I Like Knowing Stuff
I was reviewing some frontend code recently and something was wrong with it. Not visually. The thing looked right and it worked. If you were the designer or PM who asked for it, you probably would’ve looked at the screen and said, “Yep, that’s it. Ship it!”
The problem was underneath. The implementation used raw HTML elements with a bunch of CSS classes to recreate something our design system already had a component for. We had already solved this problem. The styling was there. The behavior was there. The decisions had already been made. And because I know the system, I could see it immediately.
That’s the part about AI that I keep getting stuck on. When I know something, I’m really really good with AI. Almost scary. I can tell almost immediately when it’s going down the wrong path. I don’t have to stare at the output for twenty minutes or ask another model to review what the first model did. I know what I’m looking at. I know the history. I know why we made certain decisions. I know where the bodies are buried. I can course correct almost instantly.
When I don’t know something? Man.
I have no idea if this shit is right
Not long ago, I was working on an iOS and Android app. I am not nor have I ever been an app developer. AI meant I could still work on it. I could ask questions. I could generate code. I could have AI review the code it generated. I could ask another review skill to check that review, and then another round for that one. I could keep going like that for as long as I wanted.
But there was a pretty fundamental problem with the whole setup. I couldn’t really read the code. Not well enough to be the final arbiter of whether what I had built was actually correct.
So AI would tell me something was right and… okay? I guess? Maybe? I didn’t know. I could prompt it again. I could ask it to verify its answer. I could give it more context. I could burn a bunch more tokens trying to manufacture confidence.
Eventually I did what I would’ve done before all of this. I found somebody who actually knew what the hell they were doing. I leaned on the app subject matter expert. If they said it was cool, cool. If they said something was wrong, I took their feedback back to the bot and had it try again. No thinking. Zero.
That’s the weird part. Put me in our design system and I can use AI like an extension of myself. Almost like how AI was presented to be in the early vibe coding days. Put me somewhere I don’t have that depth and AI can make me productive, but it can’t make me certain. Those are not the same thing.
I’m seeing this everywhere now. Designers and PMs are pushing code. Backend engineers are writing frontend code. Frontend engineers are writing backend code. I’ve done work in native apps despite not being a native app developer. And technically, that’s incredible. People can make things they couldn’t make before.
But I think we’ve skipped over a pretty important question, “Who knows if it’s right?”
That’s different from asking whether it works. The frontend implementation I mentioned at the beginning worked. It looked right. The AI had accomplished the immediate task. It just didn’t know that the thing it built already existed.
That’s the kind of problem expertise catches. I’ve spent years in design systems, so I don’t just see a button or a form or some CSS. I see the decisions behind those things. I see where we’re duplicating something. I see when we’re introducing another pattern that somebody is going to copy six months from now. I see when the visual output is technically correct but the implementation doesn’t belong in the system it’s entering.
That’s judgment. And increasingly, we’re asking AI to fill the gaps where that judgment used to live.
Humans wrote mediocre code too
I don’t want to pretend software development was some beautiful craft before Claude & Co. showed up. People wrote bad code. I wrote bad code. We made dumb decisions. We overengineered things. We duplicated things that already existed. We left comments that made no sense six months later. None of this is new.
The scale is new.
Humans write mediocre code at a decent pace. AI writes mediocre code at a massive scale. That difference matters.
When more questionable code enters a codebase faster, it gets harder to establish what the proper patterns even are. One implementation becomes an example for the next implementation. The bot sees that. Somebody copies this. Somebody generates that. Eventually you’re not fixing one questionable decision anymore. You’re trying to figure out which of the seventeen ways we’ve solved the same problem is now the “right” one.
I work on a design system, so this hits especially hard. The whole point is consistency. But the amount of code we can generate is growing much faster than our ability to give every piece of it the same level of attention.
And somehow we’re celebrating that. “Look how much we shipped.” “Look how much code we wrote.” “Look how many features we pushed out.” Okay. Why is that the metric?
We’re worrying about the wrong thing.
I like knowing stuff
I said this in a conversation recently and didn’t think much of it at the time, “I like knowing stuff.”
