AI Saves Time — So Why Am I Wasting So Much of It? [Raw Session]

00:00 — AI saves time, but where does it go?

Hey, welcome back to Slow Builds. So this video is about like... AI does save us time. But there's this other side of it that... You go into this void and you waste so much of it. Like, it's like I'm spending more time... With AI than I would have not with it. I'm not sure if that makes sense.

00:24 — When AI helps me think and I write the code

And I had a perfect example of this earlier today, actually. I had an issue in some code and I used AI to work through the problem. I already had a pretty good idea of what was happening. So I explained what I was seeing and I used AI almost like another developer sitting beside me. Here's the behavior. Here's what I think is going wrong and I'm looking at...

Am I looking at this correctly? Is there something I'm missing? That was useful. We narrowed the issue down and once I understood what was happening, I knew exactly where I wanted to change the code. At that point, I could have said... Okay, make the change for me, but I didn't. I just opened the code and wrote it myself because it was faster.

I knew the change was small. I knew exactly what I wanted to do and I already understood why I was doing it. If I had let AI do it, I would then have had to look through the differences. Make sure it didn't touch something else. See what it did with the tests and make sure it [unclear] decided to improve something I wasn't trying to improve.

So in that situation, AI did the part I wanted help with. It helped me get to the problem faster. Once I knew the answer, I didn't really need it anymore. And I think that's becoming a bigger part of how I use AI for coding. I'm not sure I actually want it writing every line just because it can. I even had it help me with the ticket afterwards and that became another example.

It was a small ticket. Nothing complicated. But it connected a couple of things that shouldn't have been connected at all. So I had to reread it, explain what was wrong and get it to adjust the wording. Again, I probably could have just written the thing myself.

02:16 — Explaining, reviewing, and correcting

That's where the calculation gets weird. AI can do a task in seconds. But sometimes explaining the task accurately enough, reviewing the results, correcting what it misunderstood and then reviewing it again takes longer than the task ever would have taken me by myself. And I'm seeing that more often, there are things I don't bother handing over anymore because by the time I explain exactly what I want, I might as well do it myself.

That's the part I'm starting to notice more and more. We talk about AI saving time and I absolutely think it does. I use it every day for coding, research, writing, planning, figuring things out around the house, travel, money, basically anything where I have a question or I'm trying to work through something. I'm not going backwards from that. But I think AI has removed one whole category of wasted time and quietly created another one.

And the one and the one is verification.

03:18 — Research and repetitive work

Coding is probably where I can see this the clearest. There are so many little things involved in development that I know how to do and I have done hundreds of times. If I need to handle a null somewhere, that's not really a problem I have to solve anymore. Same thing with logging, error tracking, some basic configuration, tests, documentation and or looking up how a particular API expects something formatted.

I know how to do those things. The issue was always the time. It's almost like asking what two plus two is. I know it's four. I don't need to spend my afternoon proving that it's four. Before AI, even even simple things could turn into little side trips. You forget the exact syntax for something. You go Google it. Then you're in documentation. Maybe the documentation is for the wrong version.

So you find the current one then you end up on Stack Overflow or in GitHub or reading an issue because someone had the same problem. Eventually you get to the answer for 15 minutes or 20 minutes disappear and you weren't really solving anything different. You were looking something up. AI is incredibly incredibly good at removing that part of it. I can give it the situation and ask it to find the relevant part of the documentation.

I can have it compare two versions of something. I can have and I can give it an error and ask what am I missing? If I already knew the shape of the test, I wanted it can save me more manual typing. A bunch of repetitive setup like it takes care of all that for me. That's real time saving and it's probably one of the biggest reasons I use AI so much.

04:59 — Reviewing code and protecting production

But now there's another side of that. If I let AI actually make the code change, I need to review what it did and reviewing AI code isn't just like looking at it and saying, yeah, that looks fine. I want to know what files it touched and why. Did it only change the thing I asked for or did it decide there was a cleaner way and start changing something beside it?

Did it add a gem or use a library that wasn't already there? If it did, what version did it choose and why? Does it even Does that even fit the application? Did it pick a gem or a library that is outside of our scope? Not properly vetted. Did it follow the patterns that already exist in the code base? Or did it build something technically correct that looks completely different from everything around it?

