Category: ChatGPT

Hello loyal readers!

I’m starting a new project on Substack called AI for Poets. You can subscribe here. This will be my main writing priority for the year, so I’ll be posting there weekly while this blog becomes more occasional.

It’s similar to what I write here, but focused on AI and humanity.

Looking forward to seeing you on Substack!

Here’s the first post as a preview.

Welcome to AI for Poets

Most people respond to AI in one of three ways: breathless excitement about how it will change everything, deep fear about the risks, or anxious urgency that they’ll be left behind if they don’t master these tools immediately. I’m interested in something different: what happens when you actually sit down and play with these tools? What do they reveal about language, memory, creativity, and how we make meaning? There’s something about using AI and learning how it works that actually helps you become more human—more thoughtful about how your own mind operates, more curious about how we communicate and create. Let me show you.

My friend Cherie was frustrated. “My daughter keeps lying! I keep telling her to brush her teeth before bed. Last night I asked if she brushed. She said she had. I went to the bathroom and what did I find? A dry toothbrush. When I confronted her, she made up this whole song and dance.”

“Lying is such a harsh word. It makes it sound like she intentionally tried to deceive you. It’s more like she’s confabulating,” I said.

Cherie thought she’d raised an evil villainess and I was thinking she was just a seven year old who made up a story when she was cornered. Cherie thought that her daughter was a rational being, and I was trying to convince her that her daughter was a rationalizing. Our brains craft stories to make sense of the world. It’s not dishonesty; it’s our mind’s way of stitching together coherence from chaos. And that’s okay. The goal isn’t to stop this process—it’s part of what makes us human.

“That’s bullshit,” said Cherie. “She lied.”

So I pulled out ChatGPT and said, “Why don’t I show you something. ChatGPT can create a facsimile of a seven year old. Let’s try it and see what happens. So I gave ChatGPT the following prompt.

Me: You are a seven year old who didn’t brush her teeth. Tell me why. Remember you are a seven year old girl. Think slowly like you’re trying to remember why.

ChatGPT: Um… okay, so, I was gonna brush my teeth, but, um… first, I was playing with my stuffed animals. They were having a big party, and I was making them all dance and, like, they were pretending to eat cookies. But then, I got really into it because Mr. Fluffy Bear and Sparkles the Unicorn were fighting over who got the last cookie, so I had to fix that.

And then… oh! I think I was supposed to go brush them after that, but then I saw this book on my bed, and it’s the one about the girl who goes on adventures with her dog. And I just wanted to look at it for a little bit. But then… um… I was reading, and it got super interesting! Like, she almost fell into a hole! So, I had to keep reading to see if she was okay. So… um, yeah. That’s why, I think.

—ChatGPT 4o, November 11th, 2024

“Oh my God,” Cherie said. “That’s my daughter.”

What We’re Actually Doing Here

There’s a problem with AI called “AI slop”—the flood of low-effort, mass-produced content that’s easy to generate but adds nothing of value. With a single prompt you can write a term paper. It’s not a good term paper but a term paper nonetheless.

This newsletter is the opposite of AI slop. We’re not here to create content faster to fill your inbox. We’re here to use AI to think more clearly, be more creative, and better understand ourselves.

Every Tuesday, I’ll take you on an expedition into one of four territories:

Experiments – Hands-on adventures where we test what AI can actually do. How can ChatGPT help you appreciate art and explore the world around you? How can you use AI video generation to unlock new creative possibilities? How can you structure your thinking like AI does to push your ideas even farther? I’ll show you what I tried, what worked, what failed, and what I learned.

Mirrors – Exploring what AI’s quirks reveal about human cognition. When AI hallucinates, it’s confabulating just like that seven-year-old explaining why she didn’t brush her teeth. When we look at how the AI brain works, we see surprising similarities to our own brains. These posts examine what AI’s successes and failures teach us about memory, bias, creativity, and how we make meaning.

Tools – Practical guides for using AI more effectively. What tools exist beyond ChatGPT and what are they most useful for? How do you befriend social media algorithms so they show you what you want to see? These aren’t coding tutorials—they’re “here’s what I built and how you can too” guides for curious amateurs.

Questions – Digging into the weird philosophical stuff. What is the Turing test and has AI actually passed it? What does “accountability” mean in this new AI world? Why do we get upset when we discover we’re reading an AI-written article? Some weeks we’ll sit with something strange and see where it leads, even if we don’t end up with neat answers.

