- AT&T is using custom network foundation models to move beyond traditional ML
- The models are already being used in production for energy efficiency and outage compensation
- AT&T is also making progress toward autonomous networks, with things moving faster on wireless than wireline
You’ve heard of large and small language models and, if you’re really into AI, probably world foundation models for physical AI. But what about network foundation models?
That’s exactly what AT&T has focused a good portion of its AI efforts building, AT&T VP of Network Analytics and Automation Raj Savoor said during Fierce’s virtual event entitled "AI and the Automated Network."
Savoor explained that network foundation models are AI models “trained on our own network data and configurations and KPIS and all the time series and event data.” He added these are 10 billion parameter models trained on 110 billion tokens.
Already, AT&T has been able to apply its network foundation models to boost energy efficiency and compensate for site outages.
“We were previously applying very static rules to those [cell sites] with traditional ML models, studying patterns and applying them,” Savoor said. “With network foundation models training on a very regular cadence, we are able to take advantage of much smaller intervals, more dynamically to go drive energy efficiency.”
In addition to delivering a “significant” improvement in efficiency compared to its previous ML models, Savoor said the new network foundation model approach has allowed it to spot “patterns we could not have seen with just the ML models and classic regression analysis.”
Another major use case for AT&T’s network foundation models is managing antenna tilts during outages – to achieve what AT&T calls “outage compensation.”
In a nutshell, the model allows AT&T to massively adjust its antenna tilts with capacity awareness to achieve both coverage and capacity optimization in real time. “This has been really, really useful because you just couldn’t do that at a human scale,” Savoor said.
Other use cases include mapping alarms and tickets and predicting failures and software congestion based on historical events.
These are among about five large use cases that AT&T currently has in production, though Savoor noted it is “working on a lot more of these network foundation models.”
The road to autonomy
AI is widely viewed as a key tool in an operators’ journey to increasingly autonomous networks. But Savoor noted that not all networks are created equal. Wireless networks, he said, have somewhat of an advantage because there’s a refresh cycle every 10-12 years with the introduction of new 3GPP mobile standards (i.e. 3G, 4G, 5G, 6G).
“The wireless domain definitely has a lot of the investment and line of sight to autonomous Level 4,” he said when asked how long it will be until operators achieve fully autonomous networks. “We have some processes at the sub-domain level already at 4 and many in the mid-3s approaching that. So, that is very achievable.”
But on the wireline side, the road to autonomy is more complex, Savoor said. That’s because the wireline network includes "hundreds of element types” and achieving end-to-end domain automation requires “some degree of homogeneity at the control plane layer.”
As a result, AT&T is being a bit more opportunistic in the access and in certain parts of the core network.
“We are doing some large core wireline transformations, and as I look at the vector and the target state…over the next three years I think you’ll have some significant opportunities end-to-end on the wireline,” he concluded.
To hear more insights from our conversation with AT&T’s Raj Savoor and many more telco executives, catch the replay of Fierce Network’s AI and the Automated Network Summit here.