A 2018 architecture name that escaped its own paper and became the generic word for “the chatbot.”
“GPT” was never meant to be a household word. It’s an acronym for a specific engineering choice: generative (it produces text one token at a time), pre-trained (it learns general language patterns from a huge unlabelled text corpus before being fine-tuned for anything specific), and transformer (the underlying neural network architecture). OpenAI’s 2018 technical report introduced it as one model among several competing approaches at the time — a somewhat dry, internal-sounding name for what was, in hindsight, the branch that won.
It won because it scaled well: each version (GPT-2, GPT-3, GPT-4…) got bigger and more capable with roughly the same recipe, and by GPT-3’s public release the name had drifted loose from the paper entirely — into “GPT” as a stand-in for “a large language model,” “a chatbot,” or in this essay’s usage, one provider among a list of interchangeable free-tier options next to OpenRouter, Groq, and Gemini. The pop-culture flattening is almost the opposite of the term’s origin: a precise architectural label became a generic noun for the whole category, the way “Xerox” became “photocopy.”
The transformer architecture underneath GPT is itself a separate, peer-reviewed contribution — not OpenAI’s own technical report, which was never submitted to peer review.
Source: Radford, A., Narasimhan, K., Salimans, T. & Sutskever, I. (2018), “Improving Language Understanding by Generative Pre-Training,” OpenAI technical report — not peer-reviewed; underlying architecture: Vaswani, A. et al. (2017), “Attention Is All You Need,” Advances in Neural Information Processing Systems 30 (NeurIPS) — peer-reviewed.