Telcos seek best ways to adopt generative AI technology

During a TM Forum press briefing in London, Vodafone's CTO Scott Petty and Verizon's global chief digital and information officer Shankar Arumugavelu addressed the question of whether telcos should be building their own large language models (LLMs), with Petty saying it's not something they should be involved in and that they should instead use the models available.

He added that a better approach is to select the best available model for each use case. Petty argued that each LLM has its own set of capabilities that are more suited for some tasks than others. This approach would also prevent vendor lock-in, which is something telcos are not keen on.

Arumugavelu, meanwhile, gave a more nuanced answer. Admitting that a year ago, he would have been opposed, he said the company's position has evolved. He noted that the idea is to adopt the foundation models available and then customize or build on them over time. He identified four approaches. 

The first of these archetypes, as Arumugavelu calls them, consists of taking a model as is and adding some prompt engineering. In archetype two, you add some context with company proprietary data that doesn't leave the enterprise, where the company gets the best of both worlds, he says. Archetype three involves some fine-tuning of the underlying model with the right dataset to make the model perform better at a specific task.

Small, expert models

The fourth archetype, meanwhile, involves pre-training with proprietary data added to the model to fine-tune it to a specific use case. An important aspect of this archetype is ensuring the data is not fed to the base model and stays within the organization, in what is called a frozen model, Arumugavelu says.

Fully building a model from scratch, however, requires massive amounts of data and costs hundreds of millions of dollars. Arumugavelu argues that this is not necessary for a telco's uses. Instead, he proposed building several small, highly specialized language models that he likened to a group of experts. Verizon is experimenting with this approach in 2024, he said, adding that small language models currently represent an opportunity.

Petty, meanwhile, pointed to the transformational effect generative AI could have in the long run, saying it holds the potential to change how people interface with the Internet. While most people today use search bars, he argued that "generative AI is going to change that. It will be voice interfaces, it will be graphic interfaces."

Meanwhile, the work on integrating AI into networks – whether generative or predictive – is still underway and will require collaboration. Asked by TM Forum's CTO Nik Willetts about the best ways telcos can work together on this task, Arumugavelu said work is underway at TM Forum to create blueprints on AI adoption for telcos, which could help them scale up AI use.