Tag Archives: LLM

AI Is Just A Majordomo

Posted on by
1

The IT world is on fire right now with solutions to every major problem we’ve ever had. Wouldn’t you know it that the solution appears to be something that people are very intent on selling to you? Where have I heard that before? You wouldn’t know it looking at the landscape of IT right now but AI has iterated more times than you can think over the last couple of years. While people are still carrying on about LLMs and writing homework essays the market has moved on to agentic solutions that act like employees doing things all over the place.

The result is people are more excited about the potential for AI than ever. Well, that is if you’re someone that has problems that need to be solved. If you’re someone doing something creative, like making art or music or poetry you’re worried about what AI is going to do to your profession. That divide is what I’ve been thinking about for a while. I don’t think it should come as a shock to anyone but I’ve figured out why AI is hot for every executive out there.

AI appeals to people that have someone doing work for them.

The Creative Process

I like writing. I enjoy coming up with fun synonyms and turns of phrase and understanding a topic while I create something around it. Sure, the process of typing the words out gets tedious. Finding the time to do it even more so, especially this year. I wouldn’t trade writing for anything because it helps me express thoughts in a way that I couldn’t before.

I know that I love writing because whenever I try to teach an AI agent to write like me I find the process painful. The instruction list is three pages long. You feed the algorithm a bunch of your posts and tell it to come up with an outline of how you write. What comes out the other side sounds approximately like you but misses a lot of the points. I think my favorite one was when I had an AI analyze one of my posts and it said I did a good job but needed to leave off my Tom’s Take at the end. When I went back to create an outline for training an AI to write like me the outline included leaving a summary at the end. Who knew?

People love the creative process. Whether it’s painting or woodworking or making music creative people want to feel like they’ve accomplished something. They want to see the process unfold. The magic happens on the journey from beginning to end. Feel free to insert your favorite cliche about the journey here. A thing worth doing is worth taking your time to do it.

Domo Arigato, Majordomo

You know who doesn’t love that process? Results-oriented people. You know the ones. The people that care more about the report being on time than the content. The people that need an executive summary at the beginning because they can’t be bothered to read the whole thing. The kind of people that flew the Concorde back in the day because they needed to be in New York with a minimum of delay. You’re probably already picturing these people in your head with suits and wide tie knots and a need to ensure the board sees things their way.

Executives, managers, and the like love AI. Because it replicates their workflow perfectly. They don’t create. They have others create. They don’t want to type or write or draw. They want to see the results and leverage them for other things. The report is there if you want to read it but they just need the summary so they can figure out what to do with it. Does it matter whether they’re asking a knowledge worker or an AI agent to create something?

The other characteristic of those people, especially as you go up the organizational chart, is their inability to discern bad information. They work from the assumption that everything presented in the report is accurate. The people that were doing it for them before were almost always accurate. Why wouldn’t the fancy new software be just as accurate? Of course, if the knowledge worker gave bad data to the executive they could be fired or disciplined for it. If the AI lies to the CEO what are they going to do? Put it in time out? The LLM or agent doesn’t even know what time out is.

People that have other people do things for them love AI. They want the rest of us to embrace it too because then we all have things doing work for us and that means they can realign their companies for maximum profit and productivity. The reliance on these systems creates opportunities for problems. I used the term majordomo in the title for a good reason. The kinds of people that have a majordomo (or butler) are exactly the kinds of people that salivate about AI. It’s always available, never wants to be complimented or paid, and probably gives the right information most of the time. Even if it doesn’t, who is going to know? Just ask another AI if it’s true.

Tom’s Take

The dependence on these systems means that we’re forgetting how to be creative. We don’t know how to build because something is building for us. Who is going to come up with the next novel file open command in Python or creative metaphor if we just rely on LLMs to do it for us now? We need to break away from the idea that someone needs to do things for us and embrace the idea of doing them. We learn the process better. We have better knowledge. And the more of them we do the more we realize what actually needs to be done. The background noise of AI agents doing meaningless tasks doesn’t make them go away. They just get taken care of by the artificial majordomos.

Human Generated Questions About AI Assistants

Posted on by
Reply

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.

The Dangers of Knowing Everything

Posted on by
Reply

By now I’m sure you’ve heard that the Internet is obsessed with ChatGPT. I’ve been watching from the sidelines as people find more and more uses for our current favorite large language model (LLM) toy. Why a toy and not a full-blown solution to all our ills? Because ChatGPT has one glaring flaw that I can see right now that belies its immaturity. ChatGPT knows everything. Or at least it thinks it does.

Unknown Unknowns

If I asked you the answer to a basic trivia question you could probably recall it quickly. Like “who was the first president of the United States?” These are answers we have memorized over the years to things we are expected to know. History, math, and even written communication has questions and answers like this. Even in an age of access to search engines we’re still expected to know basic things and have near-instant recall.

What if I asked you a trivia question you didn’t know the answer to? Like “what is the name of the metal cap at the end of a pencil?” You’d likely go look it up on a search engine or on some form of encyclopedia. You don’t know the answer so you’re going to find it out. That’s still a form of recall. Once you learn that it’s called a ferrule you’ll file it away in the same place as George Washington, 2+2, and the aglet as “things I just know”.