Then I have to think about the test. Are they actually testing the behavior or did it write a test that basically confirms its own implementation? Did it miss an edge case? Did it change documentation? Did it misunderstand one part of the task and then do a really good job solving the wrong problem? All that takes time and I'm not saying you shouldn't review it.

That's the exact opposite of what I'm saying. You have to review, especially when you're talking about a production application. That's been around for years. A small app I'm building myself is different. I probably know most of the applications. If something breaks, there's a good chance I know where to look and the blast radius is pretty small. A mature production system isn't like that.

There's history in there. There are old decisions that might look weird today but are there because something that happened five years ago. There are customers depending on behaviors no one thinks about anymore. There are integrations, background jobs, APIs, dependencies and all these little connections that aren't obvious from whatever file AI happens to be looking at. The more mature the system is, the tighter I want AI to be.

I don't want it deciding what would be nice. I want it doing what I ask and then I still have to verify that it's exactly that. Maybe AI saved me 20 minutes of typing but I just spent 15 minutes understanding the code it typed for me. It didn't necessarily remove the time. It moved it from creation to verification.

07:26 — Using AI around the work

I think the place AI is becoming more valuable to me is around the work instead of always doing the work. Help me find the problem. Go through the documentation. I don't want to spend an hour reading. Find the relevant API change. Compare these versions. Tell me whether this behavior changed. Bring me the links and show me the parts that actually matter. Let me talk through the idea and point out something I haven't considered.

Then I can make the decision and after I make the change I can bring AI back in and say look at this. Am I missing anything? Is there an edge case here? Does this match what we just figured out? That feels more useful to me than handing over the whole task and hoping I can trust whatever comes back because I still understand the work.

08:10 — Research is also how we learn

There's another part of this I've been thinking about too. Some of the time AI removes wasn't really wasted time. Research can feel slow but research is also how you learn. If I'm digging through documentation looking for one thing I might see something beside it that I didn't know existed. I might figure out why something works the way it does or realize there are two different approaches and understand the trade -off between them.

Sometimes you start researching one question and discover that you've been asking the wrong question. There's value in that. So I don't necessarily want to take all of that away from me. I wanted to remove the parts that are actually tedious without removing the understanding. That's why I'll ask it to show me where it got the answer. Give me the documentation explain what changed or tell me why one approach is different from another.

I don't just want the answer. I still want to know enough that six months from now when AI gives me a bad answer I can recognize that something doesn't look right because that's another problem. If AI does more and more of this stuff for us but we're learning less and less about what it's actually doing then eventually our ability to verify it gets worse too.

That's a pretty strange loop

09:24 — Checking everyday answers

and none of this is limited to coding. I probably Google less now than I have in years. If I want to know when a store closes I ask AI. If I'm looking for a phone number or an address I ask AI. If I'm trying to figure out a tax, a rule, a price, something about a trip or whether something works differently where I live, I'll probably start with AI.

That's incredibly convenient but I also find myself checking the answer. Did it find the right location? Is that really today's opening time? Is that price Canadian or US? Does it know what country I'm in right now or what province or what state? Did it apply the right tax? Is the information correct or did it find something that was true two years ago?

So sometimes all I've done is changed the order. Instead of going to Google to get the answer, I ask AI for the answer and then go to Google to make sure AI was right and then there's the fact that every AI is a little bit different. One knows something about you, another knows something else, one has some context from previous conversation, another has none.

They might be using different tools or different sources or interpret the same question differently. So you ask the same thing in two places and get two answers. Now which one do you trust? Maybe you ask a third, then you're on Reddit, then you're on the official website, then you're on a YouTube channel, then you've got five tabs open trying to verify an answer that was supposed to save you time.

We've made it incredibly easy to get an answer. I'm not sure we've made it equally easy to know whether the answer is right.

11:03 — Is this real? Another verification layer

Social media has the same problem now. You see something and before you even think about what it means, there's another question sitting there. Is this real? Maybe the photo is generated, maybe the video is real but the audio isn't, maybe the person is real and the story is fake. Maybe the event happened but someone attached an AI image to it. We have another verification layer on top of everyday life and that takes energy.

It's not just the minutes spent checking something, it's the constant little questions in the back of your mind of whether the thing in front of you can be trusted.

11:37 — GPS and learning to trust technology

The strange part is that I don't think we're going backwards from any of this. We don't phone every hotel when we're planning a trip and ask whether they have a room. We check Expedia or booking .com or go to the hotel's website. We don't usually call a travel agent every time we want to know the price of a flight. We don't grab the Yellow Pages when we need a phone number.

