You said to use AI, you didn’t say to use it well

I would find this person, I would take my glasses off and I would be like, "Are you okay? Do you need help?" Wait, from our last—

what is it? This is the worst project I have ever seen.

That would be you, good cop, bad cop.

I'd be like, "How can I help you?" You'd be like, "This is— no, it's really bad." You're listening to Autha VPN.

The show where two work besties get real about working in tech. Have you ever heard of AI slop?

I don't think I have.

Have I?

No. What is it?

I mean, I think it's a term that's been around for a decent amount of time at this point, but it's basically referring to just like bad, awful AI products, if you will. So, okay. Like things that are produced or generated by. AI. So like, I don't know if you see like a really bad, badly done image of something, you might call it like, what is this AI slop? Or like you read a bunch of articles that are like clearly written by an LLM. You're like, this is just slop. What is this? So today's story, we're going to talk a little bit about AI slop.

Oh my gosh. This sounds exciting.

It is exciting, and I'm especially excited to talk about it with you because you and I know a thing or two about AI coding tools and AI coding tool adoption and rollout. And that's just been our, that's been our bread and butter for like, what, multiple years at this point.

All right.

So this is a story about what happens when velocity and moving quickly is the only thing leadership cares about.

Oh no.

Our story is going to center around Raj. Raj is a tech lead at some company. We don't know all the details, but assume it's like a medium to large size company.

Okay.

Not like a massive company, but just like medium to large. Okay. At this specific company, he is a tech lead of a tech lead role, which is a dev slash manager hybrid. So he also has some direct reports and Raj's specific team has a very wide area of domain and ownership. So what this means for them typically is that a lot of other teams are collaborating with Raj's team. They're coming to them for pull request reviews. Like basically anytime somebody touches or needs to touch the code that lives within, uh, this team's domain, it requires a lot of like back and forth communication across teams.

Got it. That sounds like a lot of collabs, lots of talking, lots of— there's a lot. That's a lot.

It does seem like a lot. It does seem like a lot. And like I mentioned, like it's a lot of just reviewing PRs and they get pinged constantly for PR reviews. But it's normal for them. They know how to live like this.

Yeah.

So recently though, at Raj's company, much like many other companies in today's world, there's been a push to adopt AI coding tools.

Okay.

For engineering, they started to adopt tools like Cursor, and the desired goal is of course to move faster. So setting the stage, there's a goal. Let's move faster. Let's develop things quicker than we're doing them now. And there's a specific tool that was named and there's just like a general push for adoption. So like we know a thing or two about this. What are some things that come to mind when you hear engineering organization wants to adopt AI coding tools?

Like just.

Anything that comes to your mind.

The first thing that comes to mind is just add AI as a strategy. And I feel like a lot of people who are still developing, like, this is such like a hot topic, obviously. I think that sometimes people forget to tie it to business goals, or like they don't quite understand the technology, so there's like knowledge gaps. So there's like when you're first starting to think, we need to start getting in tune with the industry, there's a lot of like knowledge kind of consumption and realization and how to take this new technology and start to incorporate it into your company in like different ways. That is like the first step that you see companies do where it's not quite clear how it fits into their strategy or how it quite fits into their workflows. And so there's a very interesting, like experimental, like first step before people start to find ground. And I think sometimes you hear mistakes and weird stuff happen in there, or also can be really exciting time because things are shifting.

Yeah. And different teams and different companies handle those question marks differently. Of course, some people are like, hey, let's treat this like any new tool. We'll roll it out, give it to a few people, try and get some people to adopt it. And then if that goes well, we can roll it out to more, and other people are like, let's open the floodgates and whoever wants to try this tool can try this tool. Some companies go through formal training. Others are like, go read the docs yourself. Have fun. Raj's company is more on the side of everybody gets a tool. Whoever wants to use it, go ahead and use it. We don't really have best practices. We just want people to start. Moving faster. That's our goal. To Raj and his team, this has caused major, major problems.

Yeah, I could, I could see how that, that might happen. I always talk about go faster. Go faster is good, but to where? Right. It doesn't really matter. Like you could be going infinite speed, but like where are you going and why? I think it doesn't matter. So there's also just like. Also the inverse to, which I imagine since we're talking about AI slop is like the inverse polar opposite. If there's two opposing things, like on the opposition of speed is quality. So you often are gonna risk some quality stuff, which sometimes is a good strategy, but we'll see.