Now, what if I asked you a question that required you to think a little more than just recalling info? Such as “Who would have been the first president if George Washington refused the office?” Now we’re getting into more murky territory. Instead of being able to instantly recall information you’re going to have analyze what you know about the situation. For most people that aren’t history buffs they might recall who Washington’s vice president was and answer with that. History buffs might take more specialized knowledge about matters would apply additional facts and infer a different answer, such as Jefferson or even Samuel Adams. They’re adding more information to the puzzle to come up with a better answer.

Now, for completeness sake, what if I asked you “Who would have become the Grand Vizier of the Galactic Republic if Washington hadn’t been assassinated by the separatists?” You’d probably look at me like I was crazy and say you couldn’t answer a question like that because I made up most of that information or I’m trying to confuse you. You may not know exactly what I’m talking about but you know, based on your knowledge of elementary school history, that there is no Galactic Republic and George Washington was definitely not assassinated. Hold on to this because we’ll come back to it later.

Spinning AI Yarns

How does this all apply to a LLM? The first thing to realize is that LLMs are not replacements for search engines. I’ve heard of many people asking ChatGPT basic trivia and recall type questions. That’s not what LLMs are best at. We have a multitude of ways to learn trivia and none of them need the power of a cloud-scale computing cluster interpreting inputs. Even asking that trivia question to a smart assistant from Apple or Amazon is a better way to learn.

So what does an LLM excel at doing? Nvidia will tell you that it is “a deep learning algorithm that can recognize, summarize, translate, predict and generate text and other content based on knowledge gained from massive datasets”. In essence it can take a huge amount of input, recognize certain aspects of it, and produce content based on the requirements. That’s why ChatGPT can “write” things in the style of something else. It knows what that style is supposed to look and sound like and can produce an output based on that. It analyzes the database and comes up with the results using predictive analysis to create grammatically correct output. Think of it like Advanced Predictive Autocorrect.

If you think I’m oversimplifying what LLMs like ChatGPT can bring to the table then I challenge you to ask it a question that doesn’t have an answer. If you really want to see it work some magic ask it something oddly specific about something that doesn’t exist, especially if that process involves steps or can be broken down into parts. I’d bet you get an answer at least as many times as you get something back that is an error message.

To me, the problem with ChatGPT is that the model is designed to produce an answer unless it has specifically been programmed not to do so. There are a variety of answers that the developers have overridden in the algorithm, usually something racially or politically sensitive. Otherwise ChatGPT is happy to spit out lots of stuff that looks and sounds correct. Case in point? This gem of a post from Joy Larkin of ZeroTier:

https://mastodon.social/@joy/109859024438664366

Short version: ChatGPT gave a user instructions for a product that didn’t exist and the customer was very frustrated when they couldn’t find the software to download on the ZeroTier site. The LLM just made up a convincing answer to a question that involved creating something that doesn’t exist. Just to satisfy the prompt.

Does that sound like a creative writing exercise to you? “Imagine what a bird would look like with elephant feet.” Or “picture a world where people only communicated with dance.” You’ve probably gone through these exercises before in school. You stretch your imagination to take specific inputs and produce outputs based on your knowledge. It’s like the above mention of applied history. You take inputs and produce a logical outcome based on facts and reality.

ChatGPT is immature enough to not realize that some things shouldn’t be answered. If you use a search engine to find the steps to configure a feature on a product the search algorithm will return a page that has the steps listed. Are the correct? Maybe. Depends on how popular the result is. But the results will include a real product. If you search for nonexistent functionality or a software package that doesn’t exist your search won’t have many results.

ChatGPT doesn’t have a search algorithm to rely on. It’s based on language. It’s designed to approximate writing when given a prompt. That means, aside from things it’s been programmed not to answer, it’s going to give you an answer. Is it correct? You won’t know. You’d have to take the output and send it to a search engine to determine if that even exists.

The danger here is that LLMs aren’t smart enough to realize they are creating fabricated answers. If someone asked me how to do something that I didn’t know I would preface my answer with “I’m not quite sure but this is how I think you would do it…” I’ve created a frame of reference that I’m not familiar with the specific scenario and that I’m drawing from inferred knowledge to complete the task. Or I could just answer “I don’t know” and be done with it. ChatGPT doesn’t understand “I don’t know” and will respond with answers that look right according to the model but may not be correct.

Tom’s Take

What’s funny is that ChatGPT has managed to create an approximation of another human behavior. For anyone that has ever worked in sales you know one of the maxims is “never tell the customer ‘no'”. In a way, ChatGPT is like a salesperson. No matter what you ask it the answer is always yes, even if it has to make something up to answer the question. Sci-fi fans know that in fiction we’ve built guardrails for robots to save our society from being harmed by functions. AI, no matter how advanced, needs protections from approximating bad behaviors. It’s time for ChatGPT and future LLMs to learn that they don’t know everything.