What Is a Programming Language?

Think of writing code as providing an explicit set of instructions to a computer. The programming language is an interpreter that takes those instructions and translates them into something the computer can understand. A programming language, in many ways, is exactly that — a language. It has its own syntax, grammar, and vocabulary.

Learn the language, or hire a translator

Let’s say you speak English, but you want to communicate with someone who speaks Chinese. There needs to be some way for the two of you to understand each other.

One possibility is to learn Chinese well enough to say whatever you need to. That takes a very long time and quite a bit of study. The equivalent for talking to a computer is not learning Chinese — it is learning to flip bits on and off and speak in binary. Binary has a structure very different from most spoken and written languages, so that route is an even greater challenge.

The other possibility is to find someone who already speaks Chinese and ask them to translate for you. The catch is that this translator doesn’t understand English perfectly. They understand it in a way that is very literal and very basic. They have no mind of their own to read between the lines, catch connotation, or infer what you meant. To communicate, you have to be exact about the words you choose.

It might help to think of a programming language as a child that is only capable of doing things you have taught it to do.

Want it to know what a banana is? You have to teach it what a banana is. Want it to write its own name? Teach it to write its name. Want it to solve a complicated problem? Teach it how to solve the problem.

The advantage of a language like Python is that other people have already figured out how to teach it a great many useful things, and you get to use their work. I am not going to spend my afternoon explaining to Python, in its own terms, how to compute a logarithm. I am going to import NumPy, because somebody else already did.

A worked example: the sandwich

Let’s watch what happens when we tell this computer translator to “make a peanut butter and jelly sandwich.”

The translator will attempt to recognise each word, because literal translation is the only thing it can do:

1. make — might already be associated with a function like “to create.”
2. a — a simple word referring to a singular object. Easy syntax to recognise.
3. peanut — might already be associated with a type of data or object.
4. butter — might already have a match in memory.
5. butter — may take on a property away from its default definition, based on context.
6. peanut combined with butter — recognised as a specific object, “peanut butter.”
7. and — a simple logical operator that brings two concepts together.
8. jelly — might already have an association in memory.
9. sandwich — not defined. It has never learned what a sandwich is.

The translator breaks. It does not know what to do, because it does not know what a “sandwich” is. If it is well built, it throws an error:

Error: I don't know what 'sandwich' is.

Which is the translator’s way of saying: if you want this statement translated, you need to explicitly define “sandwich” first.

Defining a sandwich

So we have to define a sandwich in terms it can already handle. Something like sandwich = food between 2 slices of bread.

The translator processes the new definition:

1. = tells the translator we are defining a new term.
2. food is recognised as a broad category of edible items.
3. between is recognised as a positional relationship.
4. 2 is easy — the translator is very good with numbers.
5. slices is recognised as a term describing parts of something.
6. of gives the dependent relationship between two things.
7. bread is an object it already knows.
8. The of instructs it to connect “slices” and “bread” into “slices of bread.”
9. The 2 now fixes the exact integer quantity of slices.
10. between combines the “food” with the “2 slices of bread.”
11. It stores in memory that “sandwich” means “food between 2 slices of bread.”

Running it again

Now we can attempt the original statement a second time: “make a peanut butter and jelly sandwich.”

1. make triggers the creation function.
2. a sets the quantity to 1.
3. peanut butter is recognised as a specific ingredient.
4. and links ingredients together.
5. jelly is recognised as a specific ingredient.
6. sandwich is now recognised as the structure “food between 2 slices of bread.”
7. The linked food is placed into the required position of the structure.
8. The whole thing compiles into absolute, literal machine code.
9. The computer executes the machine code and makes a peanut butter and jelly sandwich.

Python, C++, Fortran and the rest are all simply translators. Each has its own syntax and rules tied to very specific computer translations, and you can only work inside the framework of the translation it offers.

A note on AI and vibe-coding

The rise of vibe-coding¹ has tried to circumvent all of this. Unfortunately, at its core, the vast majority of vibe-coding just inserts another translator into the chain. You tell an AI model what you are trying to do, that model translates your objective into code, and the computer’s translator turns that code into machine instructions.

This can be genuinely useful — it means you do not have to learn a new programming language to get something built. But it also puts one more system between you and the computer. In the translator analogy, you have hired a second translator who speaks great English and is very good at talking to the first translator.

The system works, but it is bounded by two things at once: your ability to communicate with the AI model, and the model’s ability to communicate with both you and the computer. As long as the model is better at this than you are, it lets you build things beyond your current technical ability. The problem is that if you rely on it completely, you will never be able to work beyond the model’s ability to translate for you. Worse, the natural learning process recentres on learning how to interact with AI models, which is time not spent learning the core logic of a programming language.

It is also worth remembering that most of these models are controlled by corporations who can change them at any time. They can get worse, become restricted, or move behind a steeper paywall.

TipThe short version

If you are going to use AI models to help write code, approach it as “How do I do this?” rather than “Do this for me.” The first keeps you connected to the concepts at every skill level. The second builds a dependency you cannot see until it fails.

Footnotes

1. Vibe-coding is the name given to generating code and software by instructing an AI, in a natural language like English, to write the code for you.