So yeah, Raj and his team are specifically buried in what Raj calls an endless stream of AI slop. Are you ready to hear exactly what they've been encountering? Oh my God.

Is this like a spaghetti code nightmare?

Yes. And more. Yes. And worse.

Oh my God.

So keep in mind the context that I mentioned earlier. Raj and his team typically review a lot of pull requests as part of their day-to-day jobs because they have, they own a wide area of the product that they're building. So in the past few weeks, this team has, one example, received multiple pull requests over 5,000 lines long. Oh no! Most of which should have been under 100 lines.

Stop.

They have seen pull requests with massive file changes, some of which don't even serve a real purpose. Like for example, logic that is like redirecting to different like helper functions that just don't do anything and end up like—

Oh my God, no.

Yeah.

So they didn't review their own code, basically.

That's what we'll get to. There's also, of course, as comes with any AI anything out there, there's hallucinations. So like in, in one particular instance that Raj mentioned, The AI invented an external service and mocked it out, and people are submitting this code and pull requests as if those services are real. And Raj is like digging through.

Stop. No.

Yeah. Raj is like digging through this code and like other repos and like, it's just like looking up stuff like service maps. It's just, do these endpoints exist? Is this real?

Did you even read your code? Like. Mm-hmm. I think you're right. What?

Of course, Raj is also pissed that these PRs ignore like any sort of common patterns that they at this company have decided to adopt in terms of like architecture style, like comment style, like how you even set up directories, like those types of things. And Raj is like leaving dozens of comments on these pull requests.

Oh my gosh.

As he should. And in one case, he asked an engineer, hey, like, why did you change this specific file? I'm not quite seeing how these new methods in here are actually doing anything that we care about related to what you stated in the pull request summary. What do you think a bad response is to this?

LOL, I didn't even know that was there.

Yep. The response, I don't know, cursor out of it.

My good old coworker cursor. Yep. Just, oh man. I, as someone who like uses these tools, like this is actually such a good story of what not to do because yes, I, and also I think going back to what I was saying before about there's like that period of time where people are figuring out like what do these tools do? Mm-hmm. But they're not quite replacing us just yet because they're often wrong and they require a lot of shepherding. Um, and I still review my own code that the AI did. I'm reviewing more, right, than writing when I'm using these tools. I'm changing a lot of the things. It never gets it right the first, second, or third time.

Yep.

Um, and it does make stuff up. I'll be like, and I, I ask it, I'll be like, did you, Do you even know where you got that?

I didn't.

And it's like, I don't. It's like, you're right. Yep.

Yep. I love that. That's always like the first response. You're right. You're right. Like, yeah, I know. Uh, that's why I'm still here. That's why you need me. Um, but yeah, this is the, this is an absolute mess. They're getting now 30 PRs a day from various—

Oh no.

Because not only is it a quality thing, there is a throughput thing here. People are just putting up AI slop pull requests, and they're putting up a lot of them. They're getting through their issues faster because they're not really being intentional about how they're using AI coding tools. They're just having them do the thing, not really reviewing the work, and then putting up the pull request.

I kind of want to see where this would go if like we just kept letting people merge that code. What would happen? Theoretically, like, no one would get what was happening 'cause the code wouldn't make any sense 'cause it would lack intentionality. But like the robots would be able to read it too, so they could figure it out and then they'd make up more stuff. It's like eventually would it just not matter? I would love to see if we just made a repo that just like we didn't even, just let the one shot, like the first time it added the code, just like kept doing that.

That'd be interesting.

Yeah.

We create a rule that's like, You have to accept the first suggestion always.

We should do a project like that. Let's, let's do one where they, and they review their own code from other ones. We'll do like 3 or 4 different tools and see, like, we'll tell them what the objective is for the repository and we'll just write tickets that it reads from. I love this.

Yeah.

Love this.

Let's see what happens. And it's low risk. We're not giving this to customers.

Oh no, we're giving it to cust—

no, I'm just kidding. And we're charging a premium. No, but in this case, they're doing customer-facing work at this company.

That is insane. But you know what, that's so scary. They're not reading the code. They don't know if like there was any security problems. Like, you could be exposing something.

