I have seen a lot of articles recently riding various “hype trains” regarding how Artificial Intelligence is evolving. It feels like every week there is a new model that claims to be sentient or capable of human-level reasoning.
I wanted to understand what is actually true.
So, I did a bunch of research. I read through technical documentation, looked at how the models are architected, and broke down the difference between what marketing departments say and what the code actually does. I wanted to separate the hype from the reality.
This article—and the accompanying course—is a non-technical refresher on Large Language Models (LLMs) for 2026. We are going to look at how they work, how they utilize probability, and what has actually changed in the latest generation of “reasoning” models.
The Reality Stack
One of the biggest issues I see in current reporting is the interchangeable use of terms. If you use “AI” and “Large Language Model” as synonyms, you miss the nuance of what is happening.
I like to visualize this like a set of nesting dolls (or Matryoshka dolls):
- Artificial Intelligence (The Box): This is the broad discipline, dating back to the 1950s. It is just the idea of making computers do things that usually require human brains.
- Machine Learning (The Big Doll): This is a subset where we give machines data and answers, and they figure out the rules. Amazon has a tax team dedicated to this right now. They feed the system 10,000 classified transactions, and the model learns the patterns to classify the next one.
- Deep Learning (The Middle Doll): This gets into neural nets, trying to map processing to how biological neurons fire.
- Generative AI / Transformers (The Small Doll): This is the breakthrough. These are models that don’t just classify data (e.g., “Is this a cat?”), but create new data based on probability distributions.
The Engine: Next Token Prediction
At a very simple level, LLMs are not thinking. They are predicting.
If I feed the model the input: “The capital of France is…”
The model looks at its training data and calculates the probability of the next word (or token). It sees that “Paris” has a 99% probability. “Leon” might have a 1% probability. “Potato” has a 0% probability.
It samples from that distribution and gives you “Paris.”
It is a prediction engine. It is doing this thousands of times to generate sentences, paragraphs, and code.
Navigating the Cloud (Vector Space)
This is where things get interesting. When we talk about “Prompt Engineering,” we aren’t casting magic spells. We are giving the model coordinates.
Imagine a giant, 3D cloud. In reality, this is a 15,000+ dimensional space (which hurts my brain to think about), so let’s stick to a 3D visualization. Every concept and word is a point in this cloud.
When you prompt the model, you are moving the cursor through this cloud.
- “Write an email.” The cursor drops in a generic spot. The output is average.
- “Write an email about tax accounting.” The cursor moves to the business quadrant.
- “Write a formal email about tax accounting for a client.” The cursor triangulates a very specific cluster of data.
The more context you provide, the more you constrain where the model looks for that “next token.” You are navigating it through vector space to get it closer to the answer you want.
The “Reasoning” Era: System 1 vs. System 2
Recently, models like OpenAI’s o1 or similar “reasoning” models have claimed they can think.
They can’t. The underlying engine hasn’t changed. It is still predicting tokens. What has changed is the workflow wrapped around the model.
- The Old Way (System 1): You ask a question -> The model immediately guesses the answer based on probability. This is fast but prone to hallucinations.
- The New Way (System 2): You ask a question -> The system forces the model to generate 30 different approaches. It then uses a “verifier” model to score those approaches, picks the best one, and then gives you the answer.
It isn’t thinking; it is showing its work. It is running the maze multiple times in the background before telling you the correct path.
Key Takeaways
- It’s Just Math: LLMs do not understand concepts; they understand the statistical relationship between words in high-dimensional space.
- Context is Coordinates: Prompt engineering is simply providing enough data points to move the model’s focus to the right part of its internal “cloud.”
- Reasoning is Iteration: New models aren’t sentient. They are just programmed to critique their own output before showing it to you.
Want to earn CPE for this topic?
- Compare Options: See how we stack up against others in our 2025 Flexible CPE Guide
- Understand the Format: Read how Nano-Learning works for CPAs.
- Check Your State: Ensure you are compliant with our State Requirements Guide.
- What is EverydayCPE?
Related Courses:
