We train machines to automate IT infrastructure
You heard of "natural language processing", or NLP for short?
In NLP the machine executes a program which analyzes a given document. The textual document is written in English, for instance. Since English is a language used by humans to communicate, then it is called a 'natural language', and what the machine does is called 'natural language processing' (NLP).
The languages spoken by humans, such as English Spanish Chinese, are called 'natural' languages.
Question: is there any 'non-natural' language?
Plenty: C++ Java Python Fortran SQL Perl Cobol BASIC Scala Smalltalk etc. These are languages used to communicate with machines, so they are not considered 'natural'.
What happens if we feed a textual document to a machine capable of NLP, but then the document is written in a machine language (like Java) and not a natural language (like English)?
The machine will process the document written in computer language. What the machine does is called 'text processing'. The computer-language document is called a 'program file'.
What happens if we take the above Java-program-file, which was text-processed by machine #1, and feed it to a machine #2 which understands Java language and can execute its instructions (the codes)?
Machine #1 has written a Java program, where the program is then executed by machine #2.
In reality, each machine #1 can write many programs, and instruct thousands of "machine #2"s to run them.
Thus, we have automated the tasks of 'system administrators' and 'computer programmers', among others. These are jobs conventionally filled by humans. Automation offers reduction in expenses, errors, while improvement in quality, on-time deployment, to name a few.
We also train machine #1 to create and deploy entire new machines of type #2, when the demand rises.
When there is more work-load than the existing cluster of machine #2's can handle, our machine #1 uses 'Virtual machine' and 'Container' technologies to create and add more machine #2's to the cluster.
Likewise, when there is many idle machine #2's and the workload is less than the compute capacity of the cluster, machine #1 automatically decommissions some of the machine #2's, hence reducing costs and conserving energy.