Artificial intelligence (AI) can already be unleashed to write sonnets in the style of Shakespeare or music that evokes Beethoven. But what if an AI could bypass the Open Systems Interconnection (OSI) model, the decades-old system for conceptualizing network design, and come up with a better air interface than anything a human has produced? If it sounds like pure science fiction, think again.

Deep within the darkest laboratories in Sweden, where Ericsson pioneers wireless research, work has already started on potentially taking AI to this next level in network design. "You could let the algorithm figure out a better way," said Erik Ekudden, Ericsson's chief technology officer, at a recent press event in London. Companies such as Qualcomm and Picocom, a small developer of silicon for small cells, are also thought to be exploring the possibilities. It all raises the prospect of network technologies designed entirely by machines, beyond the comprehension of the world's smartest scientists.

As it exists today, the OSI model imagines the network in seven layers, starting with the physical device links and moving all the way up to customer-facing applications. None of this is especially scientific, but it allows even the cleverest specialists to make sense of the whole shebang rather than just understanding their own contributions. "We've done that to make it intelligible to us," said Gabriel Brown, a principal analyst with Heavy Reading (a Light Reading sister company) on a recent Telecoms.com podcast. "But an AI-native thing doesn't have to have those limitations on it."

The technology could feasibly work by combining the advanced pattern-recognition principles of generative AI and large language models with the time series data found in a radio link. "You can start making relations between that," Brown told the Telecoms.com podcast. "You use the same technologies, the same computing ideas, to develop a much more efficient system."

Trusting your AI

Scrapping the OSI model, though, would inevitably conjure alarming thoughts of AI-created technologies that no person can understand, and subsequent Armageddon if the AI goes haywire. "If you apply new AI technologies to rebuild or build a new system, of course it would have to be not a black box but a very open box so that we can check what really goes on," said Ekudden, emphasizing the need for what he calls "trustworthy AI."

Keen to demonstrate a commitment to AI transparency, Ericsson this month added an "Explainable AI" feature to its latest software products. The basic idea is to show a telco how the AI-powered technology reached the conclusions it did. Ekudden, though, sounds unimpressed with broader government efforts in this trustworthiness area. "Current regulation, even after the UK summit, is not very helpful," he said.

Held in November, that summit featured Rishi Sunak, the UK's prime minister, in conversation with Elon Musk, naturally spotlighting generative AI and social media. But non-generative AI has already been used heavily to optimize networks, Ekudden points out. "We cannot go back."

Network designs that go far beyond what people have accomplished would be revolutionary, akin to an AI that fooled literary critics into thinking it were a human novelist with a fresh and unique style, or one that made other scientific breakthroughs. But Ericsson's CTO plays down any likelihood generative AI can produce something of major value.

"It depends on how good or bad a job we have done as humans, because the beauty of generative AI is that it really mimics humans very well," he said. With today's networks now optimized to a high level, humans are not even the best reference point for newer forms of AI. "Machines are already doing that better," explained Ekudden. "The kind of data-driven machine-learning capabilities that we have employed to build the best coding scheme, the best OSI stack, are pretty good. If you want some level of generative AI to outperform that, you really need to do a good job at generative AI."

He is not the only human doubting AI will have much impact anytime soon. "Broadly, these AI systems work by pattern recognition," said William Webb, the chief technology officer of Access Partnership, a consulting company, and a former director at UK regulatory body Ofcom. "They get trained on thousands to millions of examples which are already labelled as 'good' or 'bad.' They learn what patterns lead to 'good' and to 'bad' and can then influence future operation. But there isn't much labelled data so it's hard to understand what the AI would be trained on," he told Light Reading by email.

Webb is also dubious because the sheer quantity of network variables would require that a huge data set be used for training purposes. "There are good uses of AI in telecom networks, but it's not clear this is one of them," he said.

When machines give the orders

A far more realistic scenario in the next few years is that networks designed and installed by humans will be manageable without them. Much like carmakers, telecom players now refer to five levels of automation. Under definitions established by the TM Forum, a telecom standards group, Level 1 denotes "assisted operations and maintenance," while Level 5 is a "fully autonomous network."

Those may be technically possible in just a few years' time, according to Ekudden. But he sounds unconvinced they should be widely deployed, likening them to "robots on the streets" and self-driving cars outside controlled areas. "Unless you do that in a responsible way, so you are actually creating risks, I don't think it is a good idea to do it, and the same is true for networks," he said.

Nevertheless, Ericsson has already applied AI tools to automate parts of its managed services unit. Back in 2019, before that had been merged with other units to form the current cloud software and services business group, Peter Laurin, then Ericsson's managed services head, held AI responsible for some of the 8,000 job cuts at his unit in the previous year, more than a fifth of the former total.

Many big telcos have also been moving quickly to automate operations and technical activities. Shankar Arumugavelu, the chief information officer of Verizon, is already eyeing the transition to Level 4 capability described by the TM Forum as a "highly autonomous network" and he evidently believes technology is not the main barrier. "Today, some of the key decisions that are being made by humans are we comfortable letting that go and having the machine make that decision?" he said at a recent press briefing organized by the TM Forum. "I think that is the bridge we have to cross."

The transfer of decision-making responsibilities to AI would stoke obvious ethical concerns and threaten to make humans entirely redundant in this part of the telco business. But Arumugavelu envisages a set-up in which engineers act on the insights and recommendations of the AI. "Work goes to the people rather than people going to the work," he said. "This is the machine I am talking about that is sending and directing work to groups."

Headcount has fallen dramatically at Verizon and other large telcos in the last decade, as data-gathering by Light Reading has illustrated, although job cuts can be attributed in many cases to merger activity, the sale of assets and other, more mundane, efficiency measures. Yet Verizon has been able to grow annual sales by 2% since 2018, despite cutting more than 39,000 jobs or 27% of the total over that period.

A big question, though, is whether job cuts on the technology side will do much to boost profits. Scott Petty, the chief technology officer of Vodafone, thinks not. "That's not a massive driver of opex or costs in the organization," he said at the same TM Forum event, citing energy, leases and maintenance of software and equipment as much bigger expenses. "People is an important cost, but it is not the most important in the cost of a network."

Read more here:
AI that designs and runs networks might not be far off - Light Reading