AWS, IBM, McKinsey, and Nokia curate LLMs for domain-specific applications.Bigger is not always better and you have to jump in at some point
If you've used a consumer-grade generative artificial intelligence (gen AI) tool, such as OpenAI's ChatGPT, Google's Gemini, or Microsoft's Copilot, you may have received some interesting and relevant responses. Perhaps you've come back with some confusing nonsense that you have a hard time mapping to the query you originally submitted. This is one of the problems with large language models (LLMs) with tens of billions of parameters. As software sifts through vast amounts of data, it's hard to consistently find the metaphorical needle in the haystack, the magic of taking a query, putting it in the right context, and returning information that can be described as intelligent. It is difficult to do so. So what does this mean for industry-specific AI solutions, such as communications AI tools for infrastructure planning and network optimization, or the many other use cases touted on conference stages? ?
This essentially means that LLM needs to become a smaller language model that starts with an overview of the world's information, filters out the noise, and overlays domain-specific data and your own business-specific data. Masu. This difficult curation step is how companies can leverage next-generation AI across their operations and realize the increased productivity and efficiency that intelligent assistance can bring. From there, it's a matter of more training, better reasoning, building confidence in the machine, organizational buy-in, and then you're off to the races.
For carrier AI LLM, “bigger isn’t always better”
“There are two important things to remember here,” Ishwar Palurkar, CTO of Telecom and Edge Cloud at AWS, explained in an interview. RCR Wireless News. “Firstly, one model does not fit all. Second, bigger is not always better. We tend to think that the more parameters we have, the better we are at work. But , that's actually not true.'' He said a smaller model with tuning that included rapid hiring of engineers, use of search augmentation generation (RAG) technology, and manual input of instructions was better. It is possible that better results can be obtained.
Parulkar laid out a three-step process for operators to follow, adding that they should also consider price/performance, model explainability, language support, and the quality of that support. “Once you have a basic model, you need to select the appropriate dataset and determine the level of tuning needed to actually address your use case. This is a three-step approach: Learn your use case well. …getting the right underlying model, and getting the right data set to tune it…this is what really forms the bulk of the productizable use cases today, but domain-specific. I think there's an opportunity to build a foundational model. That will come a little later.”
AI and multicloud are important strategic priorities for IBM. For operators, this means moving from manual to automated processes. Stephen Rose, IBM's general manager of global industries, described four broad categories of use cases: customer care, IT and network automation, digital labor, and cybersecurity.
He said that when it comes to consumer AI and enterprise AI, especially carrier AI, the big questions are around the provenance of the data, its security, understanding bias, and the general reliability of the system. “If you really look at enterprise-grade AI, it basically starts with knowing where your data comes from, so you can trust it and apply it in more specific and unique ways. “You can,” he said. Because AI knows exactly where the data comes from.I think so [communications service providers] I see two main opportunities going forward and for the industry as a whole. ”
He continued: “One is finding a way to be willing to share privileged data. So the story is that a lot of the data was hidden behind firewalls or within organizational constraints. But now, Openness as a general concept is becoming pervasive throughout the industry, and the data fabric that can actually be built to support AI is therefore becoming more accessible within a given organization in a way that has never been seen before. I think there are opportunities within our organizational silos, but also within specific ecosystems. So I think there's a huge opportunity for us in both areas, but in areas where we have less proprietary but privileged I think if we work on data that is more privileged, and work on the openness within that data, we can do some really interesting things with AI.”
At the end of the day, Rose said, the question becomes, “Where can it actually be implemented?”
Therefore, it is clear that the quality of the data impacts the quality of the AI-powered results. In other words, when garbage goes in, garbage comes out. But here's the problem. Telcos have vast amounts of highly personalized and highly contextual data on the consumer side. On the operational side, there is a tremendous amount of network telemetry that exists and can be leveraged. The problem is that carriers have historically underutilized the data they have, whether for customer-facing outcomes or internal optimization.
The “vicious cycle” of carrier AI data entry
In discussing the data part and the AI part of data, Thomas Raju, a senior partner at McKinsey & Company, floated the idea that the network is a proxy for user experience, so improving the network corresponds to improving the customer experience. I did. “Where AI comes in is that we can use AI to understand everything that’s happening on the network and understand it according to the individual’s needs, whether the experience is present or not. So, first of all, just having this data allows carriers to improve their products. And of course, product improvements improve the overall experience for customers and differentiate them in the competitive environment. This is the first step towards bringing about the source of
Regarding the siled nature of carrier data: “The telecom industry has been plagued by a vicious cycle of bad data leading to bad or insufficient AI, which leads to less focus on data generation, which leads to bad or insufficient data, which leads to insufficient data. etc. But we are coming out of that situation.”
Returning to Parulkar's comment that domain-specific LLMs are the future. This comment was made in an interview conducted last November. Fast forward to February's Mobile World Congress this year, Deutsche Telekom, e&Group, SingTel, SK Telecom, and SoftBank announced the Global Telco AI Alliance. The companies plan to launch a joint venture to develop carrier-specific LLMs, initially focused on digital assistants and chatbots. Also, regarding language support, as Palurkar pointed out, the plan is to optimize for Arabic, English, German, Korean, and Japanese, with more to come.
“We want our customers to experience the best possible service,” Deutsche Telekom board member Claudia Nemat said in a statement. “AI can help make that happen,” she said.
Communications for Telcos Beyond AI, it has traditionally been said that telcos have penetrated deep into various enterprises to increase market share by selling private networking, edge computing, and other solutions. Subthemes have been developed that correspond to what was being considered. Nokia, which appears to be leading the way into the enterprise, trialled an industrial AI chatbot for its MXIE system, a 5G/edge bundle for industrial applications, ahead of Mobile World Congress. The product is powered by MX Workmate LLM, which Nokia touts as “the first of his OT-compliant generation AI solutions.” If you follow this thread, the industry heavyweights presenting at the Hannover Messe industry trade fair this week seem to be going all-in on artificial AI for industry.
While discussing MX Workmate, Nokia's Stephane Daeuble, solution marketing in the vendor's enterprise division, shared his perspective on Gen AI adoption. While this is focused on industrial enablement, it is also relevant to his AI for telecom companies, and indeed his AI in general. “When we got this, we thought, what are we going to do with it?” he said. Is it still early? …[But] We now have a solution that is greater than the sum of its parts. And likewise, we always launch early. We were early adopters of private wireless (back in 2011). People were like, “What are you doing?” But we were right. This is the same, it takes time. But if you don't start, it will never happen. ”