Yellow pages are a place that you used to look up phone numbers and I haven't had a paper map sitting in my car for years. Those were a thing and I think you can still get them at some gas stations. We move those things online and eventually we trusted them. If I put an address in GPS and it tells me to turn left, I'm going to turn left.

I don't pull the car over and independently verify the road before I continue. Obviously GPS gets things wrong sometimes but it's reliable enough that I've accepted the trade off. That's probably where AI is going too. At some point it becomes reliable enough that we stop checking so much.

12:35 — Getting tired of checking

The problem is this period we're in right now, AI is good enough that you want to trust. It's useful enough that you start depending on it but it's still wrong enough that you know you shouldn't blindly trust everything it says and I'm getting tired of verifying it. I think that's probably what worries me more than whether AI occasionally gets a mistake. What happens when we're all just tired of checking because that's probably how the trust happens.

If something is right half the time, you're careful. Then it gets better. Maybe eight out of ten answers are right and then nine out of ten. Eventually you start looking at that tenth answer and thinking it's probably fine and maybe most of the time it is but the consequences aren't the same every time AI is wrong. If it tells me a price in the wrong currency, that costs me more money.

If I'm traveling and it gets an important time wrong, that matters. I can miss a flight. If I'm asking about rules where I live, money or something going into production application, the stakes change again. The problem isn't necessarily how often AI is wrong. Sometimes it's which answer happens to be wrong.

13:44 — Choosing where AI fits in the process

So I don't think my answer to any of this is use less AI. It's never going to be that. I'm going to keep using more and more of it. I'm just starting to be more selective about where I want it in my process. Right now I think I like it at the beginning. Help me understand the problem. Narrow things down. Do some of the tedious research.

Read the documentation. Find the relevant information and show me what I'm missing. Then let me make the decision. The final implementation is something I already understand and I can write it in five minutes. Sometimes I'm just going to write it. Then I might bring AI back in afterward and use it as a reviewer. That seems to be where I'm landing right now.

AI is absolutely saving me time. I wouldn't use it this much if it wasn't. But I don't think the equation is as simple as we sometimes make it sound. It removed a lot of searching, repetitive work, typing and looking things up. Then it added checking, reviewing, correcting and verifying. Maybe over time, that second part gets smaller as the systems get better and we learn to trust them more.

I don't know.

14:53 — Making this video became the example

But I was actually working on this video with AI and I'm using a tool that it helped me build for teleprompting and the process of making the video ended up becoming probably the best example of what I'm trying to talk about. I had already explained how I wanted the script written. I have instructions for the channel. I have a specific format I use for Apple Notes.

I want the scripts conversational and I don't want a bunch of repetitive little sentences or something that sounds like AI trying to write a dramatic YouTube video. I told it to do it right. Follow the instructions. Get the timing right. Keep my voice and give me something I could copy straight into Apple Notes. Right at the beginning, it had these little lines.

Do this properly. Follow the instructions. Get the spacing right. Get the timing right. Keep my voice. I'm looking at it thinking this is exactly the thing I told you not to do. So I pointed that out and it fixed the opening. Then I had to explain that if the problem is obvious in the opening, it's probably throughout the entire script. Don't just fix the first paragraph.

Rewrite the entire script. It rewrote the script, but then it gave me only the script and it didn't put into the Apple Notes format that we literally just spent time establishing. And I had to correct that too. And at that point, I thought this is the video because AI helped me create all of this. But look at how much time I just spent making sure the AI did what I asked it to do.

I made an entire video of being tired of verifying AI. Use AI to help me write it. Had to correct the writing. Had to correct the format. Had to tell it to redo the whole thing instead of fixing one paragraph. And then had to tell it one more time to format the script so I could actually read it properly in Apple Notes.

And when I paste in Apple Notes, I'm still going to read through the whole thing again to make sure it got it right. Which is probably the best ending I could have come up with for this video.

16:59 — Growing pains and what comes next

So thanks for watching. AI is a funny, it's a funny thing at the moment. It does miraculous amazing things. I think there is limitations on which model you're using, how much you're willing to spend. But I also think there's a huge limitation on what you expect from it and how you use it. And we're getting there. We're going through growing pains. And but I'm excited to see where this ends up because it's definitely going to end up in somewhere great.

So thanks for watching. Bye.