Text Is the Universal Interface
The Unix philosophy declared that text streams are the universal interface between programs. Large language models have made this literally true for intelligence itself.
"This is the Unix philosophy: write programs that do one thing and do it well. Write programs to work together. Write programs that handle text streams, because that is a universal interface." Douglas McIlroy
In the late 1960s, Unix pioneers discovered that if every program reads and writes plain text, you get composability for free. Programs written by strangers, years apart, snap together through pipes. No shared memory, no binary protocols, no coordination just text in, text out. This insight built the most enduring software ecosystem in history.
Large language models have extended this principle from programs to cognition. A language model takes text in and produces text out. Because it was trained on the full breadth of human knowledge serialized as text chess games in algebraic notation, code in Python, medical records in clinical prose it can reason about anything that can be expressed in writing. The prompt becomes the new pipe. Every prompt instantiates a different conditional probability distribution, a different "mood," that turns the same general model into a specialized one.
This means entrepreneurs and engineers can now build cognitive software the same way Unix hackers built utilities: find a prompt, don't write a program. Test sets replace unit tests. Prompt engineering replaces feature engineering. The interface between human intent and machine execution is natural language, the most accessible programming language ever invented.
"The language model works with text. The language model remains the best interface I've ever used. It's user-friendly, composable, and available everywhere. It's easy to automate and easy to extend." Roon
Takeaway: Text is the only interface that is simultaneously human-readable, machine-processable, and universally composable and LLMs have made it the default API for intelligence.
See also: Good Prompts Encode Understanding | Compound AI Systems Beat Monolithic Models | The Bitter Lesson Scale Beats Cleverness