI solutions for telco-specific problems?

Meanwhile, the vendor community was predictably keen to support CSPs on their AI journeys, with services and solutions galore. For example, having announced its collaboration with the GSMA to accelerate telco AI adoption and skills with the launch of a generative AI training program and industry challenge in January, IBM discussed targeted services for the telecom sector served by watsonx, IBM's cloud-based generative AI and data platform.

ServiceNow announced an extension of its partnership with Nvidia to introduce a suite of generative AI tools tailored for CSPs, harnessing its Now Platform and Nvidia's AI software and hardware to assist customer service agents with advanced chat summarization,. As with other tools on show, the solution uses gen AI capabilities to workflows and processes, providing greater levels of contextual insight and content output for all roles interacting with CSP customers. Gen AI is changing the game for CX, but there are still lots of ways the game could play out.