In this Jam-packed podcast episode, Robyn Anderson (CEO of Jam AI) and two guests unpack how large language models (LLMs) work and why most users misuse them. They argue that LLMs are probabilistic "relators" that generate text by predicting the next word based on patterns, not by retrieving verified facts from a database. Using a vivid librarian analogy, they explain that the model "rolls dice" (temperature) to produce varied output, which leads to hallucinations when users demand exact numbers, real‑time data, or precise citations. The conversation cites IMD Business School and PwC research to define a "weakness zone" (tasks requiring absolute ground truth such as exact figures, legal citations, real‑time facts, basic math) and a "strength zone" (pattern recognition, connecting ideas, summarizing, drafting, brainstorming). Real‑world examples illustrate the cost of mis‑assigning AI— a marketing agency that asked the model for precise demographic data and received fabricated census figures, versus using the model to synthesize trends and generate campaign concepts. The episode concludes with a practical filter: before prompting, ask whether the request falls in the strength or weakness zone; if it’s a hard fact, stop, otherwise proceed. Listeners are invited to try Jam AI’s free 7‑day Club Jam trial.
Chapters
Welcome to Jam-packed: Business Made Easy with AI. My name is Robyn Anderson, and I'm the CEO of Jam AI. At Jam AI, we have a community where we are training thousands of small business owners every week on how to use applicable and everyday AI. In this podcast, we're focusing on the human side of AI. We're interviewing members of our community and understanding how they're using AI to save more time and make more money. We also practice what we preach. Every Monday, we're generating a podcast from our AI agent. You're going to love it. We have way too much content and not enough time to record, so it's helping us get this information to you. Well, buckle up, take some notes, and listen to what we're about to tell you. And if you want to know more, you can join us free for 7 days in Club Jam by going to jamout.ai for more information.
Over a billion people use AI every week. Most of them are using it wrong. Not because they are bad at it, because nobody explained how it actually works. That changes today. Welcome to The Human Side of AI from jamoot.ai, and let's jump right into this deep dive.
Yeah, I mean, that scale is just, it's genuinely massive when you think about it. You have a billion people trying to fit this technology into their daily workflows, right?
Their creative projects, their strategic planning, all of it. But we're pulling from some really insightful sources today, like IMD Business School and PwC, to figure out what the actual rules of engagement are. Because our mission today is to give you, the listener, that lightbulb moment so you can stop fighting the AI's flaws.
Exactly. Because right now, so many of us are operating completely misaligned with how the tool actually functions. Like, we tend to treat it exactly like a search engine.
Which, I mean, makes perfect sense, right? Because that is how our relationship with computers has worked for decades.
Oh, 100%. You type a query and it goes to a massive internal filing cabinet and it just hands you back a verified fact.
Right. You ask for a file, it gives you a file.
Exactly. But here is the critical distinction from that IMD Business School piece. AI language models are not built to know things. They are built to relate things to each other.
Okay, wait, not built to know things?
Right. They don't have a database of hard facts. So we are expecting a database, but we are actually interacting with a relator.
A relator? Okay, let's unpack this for everyone listening, because I think I have an analogy from one of our tech explainer sources that really nails this.
Yeah, let's hear it.
So imagine a brilliant librarian, right? They possess a genius-level intellect and they have literally read every book ever written in the entire world.
Okay, I'm picturing it.
But when you walk into their library to ask a question, there is not a single physical book in the room. The shelves are completely bare.
Wow. Okay. That is a great image.
Right. So when you ask them a question, they aren't looking up text in an encyclopedia. They are just guessing the most likely next word based on the patterns of everything they have ever consumed.
That is, yeah, that's a perfect way to look at it. They are acting as, uh, prodigalistic storytellers.
Probabilistic storytellers. I like that.
Yeah. And that prodigalistic nature, which the Information Difference article really highlighted, explains a frustration I think you and pretty much everyone listening has run into.
Which is what?
It is exactly why two people can ask an AI the exact same question word for word and receive completely different answers.
Oh, because it's not pulling the exact same file from a cabinet.
It's—
wait, so it's literally rolling dice.
Basically, yeah. It's rolling the dice on probability. It's a feature they engineer into it called temperature, which adds a little bit of chaos so the AI doesn't sound like a boring, repetitive robot.
So it picks like the second or third most likely word sometimes just to mix it up.
Exactly. And that tiny change alters the whole math of the sentence. Which brings us to the really big issue. What happens when we force this probabilistic guesser to give us exact hard facts?
It hallucinates. It just makes things up.
100%. It hallucinates.
And we have this amazing example from The Sources to make this super concrete for you guys listening. The weight loss meal plan example.
