AT&T CTO Jeremy Legg this week revealed some of the key lessons garnered as an early mover in the generative AI (genAI) space, including how the technology can best be used and which large language models (LLMs) are suited for specific tasks.
Speaking at an industry event, Legg said AT&T has more than 600 AI and ML models in production, with more than 60 new genAI use cases earmarked for 2025.
By leveraging the billions of dollars of investments companies are making in LLMs and chips from companies including Nvidia, AT&T aims to “become the best practitioner of AI in our industry,” Legg said.
Tips
He said there are several knowledge points AT&T learned about genAI.
The first is each company should host and control their own data.
While LLMs from companies including Facebook, OpenAI and Microsoft are trained on information available on the public internet, AT&T’s network and customer data is kept behind a walled garden.
By training the LLMs on the corpus of information that sits on the internet, Legg says those companies democratise large amounts of information across many industries.
“But what I would also argue, what it’s done is it has commoditised that information so that everyone has the exact same set of access.”
“What we are in the midst of doing is actually embedding that information into these LLMs in a proprietary way so that it does not go out into the broader internet.”
By keeping the data within the company, AT&T also avoids the risk of it falling into the wrong hands.
He explained genAI users should be aware not every model is equal: some are better at mathematics, others excel at image interpretation and others text conversion.
“You have to know how to score the models and know how accurate the models are so that you use the right model for the right task.”
In house nous
Another lesson is being an expert on your own data.
GenAI users must understand how to ask the right questions to get the information back in an accurate way, Legg explained.
“The way you ask a question is incredibly important to the way that the response actually comes back.”
“All of these models are parameters that have weightings on them. How you weight a parameter will drive the response that you get back. Are you putting the right weighting on it? These models aren’t plug and play.”
Among its genAI uses cases, AT&T is using it to assist in thousands of fibre installations.
It helps with scheduling technicians’ visits and optimises their journeys to minimise mileage, in turn increasing the number of homes which can be serviced each day.
AT&T plans to embed the same genAI capabilities into its network to identify and fix problems before consumers are affected.
“The things that we talk about internally around genAI and how we use genAI, it’s around penetration rates. It’s around ARPU. It’s around lifetime value.”
