pip install

Python is a programming language that is especially often used for AI programs. Much of what you’d want to build with it has already been built by someone else and put online for free. Such ready-made code building blocks are called packages. “pip install” is the command used to download such a package and make it ready for use. You don’t type it into a normal program window, but into the command line — a text window in which you give the computer commands directly as text. “pip install numpy”, for example, fetches the computation package NumPy and places it wherever Python can find it.

Why hardly any AI project starts without this command

Almost the entire AI world runs on Python. The major tools for it — PyTorch from Meta, TensorFlow from Google, the Transformers library from Hugging Face — are themselves just packages. Anyone who wants to try out a language model practically always begins with a few lines of “pip install”. Without this system, everyone would have to download the code by hand from a website, unpack it, and copy it to the right location.

The second reason is division of labor. Nobody programs a matrix multiplication or a PDF reader from scratch again. You install the appropriate package and build on top of it. The official directory, the Python Package Index (PyPI for short), holds over 600,000 such packages. This collection is one of the main reasons why Python has become the standard language for data analysis and AI.

But this has a downside. If a project is built on thirty foreign packages, it also depends on thirty foreign developers. If one of them fails or contains a bug, this suddenly affects a great many programs. In security reports, this effect regularly appears under the keyword supply chain risk.

What actually happens when you type the command

pip is itself a small program that comes bundled with Python. If you enter a package name, it first asks PyPI whether this name exists. Then it looks for the matching version for your own operating system and your own Python version. It then downloads the file and places it in a fixed folder that Python automatically searches through on startup.

Things get interesting with dependencies. Many packages themselves need other packages in order to run. pip reads out this list and installs everything necessary right along with it. A single “pip install” can therefore pull in twenty more packages. If two packages require different versions of the same building block, there’s a conflict — developers half-jokingly call this “dependency hell”.

There are virtual environments to counter this problem. These are isolated folders in which each project gets its own set of package versions. This way, Project A can use an old version and Project B a new version of the same package without interfering with each other. Almost every tutorial therefore recommends setting up such an environment before the first “pip install”.

Where you’ll encounter the line in guides and repositories

If you look at an AI project on GitHub, the description file almost always has a section called “Installation”. Below it follows a line like “pip install -r requirements.txt”. The “-r” means: don’t take a single package, but all of those listed in this file. This way, a developer passes on exactly which building blocks his program needs.

The command also shows up in online notebooks like Google Colab, where you run Python in the browser. There, it’s usually preceded by an exclamation mark. In tutorials on ChatGPT interfaces, you typically find “pip install openai” as the very first step. Anyone who skips this immediately gets an error message when running the code saying that the module was not found.

A common misconception is that pip installs programs like an app that you can click on afterward. That’s not true. pip places code that other Python programs can use — afterward, usually nothing is visible at all. You also shouldn’t confuse pip with conda, an alternative management program that additionally manages Python itself and non-Python components.