Oh yeah. That one is classic.
So if you ask an AI to create a 7-day weight loss meal plan, it does a stellar job, right?
Yeah. It organizes the structure, the timing, the meals perfectly because it completely understands the relationships between those concepts. Morning relates to oatmeal, weight loss relates to lean protein. It's great at that.
But the moment you ask it for, uh, exact calories and macros for every single meal, yeah, it fails. It gets it wrong pretty much every time.
Yeah. It will confidently tell you an apple has 400 calories, right?
And the macros will add up to some wild number that doesn't match the total. And it— what's fascinating here is that being bad at basic math or making up specific numbers, that is not a bug that needs fixing.
Right. It is the literal architecture of the model. It gets the relationship right, but it makes up the specifics because it doesn't actually have a calculator in its brain.
But wait, I have to push back here on behalf of everyone listening. If it's inherently bad at basic math and it just hallucinates facts when you ask for them, Why on earth are business leaders and agencies relying on it?
That's the million-dollar question, isn't it?
I mean, are we just setting ourselves up for failure?
Well, no, we aren't setting ourselves up for failure. We are just— and this is where that PwC piece comes in— we are simply giving it the wrong tasks.
Okay. Giving it the wrong tasks. How so?
PwC has this great framing: to succeed with AI, you have to keep it strictly in its strength zone and completely out of its weakness zone.
Oh, strength zones versus weakness zones. I love this. Let's define the weakness zone first, just to clear the air. What is the quicksand we need to avoid?
The weakness zone is anything that requires absolute ground truth. So specific data points it cannot verify, exact numbers, real-time facts, basic math.
Basically anything where being 80% correct is disaster.
Exactly. If you need a specific citation for a legal brief, do not use a relator.
Okay, so that's the weakness zone. Let's pivot to what it is actually good at. The strength zone.
Right. This is where it shines. Pattern recognition.
You know, it's like, it's like hiring a visionary creative director to do your tax returns.
Oh, that's exactly it.
Right. You wouldn't blame the creative director for messing up the math on your W-2. You would blame the person who assigned them the wrong job.
Yes. And we see this happen in the real world all the time. There is a great scenario from the TechWise source about a marketing agency.
Oh, right. The marketing agency. Walk us through that one.
So imagine an agency trying to save time. They use AI to pull accurate demographic data metrics for a specific zip code. They want the median income, the exact age breakdown, which is deep in the weakness zone, deep in the weakness zone. So the AI just hallucinates a very plausible-sounding census report.
And then the agency drops that fake data into a client pitch.
Exactly. And when the client fact-checks it, the agency looks terrible and they blame the AI.
When really it was user error. They assigned the creative director to do the taxes.
Precisely. Now imagine that same agency using the AI in its strength zone.
Okay. What does that look like?
Instead of asking for hard metrics, they feed the AI 3 different consumer trend reports and ask it to brainstorm campaign angles or draft structural outlines based on the recurring themes.
Because it excels at finding those relational patterns.
Right. The AI will thrive there. It will connect dots between those reports that a human might have totally missed and output an incredible draft.
It's just a complete mindset shift. We have to stop using it as a digital filing cabinet and start using it as this massive reasoning engine.
It really is. And honestly, trying to make AI perfectly accurate at everything, trying to turn it into a flawless database is just a never-ending chase. The tech companies are spending billions on it, but the core architecture is still a probabilistic storyteller.
Which brings us to the ultimate golden rule for your daily workflow. For all the business owners and everyday users listening to this deep dive right now, This is the practical takeaway.
Yes, this is the filter you need to apply every single time.
Before you type a single prompt into an AI, before you hit enter, stop and ask yourself one simple question. Am I in the strength zone or the weakness zone?
It is that simple. Are you asking for a hard fact or an exact number? If yes, you are in the weakness zone. Stop typing.
But if you are asking it to connect ideas or draft an outline or summarize a meeting, then you are in the strength zone and you are about to save yourself immense frustration and a ton of time.
Adopting this single filter changes everything.
It really does. It stops you from finding the tool. And you know, if we stop treating AI like this flawed encyclopedia of facts and start treating it as an engine for mapping relationships, it really begs a bigger question.
It does. And this is something for you to mull over as you go about your day. Think about your own business or your own creative projects. What entirely new connections in your work are just waiting to be discovered simply because you haven't asked AI to connect the dots yet?
What entirely new connections are waiting? I love that. It's a huge opportunity if you just use the tool the right way. Well, that wraps up today's deep dive. This Friday on The Human Side of AI, we are sitting down with a real person building a real business using AI right now. You do not want to miss it. Find us wherever you listen. See you Friday.
