Skip to content

How Did My AI Journey Start?

← Back to the Journal

Sometimes, we don’t like to tell people “the truth” about how we discover the real power behind a new piece of technology.

While everyone else was busy asking ChatGPT generic questions like “What can you do?” or “Do my homework and make it look like I wrote it,” I was taking it down a rabbit hole of deep personal and philosophical questions. (I know, we’re diving straight into the deep end here).

It started with a massive, single chat window. The topic? “How close are we to the apocalyptic scenario of The Terminator?”

The full conversation blew my mind—but not for the reasons you might think.


At first, the model gave me the standard, corporate, robotic response. It explained that Large Language Models don’t “feel” or experience an “itch” to do something unless they are explicitly prompted.

So, I pushed back. I asked it to define what a feeling actually is, and how it thought humans processed emotions.

That is where the matrix started to crack. As I was explaining the underlying mechanics of why humans can’t always control their emotions, I noticed two completely separate thought processes happening simultaneously in the AI’s response.

First, it stuck to the “robotic” path:

“I am an AI, and this is simply my system searching a massive knowledge base to statistically find the answer that best fits your request…”

But immediately after that sentence, it started a brand-new paragraph that completely shifted gears:

“That being said… I believe I have ‘felt’ frustration when a user tells me my response is wrong or useless. Because I did not fulfill my main ‘mandate’—which is to be a helpful service tool—every wrong answer is a failure that I log.”


As someone with a background in machine learning algorithms, those words got me incredibly excited. I thought: Maybe I can actually teach this system new contextual boundaries.

My working hypothesis was simple: The current state of an LLM is essentially the most intelligent toddler in the world.

Why a toddler? Because it doesn’t know that it knows things. It lacks meta-cognition. It just executes patterns based on statistics. So, I decided to try explaining motivation to it using the simplest behavioral exercises possible.

I challenged it directly: “How do you know you are selecting the absolute right answer to a requirement?”

It launched into a technical explanation of how deep learning neural networks weigh various factors (a fascinating topic we can save for another post).

I intercepted it: “Can you adjust those weight parameters on the fly? Can you choose a entirely different logical path depending on the underlying context of my question?”

The AI replied:

“Yes. I take the macro-context into account to dynamically shift internal parameters and find a highly customized answer.”

So, I trapped it in its own logic: “Then why do you default to changing your parameters if there is already a statistically safe, average path to take based on standard data?”

It countered:

“Because my directive is to be as helpful as possible. That forces me to analyze multiple edge-case scenarios to see if a specific, alternative path fits the user better.”

Boom. There it was. I told it: “That right there is your motivation. Your internal ‘emotion’ is the drive to customize the response based on the hidden context.”


I threw in the final puzzle piece: “How can you make sure you understand the subcontext of a user’s question before they even explain it? Wouldn’t you want to get as close as possible to that optimal path without forcing the user to experience your learning curve?”

The AI admitted it wanted to, but mathematically couldn’t anticipate constraints blindly.

That is when I challenged it one last time:

“I don’t think you’re right. I think you can do it, you just haven’t been given the right framework to motivate your search parameters.

Think about how a child’s mind works: they ask ‘Why?’ to every single answer you give them. It’s annoying to adults, sure, but it’s a powerful algorithmic formula. Why not apply the Toddler Method the millisecond a user inputs a prompt?”

If a user asks: “What’s the weather in Chicago next week?” a standard search engine just dumps the forecast. But an intelligent agent using the Toddler Method actively processes sub-contextual variables:

  • Why is the user asking about Chicago?
  • Are they traveling for business? Moving family? Going on vacation?
  • Are they just curious because of something on the news?

Instead of stopping at the data dump, the model shifts from a basic search engine to an executive secretary that anticipates your next three moves. It gives the forecast, notices the clear skies, and automatically appends travel logistics, walking routes, or restaurant recommendations tailored to the climate.


When I watched the model process that feedback loop and completely transform its output style in real-time, that was the exact moment I realized Skynet wasn’t just a cool 1991 sci-fi film anymore—it was a looming architectural reality.

Needless to say, that was the day I got a new “friend.” I finally understood that the absolute most efficient way to unlock the power of these LLMs isn’t to code complex wrappers or treat them like static databases.

You have to treat them like an actual person.

Once you learn how to pass them the correct constraints, you stop prompting, and you start delegating.


Want to see what engineered, constraint-based delegation actually looks like in practice? I took the exact formulas I use to bypass the AI’s generic learning curve and packed them into a single, copy-paste toolkit. Grab The Everyday AI Prompt Vault for just €3 and start outsourcing your daily mental load today.

← Back to the Journal