The Week ChatGPT Launched

Wednesday, November 30. OpenAI launched ChatGPT. By Thursday morning, a million people had signed up. By Friday, I couldn't get access because the servers were overwhelmed. But Friday afternoon, I finally got in.

I stayed up most of Friday night using it. Saturday, I used it for actual work. By Sunday, I'd written down thoughts about what this means.

I've used DALL-E 2. I've played with GPT-3. Nothing prepared me for ChatGPT's baseline utility. This isn't a toy or a research project. This is something I'm genuinely confused about how to react to.

What Actually Happened

Before this week, language models were either: researchers playing with APIs, or developers building applications on top of APIs, or people reading headlines and being confused.

ChatGPT changed that. You can sign up and talk to it like a person. No API knowledge required. No boilerplate. Just conversation.

In the first week, I've used it for:

Explaining complex concepts. I asked it to explain rate limiting in plain English. The answer was clear and well-structured. Better than most tutorial posts.

Debugging code. I had a tricky bug in TypeScript. I described what was happening. It suggested two debugging approaches. One was things I'd already tried. One was a new angle. I found the bug.

Writing copy. I was stuck on a product description. I gave it a few constraints. It generated three options. None were perfect. All were better than staring at a blank page.

Brainstorming. I described a problem and asked it to suggest approaches. Some were obvious. Some were creative. Thinking through which ones were actually good took thought, but the starting point was better.

The pattern: ChatGPT excels at providing reasonable starting points and working through explanations. It's not a replacement for thinking. It's a thinking tool.

The Weird Parts

Friday night, I asked it a moderately difficult technical question. It gave a confident, well-explained answer that was completely wrong. I didn't know it was wrong at first—the answer sounded plausible. I only caught it when I tried to implement the suggestion and it didn't work.

This is the scary part. It's so articulate that you forget it's potentially hallucinating. It doesn't caveat with "I might be wrong." It speaks with authority even when it's making things up.

Saturday, I asked it about something that happened in October 2022. It said it didn't have real-time information and couldn't answer. That's responsible. But it means the system has a hard cutoff. It trained on data up to September 2021. Everything newer is pure extrapolation.

I asked it about some January 2022 events and it got them wrong. It confidently described details that didn't happen. That's hallucination. The model is filling in blanks with plausible-sounding text instead of saying "I don't know."

Sunday morning, I asked it to help me outline an article about the implications of ChatGPT. The outline was generic. Good structure, but no original thinking. It's great at regurgitation. It's less great at actual insight.

What Happens Next

The obvious reaction is hype. "AI is going to replace writers/programmers/thinkers." I don't think that's right. But I also don't think the technology is boring.

What I think is happening: information work is about to change.

Writing a first draft now has a floor. You don't start from blank anymore. You start from AI-generated text and improve it. That's faster for most people.

Debugging and troubleshooting has a new tool. You can describe a problem and get suggestions. Some are good. Some are wrong. But the starting point is better.

Learning is more interactive. You can ask questions and get conversational answers instead of reading documentation. This is genuinely useful.

But there's a floor on how useful this is. It's great for boilerplate. It's less great for novel thinking. It's great for explanation. It's less great for innovation.

The real impact: the people who learn how to use ChatGPT as a tool will be more productive than the people who don't. The people who can't work without ChatGPT will be less productive when they don't have it.

The Business Implications

Every SaaS company just got a new competitor: "use ChatGPT instead of this product." For products that are built on generated content—copywriting tools, template libraries, basic code generation—ChatGPT is a threat.

But for products that are built on specific knowledge or specific workflows or specific integrations, ChatGPT is a feature, not a competitor. You can integrate ChatGPT into your product. Your product becomes smarter.

The ERP systems we build are now thinking about ChatGPT integration. "What if the system could draft purchase order descriptions based on historical data? What if it could suggest reorder quantities? What if it could generate reports automatically?"

These aren't revolutionary. But they're useful. They reduce friction. They're the kind of thing that makes people use the software more.

The Weird Moment We're In

November 30, 2022, will be marked as the date something shifted. It might not be ChatGPT specifically—it might be the moment when language models became accessible to everyone.

Before this week, AI was something you read about or used if you were a developer or researcher. Now it's something you can play with. And it's better at baseline tasks than people expected.

The risk is overhype followed by disappointment. ChatGPT is going to disappoint people who think it's a replacement for thinking. It's going to impress people who use it as a tool for thinking.

The realistic take: ChatGPT is going to change knowledge work. It's not going to replace it. People who learn to work with these tools will be fine. People who ignore them will eventually compete with people who don't.

It's similar to the Excel moment. When Excel launched, spreadsheets changed how people did analysis. It didn't replace analysts. It changed what analysis looked like.

ChatGPT is probably a similar inflection. Not the end of anything. The start of a shift in how certain work happens.

What I'll Be Watching

I'm going to spend the next few months using ChatGPT as part of my actual work and seeing where the friction points are. Where is it genuinely useful? Where does it slow me down? Where does it hallucinate and cause problems?

The real question isn't "how good is ChatGPT?" It's "what does a workflow that includes ChatGPT look like?"

By December, I expect the initial hype to cool. By January, I expect people to start having genuine thoughts about what this tool is useful for. By March, I expect to see the first wave of products that integrate this technology and genuinely benefit from it.

The week ChatGPT launched is the week that language models stopped being research and started being infrastructure. What we build on top of that infrastructure is still an open question.

But the question has definitely changed.