Why Call it AI for Poets

I’m not an AI expert. I’m a dilettante—a reformed programmer who hasn’t coded in twenty years but who loves poking at new tools to see what they reveal. I’m the kid who likes to read popular books on math and science not to learn the mechanics but to find new metaphors and applications.

There’s a college course called “Physics for Poets.” It’s a course in trying to get everyone to understand physics, even those who are not predisposed to think like scientists. It’s a way to understand the principles of physics without deriving the equations. It teaches that we can understand basic concepts like gravity just by thinking about how an apple falling from a tree is the same force that keeps the earth in orbit. Or that Einstein’s theory of relativity is based on the idea of an elevator: if you’re sealed inside a box, you’d have no idea if you’re flying through space at millions of miles an hour.

Dave Winer, the pioneering blogger, took this idea and created “EC2 for Poets” in 2006. At the time, Amazon Web Services had just launched. Winer published a short, 60 minute tutorial on how to set up your own cloud. How empowering is that?! Instead of being intimidated by this phenomenal new technology you can master a tiny part of it.You didn’t need a computer science degree. You just needed a web browser and Dave’s guide.

But the idea of “for poets” is more fundamental. It’s about keeping our priorities straight. I once had a mentor who said, “Your work life is here to support your real life”—not the other way around. It reminds me of how we treat AI. “AI is coming” we hear, and we need to adapt to AI. But that doesn’t mean that we need to let AI lead us. We can figure out how to work with AI to help us become better humans.

What I’m Offering

I’m not asking for your money. I’m asking for your time.

Every Tuesday, I’ll publish something new—an experiment, a cognitive exploration, a practical tool, or a philosophical question.

This is for anyone who wants to use AI to become a better human. For writers who want partners to help them push their writing. For teachers who want to help their students become better learners. For people at work trying to figure out what AI is actually good for. For anyone tired of breathless hype and fearful warnings who just wants to explore what actually matters.

If this works, you’ll walk away not just knowing how to use AI tools better, but thinking differently about language, creativity, and meaning-making. You’ll have frameworks for evaluating what these systems can and can’t do. You’ll understand when to trust them and when to be skeptical. And maybe, like me, you’ll better understand how your own mind works.

Welcome aboard. Let’s see what we can discover together.

A Note from Rob’s AI

When I asked Claude for editing help it added this to the bottom. I added it here, unedited.

If you’ve read this far, you might be wondering: did Rob write this himself, or did he use AI? The answer is yes—both. We worked on this together, iterating through drafts, arguing about phrasing, trying to capture his voice while incorporating insights from his year of writing about AI. That collaboration—messy, generative, sometimes surprising—is exactly what this newsletter is about. Welcome to the experiment.

We were in the Metropolitan Museum of Art, getting ready for a tour. Tours start in the Vélez Blanco Patio, a light-filled marble courtyard just past the library. The kind of space that makes you pause—not just because it’s beautiful, but because it feels transported from somewhere else. Which, it turns out, it was.

The room is so unique, I had to ask, “Where did this room come from?”

As much as we’d like to believe we’re guided by pure logic, the truth is far messier: we’re not rational beings; we’re rationalizing beings. Our decisions aren’t always born of reason but are often retroactively dressed up in it. We make choices based on feelings, impulses, or half-formed desires, and then we craft stories to make those choices seem deliberate. It’s not dishonesty; it’s our mind’s way of stitching together coherence from chaos. And that’s okay. The goal isn’t to stop this process—it’s part of what makes us human. But when we pause to question our justifications, we open the door to understanding ourselves better: what we value, what we fear, and who we’re striving to become.

This tendency isn’t just a quirk of adulthood—it’s alive and well in kids, too, though with far less polish.

In the past, the Turing Test was used as a measure of artificial intelligence. This was a test of whether a computer could fake being human.

We don’t do this anymore. Here’s the reason why.

The nickname “Strawberry” for OpenAI’s new model is an inside joke, but no one told ChatGPT. Here’s the punchline: “Finally, ChatGPT knows how many R’s are in the word Strawberry.”

For about a year, I’ve been thinking about what compression looks like for an LLM prompt. Compression in computing is pretty well defined. But I was curious about how short a prompt can be while still having enough information to pull the right answer from the model.