AI might take my job, but I've never loved my job more

Last week I watched Karpathy explain how LLMs process text. Tokens, vectors, the whole pipeline. I wanted to play with it myself. So I built an interactive version.

It took me two hours.

Here's What Happened

I watched his video. Text becomes binary. Binary becomes tokens. Tokens become vectors. Cool concept, but I learn by touching things.

I opened Claude. Typed:

"I want four panels. User types text in panel 1. Show the UTF-8 bytes in panel 2. Show hex and tokens in panel 3. Vectors in panel 4. Everything updates live as they type."

It spit out React code. I ran it. It worked. Kind of.

The layout was wrong. I said "make the panels stack vertically on mobile." Fixed.

The tokens weren't highlighting properly. "Color-code the token boundaries." Fixed.

The vector display was ugly. "Make it look like a heatmap." Fixed.

Two hours later, I had something people could actually use to understand tokenization. Not a slide deck. Not a diagram. A tool.

This Shouldn't Be Possible

I'm 39. I learned to code the hard way. Books. Documentation. Hours debugging semicolons. Building something like this used to take days of planning, setup, architecture decisions.

Now? I just describe what I want. The AI handles the boring parts. I handle the "does this feel right?" parts.

Karpathy calls this "vibe coding." Start with a feeling. Let the implementation emerge. No upfront architecture. No detailed specs. Just dialogue between you and the machine.

What It Actually Feels Like

Working with AI is like having a brilliant intern who never sleeps. Super fast. Occasionally completely wrong. But always willing to try again.

Example from today:

That whole exchange? Three minutes.

The Tools Make It Real

I use Claude for most coding now. Sometimes Cursor. Sometimes multiple threads:

Each conversation has its own context. I can jump between them like browser tabs.

Yes, it's expensive. Yes, it occasionally suggests nonsense. No, I don't care. The speed gain is insane.

My Favorite Part

The best thing? I can build tools to understand concepts as I learn them.

Confused about attention mechanisms? Build a visualizer. Want to understand embeddings? Make them interactive. Learning by building, except now building takes hours instead of weeks.

I built another tool to show how training data gets cleaned. Upload messy text. Watch deduplication happen. See quality scores change. It's like pressure-washing a dataset.

These aren't products. They're interactive notes. Ways to think with my hands.

When I Got Stuck

Here's the thing about having multiple AI collaborators. When I ran out of ideas for the UI, I took a screenshot. Opened ChatGPT o3. Told it:

"I'm not happy with this design. You're a UI/UX expert. How would you improve it?"

It returned a detailed plan. Typography changes. Better visual hierarchy. Interaction patterns I hadn't considered. I threw the whole plan into Cursor and watched it transform.

This is what it produced => [link]

Three different AIs. Each doing what they're best at. Claude for the initial build. ChatGPT for design critique. Cursor for implementation. Like having a whole team, except they never sleep and never judge my 2am coding sessions.

Try This Tomorrow

Pick something you want to understand better. Something with moving parts. Then:

Don't plan. Don't architect. Just start. You'll throw away the first version anyway.

The Part Nobody Talks About

Sometimes the AI completely misunderstands. Sometimes it adds 500 lines of unnecessary complexity. Sometimes it breaks everything.

So what? Delete and try again. The cost of being wrong is basically zero now.

I spent years treating code like it was precious. Now I treat it like clay. Squish it, reshape it, start over if needed.

What's Actually Happening

We're not becoming worse engineers. We're becoming faster builders. The gap between idea and reality just collapsed.

That interactive tokenizer? A few people told me it finally helped them understand LLMs. That's worth more than perfect code.

The data cleaning visualizer? Someone used it to debug their training pipeline. Actual utility from a two-hour build.

You Can Do This

I'm not special. I'm not even particularly good at prompting. I just started building instead of planning.

If you've ever thought "I wish I could see how X works" - you can build that. Today. In a few hours.

The tools are all free to start:

Send Me Your Stuff

Seriously. I want to see what you're building. The broken experiments. The "I can't believe this works" demos. The tools that only make sense to you.

We're in the weird early days where a half-baked idea can become a working prototype before lunch. Let's see how far we can push this.

Yes, Most of It Will Be Garbage

I know what you're thinking. If building is this easy, won't the internet fill with trash? Half-working demos? Broken apps? AI slop everywhere?

Yes. Absolutely. The flood is coming.

But here's the thing: We'll also get brilliance from unexpected places. From the kid who couldn't afford bootcamp but had a vision. From the designer who always wanted their mockups to work. From the expert who knew exactly what their field needed but couldn't code.

You're not an engineer because you memorized React patterns. You become one by executing ideas. By taking what's in your head and making it real.

The barriers just fell. The gatekeepers lost their gates. Some people will ship garbage. Others will ship magic. The difference? Taste. Ideas. Knowing what's worth building.

I'm all for removing friction while we learn to build responsibly. That wisdom comes from practice. From millions creating, failing, sharing what works. Not from keeping the tools locked up.

The foundation might be shaky. But watching people build things they never could before? That view is worth it.

Building something? Hit me up. The weirder, the better.