ChatGPT first came out in November 2022. And the reaction at the time was basically "huh, that's neat." People who'd gotten completely used to Google and Naver search saw ChatGPT occasionally say something bizarre and figured it was just a fun toy.
But then at some point it got updated, and people genuinely started being blown away. Ask "what's apple in English?" and instead of just spitting out "apple," a whole cluster of related information would come tumbling out. Then something different happened. Right after that, someone just typed "and grape?" — no "what's this in English," nothing. And ChatGPT answered: grape. It had understood the context.
The basic structure underneath ChatGPT and today's AI language models was actually built by Google. In 2017, a Google research team published a paper called "Attention Is All You Need" and introduced the world to a structure called the Transformer. ChatGPT, Google's own Gemini, nearly every AI language model today runs on top of that structure.
But it was OpenAI that took that blueprint and built "an AI that talks like a person" first. Why couldn't Google do it first? Because for Google, putting AI out into the world carelessly risked shaking trust in its search product — a much bigger risk for them to take.
Meanwhile, OpenAI went a different direction. "If it's going to talk like a person, it needs to be trained on how people actually talk." The way characters speak in movies, dramas, documentaries. The jargon in medical dramas, the legal language in courtroom films. Once it was trained on that kind of conversational data, it started sounding genuinely natural, like talking to an actual person.
Change how you use the tool, and what comes out changes too. So when I'm teaching my kid to use GPT, I tell them the same thing. Don't just throw one word at it — put in what happened today, like you're just talking.