Automation in Telecom Networks Improving, but 'Lazy' AI Is a Thing
Carriers have a long history of using statistical methods and machine learning when analyzing their networks, but generative AI means a “step function in capabilities,” Raj Savoor, AT&T vice president-network analytics and automation, said Wednesday during an RCR Wireless webinar. However, another speaker warned of “lazy” AI.
Sign up for a free preview to unlock the rest of this article
Timely, relevant coverage of court proceedings and agency rulings involving tariffs, classification, valuation, origin and antidumping and countervailing duties. Each day, Trade Law Daily subscribers receive a daily headline email, in-depth PDF edition and access to all relevant documents via our trade law source document library and website.
Carriers can pick from numerous models for anomaly detection but should be “very, very careful and ensure there are guardrails on picking” the right one, Savoor said. Networks are becoming more disaggregated and virtualized, and functions are “highly distributed.”
“There are a lot more configurations” and network changes are made constantly, Savoor said. “Each change process introduces the risk of potential impairments.”
Still, Gen-AI is a powerful tool for identifying patterns and developing insights about networks, he said, adding, there are “real opportunities to advance autonomous networks and drive further efficiency.”
One potential problem with AI is the tendency for it to become “lazy,” warned Stephen Douglas, head of market strategy at Spirent Communications, a telecom testing company. “It’s a behavior we’re starting to see today in some of the large language model implementations,” he said: The AI models either “slow down in terms of how long it takes them to provide the outputs, or they provide … abbreviated outputs where in the past they would have been giving sort of a more fully detailed answer.”
AI begins behaving as humans sometimes do when they’re tired, Douglas said. No one has figured out yet what causes lazy AI, though there are theories, he said. One possibility is that training data “is introducing human-like behaviors into the model.”
One implication is that carriers still can’t fully trust AI, Douglas said. “We’re starting to see these types of new novelties … the longer the large models are out there.”
Savoor said a top AI focus of AT&T is cutting energy costs, as it's costly to power a highly distributed network. For example, it's a challenge “ensuring the right balance between performance and energy efficiency.” At times, network use declines. With automation, carriers can better tune networks to adjust to usage patterns, he said.
With open and virtualized radio access networks, providers can tune the network to performance states, sleep states and “deep sleep state, where you can flush the cache, turn off the clock,” Savoor said.
AI has been a “game changer” in helping AT&T optimize coverage capacity and better adjust the network when cellsites go down, Savoor said. AI means increases in intelligence at the edge of the network, he said.
AI accuracy is increasing, Douglas said. “At the end of the day,” accuracy is still based on the data that’s fed into the model. Sometimes small-language models are just as powerful as large-language models, depending on what an operator is trying to do, he said.
One growing use of AI comes through feeding all the configurations and other details about network gear into an AI system, Douglas said. When a problem occurs, a field engineer can quickly get answers about what’s happening, he said. “Every operator should be doing that -- it’s a bit of a no-brainer.”
Another use, which started in Canada, is tying AI into drones monitoring network facilities, Douglas said. AI can find problems the human eye would never see, he said: “It could be hairline cracks on an antenna. It could be an issue with some of the cabling you just could not pick up.”