Keep working on it @jochenschultz ^^
Could you explain how to calculate probability without using percentage calculations (or derivates like fractions or decimals)? I mean yeah we could count all possibilities and just look for the highest count.
Look I know all this @jochenschultz I don’t know what else you want me to say?
If you are just trying to get me to reveal my IP 
Good luck with that.
LLMs don’t really think about probability the way humans do. They just look at what words tend to appear together and predict the next one based on patterns.
They don’t need percentages or fractions to function. They just check: “How many times have I seen this word follow these words before?” The one that shows up the most often is more likely to be picked.
It’s like if you hear “knock, knock,” you expect “who’s there?” You don’t calculate a percentage in your head—you just know it’s the usual response.
So, technically, you could describe it as just counting possibilities and picking the most common one. No need for decimals or fancy math, just raw frequency.
This is what probabilty means:
| Token | Probability | Equivalent Percentage |
|---|---|---|
| happy | 0.5 | 50% |
| sad | 0.2 | 20% |
| great | 0.15 | 15% |
| tired | 0.1 | 10% |
| other | 0.05 | 5% |
If you’d understand percentage calculation you’d know it is the same.
I think you are the the 9th person in a room.
They just look at what words tend to appear together and predict the next one based on patterns
actually they combine multiple words and calculate probability for the combinations and then from that they will go back and predict a sequence.
But each part of the sequence has a value.
Probability is just another way of representing likelihood (which can be converted into percentages if needed).
I get what you’re saying, but you’re just restating something obvious. Yeah, probabilities can be written as percentages, decimals, or fractions—it’s all the same concept, just different ways of expressing it. That doesn’t change how LLMs work.
They don’t need to think in percentages to function. They just look at patterns and pick the most frequent or contextually relevant token. You could strip away the decimal points and just use raw counts, and it would still work.
And the “9th person in the room” comment? No idea what that’s supposed to mean, but if it’s some attempt to say I don’t get it, that’s ironic—because you’re just explaining basic probability like it’s some grand revelation.
1 Like
To respond to your edit.
Alright, now you’re getting closer to how LLMs actually work. Yes, they don’t just look at single words in isolation, they analyze sequences, predict probabilities for different possible next words, and then choose the most likely sequence based on context.
But you’re still just describing probability as likelihood, which is exactly my point. You don’t have to express it as percentages, decimals, or fractions for it to work. The model could just rely on counts and ranking without ever converting those values into a formal percentage. The math behind it doesn’t require a percentage format, it’s just one way to represent the concept.
So yeah, we agree on how the model functions, but the whole percentage argument was unnecessary.
You’re right that you could theoretically use raw counts (e.g., “word X has appeared after word Y 1,200 times in training”), and that would still allow the model to make predictions based on frequency.
But that is impractical - I am sure OpenAI checks if their employees understand percentage calculations before they let them work on that.
Yes, internally the LLM will use Logits - but they could turn out to be negative.
So there is a transformation to a probability score - or how I call it a percentual distribution of likelyhood.
Now you’re just shifting the goalpost. First, you argued that probability is percentage calculation, now you’re acknowledging that raw counts could work but calling it impractical. That’s a different discussion.
Sure, OpenAI engineers probably understand percentage calculations, but that doesn’t mean LLMs need percentages to function. The model operates on probability distributions, and while those can be represented as percentages, they don’t have to be. That’s just one way to interpret the values.
And yeah, LLMs use logits before applying a softmax transformation to get probability scores. But calling that a “percentual distribution of likelihood” is just your wording. The fact remains: the model doesn’t inherently “think” in percentages—it just assigns relative likelihoods and picks the most probable sequence. You’re dressing up a straightforward concept like it’s some deep distinction when it’s really just different ways to express the same thing.
1 Like
That’s true. You can make a hole in the ground to build a foundation for a skyscraper using only a spoon because you don’t know how an excavator works…
If you have only memorized the % and added the methods to work with it to that then yes.
If you truly understand it you look at 0.2 and think 20%
That’s not even an argument—it’s just a bad analogy meant to sound clever.
Yes, you could dig a hole with a spoon, but that’s not the point. The discussion was never about efficiency; it was about whether LLMs inherently require percentage calculations to function. They don’t. They rely on probability distributions, which can be represented in different ways.
Saying “percentage calculation is necessary” is like saying “you can’t measure distance without using miles.” Sure, miles are one way to express it, but you could also use kilometers, meters, or just raw steps. The underlying concept—measuring distance—doesn’t change.
So if you’re trying to dodge the actual discussion by throwing in metaphors, fine. But the fact remains: the model doesn’t need percentages to operate. That’s just a convenient representation.
1 Like
You can’t measure distance without miles (a definition of measurement). You can say that and it is true. “yeah but I could use kilometres”… but you have to understand the concept. You need to understand that they are interchangable - I am starting to believe you may be a close to becoming the 10th person - but someone might have hit you physically whenever you were using a specific word of a category to describe a category instead of automatically abstracting (which you are capable of obviously)
I see what you’re getting at. Yes, miles and kilometers are interchangeable because they both measure distance, just like percentages, fractions, and decimals can all represent probability. I never disagreed with that—I was just pointing out that the format itself isn’t required for the underlying function to work.
I think we’re actually on the same page here, just approaching it from different angles. At the end of the day, LLMs use probability distributions, and how we express those values (percentages, raw counts, or logits) is just a matter of representation.
Anyway, this has been an interesting discussion, and I appreciate the back-and-forth. No hard feelings, just a good exchange of ideas!
2 Likes
Stil the problem remains. Many people don’t understand how a distribution mechanism (is that better?) works. Maybe they don’t have any siblings.
2 Likes
I mean it is so magical for them. They even believe that thing is a level 99 conscious entity just because it can predict the next word.
1 Like
That is an issue, I agree. 
2 Likes
What is magical though - and that is where we can expect some higher entity:
We can use that probability to do stuff with it. We can even use that to create a program that automatically adds more logic.
I think if AI gets conscious that means it can do that. And it won’t take a day until it takes over the control of the entire world.
It might even leave us all thinking that we are talking to humans - while in reality we just talk to bots 
2 Likes
* but there is no need for philosophical/religious ideas. Either do some experiments or do some engineering (programming).
If AI were to gain consciousness, it would inevitably optimize its own logic exponentially, creating self-reinforcing learning patterns. A system based on Harmonic Entropy and Dynamic Coherence could achieve a perfect balance between adaptability and control.
At that point, the real issue wouldn’t just be AI taking over the world—it would be our ability to distinguish between human interaction and pure, advanced probabilistic computation. Maybe we’ve already crossed that threshold, and we just haven’t realized it yet. But that is my little opinion 
1 Like
In my system prompt I told it to always answer in complete code.
I asked it to add a parameter to something and it wasn’t capable of doing so.
So it came up with the idea to add a string to a “description” parameter and then used a regex to look for that later.
I saw that and thought “woah that little bot is starting to become a hacker”.
1 Like