If you have ever used a neural network, you have probably encountered a situation where it starts acting dumb and fails at an elementary task, such as adding numbers.
You ask it to solve an expression, but it makes a mistake in a routine operation that you delegated to save 30 seconds. It cannot even double-check its own answer.
The complaint is valid, but it should be directed not at the neural network, but at the person using it incorrectly.
A neural network is not a calculator.
Broadly speaking, it is a probabilistic predictor of the next token. And the problem lies precisely in how it works with probabilities.
Without going too deeply into the details, before producing the next token, the neural network evaluates many options that could fit in that position. Then it chooses... no, not necessarily the most probable one. It may choose one of several highly probable options.
Why?
Because it has a parameter called “temperature.” Put simply, this controls how variable the response can be. Among the most probable options, the model does not always choose the most obvious one.
This is why a neural network can answer the same question differently each time.
But this is also where the flaw lies: when you give it a mathematical expression, it does not necessarily calculate the sum according to the same rules as a calculator. It selects a probable token that should continue the provided text. That is why the model may sometimes produce an incorrect answer.
So What Should We Do?
This problem was solved a long time ago.
A neural network is not needed to add numbers: that is what a calculator is for. What it does well is transform one type of information into another.
For example, we might not want to retype an expression from a notebook just to enter it into a calculator. Instead, we can simply take a photo and ask the neural network:
“Use a calculator and calculate the value of this expression.”
This prompt directly tells the model which tool it should use.
A calculator does not hallucinate: it performs mathematical operations according to predefined rules. A neural network, meanwhile, helps us work with information.
If we understand when and how to use it, our productivity will increase rather than decrease.
The calculator example is the most basic one, but the same principle applies to other tasks.
A neural network is sensitive to how a request is phrased. To increase the probability of getting a good result, we need to give it suitable tools and specify a clear response format. This skill develops over time.
The task for anyone who wants to work more efficiently is to understand what result they need, describe it correctly, and verify the answer they receive.
This is only one aspect of working with neural networks, but at the beginning, it is probably the most important one. This is where I most often see everyday users run into problems.