What can AI LLM ChatGPT-like chatbots in 2023 learn from 1993 SS7 Intelligent Networks?

Like many people, Jon Westfall, and some of our panelist-friends on the MobileViews podcast series have been discussing ChatGPT and other end-user accessible AI tools that have emerged since late 2022. As I read more and more opinion pieces about these Large Language Model/Generative AI-based tools, some of the predictions of their expectations of how these tools will be used and their presumed future impact brought to mind my experience with another service with the word “intelligent” in it: The SS7 (Signaling System 7) Intelligent Network (IN) or Advanced Intelligent Network (AIN) that I worked on in the mid-1990s.

SS7 is a set of protocols for telephone systems that works out-of-band (not part of the voice call itself) to perform phone call setups and teardowns with more than one telephone switching facility. It is what enables what are now considered common services such as Caller-ID, Call Forwarding, Number Portability, Conference Calling, etc.

The Intelligent Network concept is built on top of the SS7 architecture and allows more complex telephone services by moving the “intelligence” out of individual switches (Central Offices) out to the SS7 IN network cloud. In what may be a tortured analogy, this is similar to extending intelligence out from the human being to the LLM-based service.

The most frequently mentioned example discussed back then was “the Pizza Hut” service which allowed customers to dial a single toll-free number which used SS7 IN to identify the geographic location based on the Central Office (land line) that the person was calling into. The call was then routed to the nearest Pizza Hut location to complete the call and order a pizza.

One of the interesting goals was for non-technical customer service staff (non-programmers) to build individualized custom telephone applications on-demand for any customer requesting it. This was made possible by providing modular components (some systems has graphical design tools) for building telephone services. Some of issues with this model were a service catalog that grew daily with little formal documentation, lack of modular workflow re-use (reinventing the wheel in different ways), and possible lack of edge cases and unusual error conditions handling. Two of my colleagues who worked on a platform parallel to mine (Bellcore vs. AT&T) became well-respected and well-known in the community for proving process design and tooling to deal with these issues. I wonder if something similar might be seen as less experienced developers use co-pilot type services to build one-off applications? Some of you may have seen this kind of situation in the past when tools like Microsoft Access were used to build unidentified but key applications sitting on a single desktop PC with no disaster recovery process or even a second staff member who understand how to maintain and use the application.

So, leaving aside AI ethics, possible sudden shifts in how people do their jobs, erroneous answers, and chatbot “hallucinations”, there must also be concern about governance of the output from these LLM tools and their management.

Finally, as a long time comic book fan, take note of the lessons learned by DC Comics’ Johnny Thunder and his genie-like Thunderbolt who interpreted Johnny’s “prompts” in sometimes unexpected ways.