GPT-3 demos spread beyond the research community

Roughly a month after OpenAI opened GPT-3 to a private beta through its API, examples produced by early testers began circulating widely outside the research community that had followed the model’s paper. Beta users posted short stories written in the style of named authors, press releases, guitar tabs and code, and the clips spread fast enough that MIT Technology Review described the reaction as treating a language model as “shockingly good” rather than a research curiosity.

The demo that travelled furthest was from developer Sharif Shameem, who showed GPT-3 turning plain-English descriptions — “a button that looks like a watermelon” — directly into working JSX layout code. Coder John Carmack said the code-generation examples had unsettled him. Other testers produced pastiche, such as a short story written in the voice of Jerome K. Jerome from only a title and opening word.

Coverage was not uncritical. MIT Technology Review’s own account cautioned that GPT-3 remained “a kind of vast, eclectic scrapbook” of internet text rather than a reasoning system, that it still produced biased and offensive language, and that its outputs often read more like a student bluffing through an exam than genuine understanding. Even so, this was the moment a large language model’s outputs became something ordinary technology users encountered and shared directly, rather than a benchmark score reported in a paper — prefiguring the wider public reaction ChatGPT would produce two years later.