I’ve taken a number of briefings in the last few months that all mention how companies are starting to get into AI by building an AI virtual assistant. In theory this is the easiest entry point into the technology. Your network already has a ton of information about usage patterns and trouble spots. Network operations and engineering teams have learned over the years to read that information and provide analysis and feedback.

If marketing is to be believed, no one in the modern world has time to learn how to read all that data. Instead, AI provides a natural language way to ask simple questions and have the system provide the data back to you with proper context. It will highlight areas of concern and help you grasp what’s going on. Only you don’t need to get a CCNA to get there. Or, more likely, it’s more useful for someone on the executive team to ask questions and get answers without the need to talk to the network team.

I have some questions that I always like to ask when companies start telling me about their new AI assistant that help me understand how it’s being built.

Question 1: Laying Out LLMs

My first question is always:

Which LLM are you using to power your system?

The reason is because there are only two real options. You’re either paying someone else to do it as a service, like OpenAI, or you’re pulling down your own large language model (LLM) and building your own system. Both have advantages and disadvantages.

The advantage of a service-based offering is that you don’t need to program anything. You just feed the data to the LLM and it takes off. No tuning needed. It’s fast and universally available.

The downside of a service based model is the fact that it costs money. And if you’re using it commercially it’s going to cost more than a simple monthly fee. The more you use it, the more expensive it gets. If your vendor is pulling thousands of daily requests from the LLM is that factored into the fee they’re charging you? What happens when the OpenAI prices go up?

The advantages of building your own system are that you have complete control over the way the data is being processed. You tune the LLM and you own the way it’s being used. No need to pay more to someone else to do all the work for you. You can also decide how and when features are implemented. If you’re updating the LLM on your schedule you can include new features when they’re ready and not when OpenAI pushes them live and makes them available for everyone.

The disadvantages of building your own system involves maintenance. You have to update and patch it. You have to figure out what features to develop. You have to put in the work. And if the model you use goes out of support or is no longer being maintained you have to swap to something new and hope that all your functions are going to work with the new one.

Question 2: Data Sources

My second question:

Where does the LLM data come from?

May seem simple at first, right? You’re training your LLM on your data so it gives you answers based on your environment. You’d want that to be the case so it’s more likely to tell you things about your network. But that insight doesn’t come out of thin air. If you want to feed your data to the LLM to get answers you’re going to have to wait while it studies the network and comes up with conclusions.

I often ask companies if they’re populating the system with anonymized data from other companies to provide baselines. I’ve seen this before from companies like Nyansa, which was bought by VMware, and Raza Networks, while is part of HPE Aruba. Both of those companies, which came out long before the current AI craze, collected data from customers and used it to build baselines for everyone. If you wanted to see how you compared to other high education or medical verticals the system could tell you what those types of environments looked like, with the names obscured of course.

Pre-populating the LLM with information from other companies is great if your stakeholders want to know how they fare against other companies. But it also runs the risk of populating data that shouldn’t be in the system. That could create situations where you’re acting on bad information or chasing phantoms in the organization. Worse yet, your own data could be used in ways you didn’t intend to feed other organizations. Even with the names obscured someone might be able to engineer a way to obtain knowledge about your environment you don’t want everyone to have.

Question 3: Are You Seeing That?

My third question:

How do you handle hallucinations?

Hallucination is the term for when the AI comes up with an answer that is false. That’s right, the super intelligent system just made up an answer instead of saying “I don’t know”. Which is great if you’re trying to convince someone you’re smart or useful. But if the entire reason why I’m using your service is accurate answers about my problems I’d rather have you say you don’t have an answer or you need to do research instead of giving me bad data that I use to make bad decisions.

If a company tells me they don’t really see hallucinations then I immediately get concerned, especially if they’re leveraging OpenAI for their LLM. I’ve talked before about how ChatGPT has a really bad habit of making up answers so it always looks like it knows everything. That’s great if you’re trying to get the system to write a term paper for you. It’s really bad if you try to reroute traffic in your network around a non-existent problem. I know there are many systems out there that can help reduce hallucinations, such as retrieval augmented generation (RAG), but I need that to be addressed up front instead of a simple “we don’t see hallucinations” because that makes me feel like something is being hidden or glossed over.

Tom’s Take

These aren’t the only questions you should be asking about AI and LLMs in your network but they’re not a bad start. They encompass the first big issues that people are likely to run into when evaluating an AI system. How do you do your analysis? What is happening with my data? What happens when the system doesn’t know what to do? Sure, there’s always going to be questions about cost and lock-in but I’d rather know the technology is sound before I ever try to deploy the system. You can always negotiate cost. You can’t negotiate with a flaw AI.