That’s really what this is about for me. Knowing something gives me confidence. When I pair with AI in an area where I have expertise, I’m not sitting there hoping it’s right. I’m evaluating it while it’s happening. The second it takes a weird turn, I can stop it.
When I’m outside my expertise, I become way more insecure. I prompt more. I verify more. I ask AI to check AI. I waste so many tokens. And sometimes I don’t even know something is off until I’ve gone pretty far down the wrong path.
Get Quinton Jason Jr’s stories in your inbox
Join Medium for free to get updates from this writer.
The model didn’t suddenly get worse. I changed.
That’s what makes me uncomfortable with this idea that AI lets everybody become a generalist. Sure, it does. I’m living proof. But being able to produce something in a domain and actually knowing that domain are two very different things.
And the scary part is that the output can look the same.
The other thing I’ve been thinking about is how I got good at the things I know. It wasn’t because somebody told me I needed 10,000 hours. I was just curious.
I would read the spec because I wanted to see the next cool thing. Not because I needed it for a ticket. I just wanted to know. I’d go to meetups about things I was interested in. I’d go to conferences. I’d hear something and want to understand how it worked. One thing would lead me to another thing. Eventually all those little pieces started connecting.
That’s how I became an SME. Exploration. I spent the time because I thought the shit was cool.
And that’s the part of this AI shift that I don’t know what to do with yet. Because now every time you hit something you don’t understand, you don’t necessarily have to explore it. You can ask. You get an answer. Keep moving. And that feels great. I do it every day.
But I wonder what happens when we remove too many of those little moments where we would’ve had to dig. What happens when I don’t read the spec because the bot can summarize it? What happens when I don’t figure out why something works because the bot already fixed it? What happens when I don’t need to wander into the meeting, read the documentation, try something stupid, break it, fix it, and accidentally learn three other things along the way?
Those things were inefficient. They were also how I learned.
This is the part that scares me as a teacher
I teach college students too, and this is where all of this gets really messy for me. I’m trying to prepare students for an industry where knowing how to use AI is quickly becoming the bare minimum. At the same time, I’m increasingly convinced that AI is most powerful when you already know something.
So what the hell am I supposed to tell them?
Use AI because employers are going to expect it, but don’t use it so much that you skip the part where you become good enough to know when it’s wrong. Learn to code even though you may work somewhere where everybody uses AI to write code. Learn the fundamentals even though the machine can recall them faster than you ever will. Take the slow path long enough to develop judgment, because the industry you’re entering is obsessed with going faster. Oh you’re looking for a job in UX? First, good luck, then use AI for that too. Use AI to fill in your gaps even though you don’t actually have the artifacts to show true value.
That’s a weird pitch. But I don’t know another way around it.
Expertise was always about time. You got good at one thing. Then you started fanning out. Slowly. Strategically. The things you learned in one area helped you make sense of the next one.
Now we’re reversing that. We’re fanning people out first and using AI to fill the gaps. And I’m not convinced those two paths end at the same place.
Production isn’t the scarce thing anymore
Maybe that’s the shift I’m actually trying to understand. Making something is getting cheaper. Code is getting cheaper. Information is everywhere. Answers are instant.
The scarce thing might be knowing enough to assess what comes back. Is this actually right? Does this belong here? Are we solving something that has already been solved? Is this a good pattern or merely a working one? What happens to this decision six months from now? What doesn’t the bot know?
Those aren’t really generation questions. They’re judgment questions.
And judgment is hard to prompt your way into because judgment comes from all that annoying shit we keep trying to optimize away.
Time.
Repetition.
Mistakes.
Curiosity.
Knowing stuff.
I’m going to keep using AI. I’m in way too deep in to pretend otherwise. I’ve built tools around it. I’ve changed how I work around it. I’m teaching students how to use it. I don’t want to go backward. I just don’t want us to get so obsessed with how much more we can do that we stop caring whether anybody actually knows what they’re doing.
Because when I know what I’m doing, AI makes me better. When I don’t, it mostly makes me faster.
I’m starting to think that distinction matters a lot more than we’re acting like it does.
