ChatGPT and the week the world changed

OpenAI released ChatGPT on November 30th. It’s December 7th and the world is different.

I’m not being dramatic. I’ve watched technology launches for years. The iPhone in 2007. The iPad in 2010. Pokémon Go in 2016. None of them moved this fast. SimilarWeb estimates ChatGPT hit 1 million users in 5 days. For context, Netflix took 3.5 years to reach 1 million subscribers. Twitter took 2 years.

Five days.

The conversations

My mom called me on Sunday. She’d used ChatGPT to write a letter to her homeowners association about a drainage issue. She read me the letter. It was better than what she would have written, and she’s not a bad writer. “It sounded like a lawyer,” she said, pleased.

My barber brought it up during a haircut on Tuesday. “Have you tried that AI thing?” He’d used it to write a social media caption for his shop. He showed me on his phone.

A colleague sent our team chat a screenshot of ChatGPT writing a Python script to parse CSV files. “I’ve been trying to do this for two hours,” he wrote. “The AI did it in 10 seconds. It works.”

My fifteen-year-old niece used it to help with a history essay. Her teacher hasn’t caught on yet. I have mixed feelings about this.

In one week, a technology that existed only in AI research circles became something my mom, my barber, my colleague, and my teenage niece all use.

What it is

ChatGPT is GPT-3.5, fine-tuned with reinforcement learning from human feedback (RLHF). The model itself isn’t dramatically more capable than what OpenAI had before. What’s different is the interface. It’s a chatbot. You type a question. It responds. You follow up. It remembers context. Natural. Easy. No API key. No playground. No prompt engineering.

The interface was the missing piece. GPT-3 was available for over a year via the API. Researchers and developers used it. Nobody else did, because the API requires technical knowledge.

ChatGPT requires nothing. Just words. Just conversation.

The technology was ready in 2020. The interface wasn’t ready until now. That gap, between capability and accessibility, is where revolutions live.

What it does well

Writing. It writes well. Not perfectly. Not with voice or personality or the specific kind of insight that comes from living a life. But well. Grammatically correct, well-structured, coherent prose that’s better than what most people produce.

Code. It writes functional code in dozens of languages. It explains what the code does. It debugs. It refactors. It’s not replacing programmers, but it’s replacing the part of programming that’s tedious (looking up syntax, writing boilerplate, translating between formats).

Explanation. Ask it to explain quantum mechanics to a 10-year-old. It does, and the explanation is surprisingly good. Clear. Accurate enough. Accessible.

Summarization. Give it a long document and ask for a summary. It condenses without losing the important parts.

What it does badly

Facts. It’s confidently wrong about specific facts with alarming regularity. It’ll tell you a book was published in 1987 when it was published in 1992. It’ll attribute a quote to the wrong person. It’ll invent citations that don’t exist. The confidence is the dangerous part. The errors sound as polished as the correct answers.

Math. Basic arithmetic is fine. Anything beyond that gets unreliable. I asked it to multiply 17 by 383 and it got it wrong. The model doesn’t calculate. It predicts what the answer probably looks like, and sometimes its prediction is wrong.

Originality. It can combine existing ideas in new ways. It can’t generate genuinely new ideas. Everything it produces is a recombination of patterns in its training data. That’s not nothing. But it’s not creativity in the way humans understand creativity.

The speed of adoption

This is the part I can’t stop thinking about. One week. One hundred million users within two months. The fastest adoption of any technology in history.

The reason isn’t that ChatGPT is the best AI model. It might not be. The reason is that it’s conversational, free, and immediately useful. Those three things together are a combination that’s almost impossible to resist. You don’t need to understand how it works. You just talk to it and it talks back.

Every previous AI breakthrough required some level of technical literacy to access. ChatGPT requires literacy. Period.

I think we’ll look back at this week the way we look back at the release of the iPhone. Not because ChatGPT is the final product. It isn’t. But because it’s the moment the technology escaped the lab and entered the living room. And once it’s in the living room, everything changes.

My mom is writing letters with it. My barber is writing captions. My niece is writing essays. The genie is out. And I’ve never seen a genie move this fast.

Related thinking: