Hello, today is Friday, July 14, 2023. I’m Joe Schmidt from TC2, and this is Staying Connected. Back in the end of May, my colleague David Lee, who is TC2’s technology director, and I recorded a podcast that looked at ChatGPT’s implications for networking and communications. There was so much to cover that we couldn’t fit it all in, so today we are recording part two on this very interesting topic.
Dave is joining me again because he’s the guy at TC2 who’s been following the developments of ChatGPT and AI and machine learning’s impact on networking and communications long before ChatGPT became ready for prime time. Hey David, thanks for joining me today. Joe, it’s always a pleasure, and happy Bastille Day. Well, thank you David.
Okay, so how about if you provide a brief summary of what we covered in our part one podcast? Sure Joe. There’s been maybe a thousand, five thousand new articles and video clips since our last podcast about a couple of months ago, so I will briefly summarize the key points from part one. Firstly, it’s worth reminding ourselves and our listeners that the term GPT means generative pre-trained Transformer. I put the emphasis on pre-trained since we will discuss the importance of that compared to all the hype surrounding AI and other GPT systems. In part one, we also discussed and provided some examples of how ChatGPT can provide terribly errored but confident responses, even if the requester attempts to clarify or re-prompt ChatGPT. And this has not changed from the last podcast, but the GPT providers are working on that. Once you have a better because I saw in the news yesterday that the FTC is now going in for ChatGPT for doling out some wonky information. More specifically, the FTC wants to know if the AI tool harmed people by generating incorrect information about them, collecting data, or publishing false information. Well, we said on the last podcast, Joe, that there is going to be more government oversight of ChatGPT. That’s already happening in EU and EU countries, etc. In the UK, though, I would have thought that Congress would have acted before the FTC. Yeah, I agree with that.
So, David, we also discussed GPT writing computer code, which I think is a sweet spot for this technology. You’re right, Joe. Some university departments already have incorporated this into their class curriculums. They’re working it over the spring semester of this year, and there have been some books written by those very same professors already. In terms of how one can use GPT to write computer code more efficiently and integrate it into their curriculums. One other topic we just described was how a handful of companies have been using AI and machine learning in their products in the past but are now using GPT to improve their natural language queries about the network, such as ‘Why is my network slow at the Peoria, Illinois site?’ with conversational responses to find out the root cause and optimize IT operations.
That’s a good overview from part one, and obviously, the complete podcast is available, along with over 200 other Staying Connected episodes. That’s a nice plug, don’t you think, David? Yeah, yes, nice plug, Joe. Okay, back to this podcast. You mentioned at the beginning the importance of ‘pre-trained’ when referencing GPT. Can you elaborate on that, David? Sure, Joe. As mentioned, GPT is a knowledge base or repository of information that is a large language model or LLM. The importance of the pre-trained concept is to understand what knowledge base is being used when the Transformer responds to a query. For example, ChatGPT from OpenAI is using the entire public internet as its knowledge base. We know that the internet now even has ChatGPT hallucinations or fake stories fed back into the overall internet. So, when you do a query today, ChatGPT then takes that information and creates yet more false content or hallucinations, meaning garbage in, garbage out. On the other hand, a company or an organization using a pre-trained knowledge base that is limited with specific information and trained to respond to only the business or activity that is relevant will yield a reliable and useful response.
Can you give some examples concerning your pre-trained point and the knowledge base? Sure, Joe. As a matter of fact, in a recent Wall Street Journal article, there was a startup e-commerce company facing consumers that essentially created an AI or GPT customer service agent and laid off a couple of dozen of its human agents completely. They used their data scientists to pre-train a knowledge base, and with backtesting, the company’s AI-based agents were able to provide the answers faster and more accurately than a human could. Note that in this particular use case, the customer queries and responses required were fairly basic, but they did get natural language, human-like responses to their queries. As an interesting GPT story, I have heard that these AI agents or GPT agents can also discern sarcasm and respond appropriately. For example, hearing an airline customer say ‘Thanks for the upgrade’ versus ‘Thanks for canceling my flight.’ Both are thanks, but there’s some sarcasm obviously in the second one. But the bottom line is that a well-trained GPT agent with a limited LLM and guardrails can be a pretty effective agent that can either assist the human agent in a contact center or potentially replace the human agent, as this startup did.
So, do you by chance have some other recent examples and maybe some best practices? Yes, and expanding on the use cases mentioned previously, enterprises can employ now or in the near future AI-enabled NetOps tools that can pinpoint an issue with, for example, the cloud provider, allowing NetOps to reroute around the issue or contact the cloud provider for faster resolution. Eventually, the AI and ML-driven solutions will be able to initiate a script approved by a human operator, of course, that solves a known error or issue that it discovered. But not every AI solution or product is fit for purpose. As an example, we were working with one of the top five U.S. banks recently, assisting them in refreshing their network infrastructure. One of the key factors for the network management and visibility solution that comes with the OEM’s equipment is that the network management tool or system can be deployed premise-based in the client’s data centers for security reasons. This is in contrast to these newer solutions which have optimized or only work in a cloud-based environment for scalability reasons. For a different use case, using a customized and well-trained LLM and AI system can monitor conversations between contact center agents or financial advisors and their customers with real-time alerts or customized alternative scripts that meet the latest regulatory policies.
Another example would be using an AI solution to act as a firewall between, say, ChatGPT and the eventual content presented to your employees. And that’s good stuff. Okay, David, I will give you the final word. Thanks, Joe. Since our part one, our Fortune 500 clients continue to ask us about AI systems and the implications to their networking infrastructure and communications technology. In addition, I still observe our enterprise clients asking their IT executives to optimize their spending on IT ops. So, budgets are being reduced. At the same time, it is expected to be proactive instead of reactive. So, AI and ML can provide more automation instead of manual processes with large IT staffs.
Okay, thank you, David. I really do appreciate you sharing your insights on this, and I know there is so much more to come. If you want to hear more about ChatGPT and other AI systems and how they might impact you and your company, feel free to contact David, me, or any of our LB3 and TC2 colleagues by giving us a call or shooting us an email. You can also stay up to date by subscribing to Staying Connected, by checking out our websites, and by following us on LinkedIn.