We'll get there.

Oh my God, this is just bad. This is making me sad.

This isn't funny.

This sounds like terrible, like someone's playing a joke on me.

That, listen, Raj agrees.

Poor Raj.

They're getting around 30 pull requests a day from various teams. How many devs do you think they have on this team?

Uh, probably 5.

Yeah, 6. So it's pretty good. But on top of this, they have their own code they need to write, other things they need to do outside of PRs. Like everybody knows engineers don't just sit and write code all day. You have tons of other aspects of your job. One junior engineer dropped a 10,000-line PR.

Shut up. That is crazy. Yeah.

In this case, Raj immediately just rejected it and commented and it was like, I don't even know what you're attempting here, but I'm not going, I'm not going through this. Like just immediately, no. Immediately, no.

I think I pushed maybe 1,000, 2,000. Like after I start hitting that, it's crazy, but. 10,000 lines.

I wouldn't even comment on this PR. I would find this person, I would take my glasses off, and I would be like, are you okay? Do you need help?

Wait, from our last one, uh, what is it? This is the worst project I have ever seen.

That would be you, good cop, bad cop.

I'd be like, how can I help you?

You'd be like, this is No, it's really bad.

But on top of that, they probably didn't even read. That is crazy. That is great. 10,000 line PR. I might do that as a prank. We should send that to somebody. We should make a 10,000 line PR.

Too bad you waste AI resources every day. Do it for people.

Shult today if you try hard enough.

You don't care what days you pull your pranks.

No, I haven't done one in a while. It's time.

Yeah, leave me alone. Anyways, you brought up a really good point, which is that it sounds like these types of PRs might actually be slowing us down, right? Because now we're putting up work that is requiring much more intense review, if not full-blown rejections, and a lot more iterations than if somebody had used a little bit more critical thinking to just do the task themselves in the first place. Again, possibly with an AI coding tool like Cursor, but we're getting to this territory of, okay, are we actually moving faster? So Raj brings this to management and is like, hey, you all said that this was a goal and I'm finding this, this, this, and this. This is what's happening. Clearly, I don't think we're moving faster. And Raj is told, listen, I just feel like you're being a little too strict. People are learning, they're experimenting, like just hold off. You're being too strict. Meanwhile, things get real bad. A PR comes across Raj's desk and at first glance it looks innocuous. Like he's just, okay, whatever. Maybe this PR is fine. Mm-hmm. Then he notices something. This PR included a very subtle change in middleware that actually caused it to abort early. Which was skipping a shit ton of like app authorization logic. Is this person authorized to—

stop.

Yeah.

And so basically what would happen is like anyone could just authenticate.

Yes.

Anyone.

Yep. It just bowed out really early.

Oh my God. From like Anonymous to my grandma. Anyone could— that's crazy. Yep. And they were just gonna merge that.

Most companies who are handling this type of logic also have tests, right? You would think, okay, great. Like we'd have some tests that make sure that we're not changing this like extremely important, like security-related logic that is core to the safety of our business.

Does this company have the tests?

They do have the tests, but this PR changes the tests. Stop.

That's crazy. And so there's like a bunch of bots. I said it's not me that's wrong, it's you. It's the test. Yep.

And so when this PR runs against old tests, the new tests that are changed, everything's green, everything's passing. At this point, I'm just picturing Raj like flipping the table and being like, like rage quitting and being like, I'm done. I am so done. He goes to his fellow managers again and is, listen, this has reached like a critical mass. Like the only reason why— he probably didn't say this, I'm now writing this in— he's like, the only reason we're still in business is because I caught this bug. I mean, it would've been a terrible—

yeah, no, that would've been security-wise horrible. I don't know what kind of data they storing because if it was something like very sensitive, yeah, this could have. Damaged the business significantly.

Yeah, 1000%. And he escalates again to leadership and they're just, I don't know what to tell you. Raj then escalates to the director. Unfortunately, the director also doesn't offer any support. Instead, the director poses a question, which is like to the effect of, I'm wondering if perhaps we're looking at this the wrong way and maybe it's possible to restructure the code base to make the AI more effective.

What? Not the reframe.

Yep. Putting it back on you.

Have you considered changing the entire code base?

Mm-hmm.

Like, oh, okay.

So I thought a goal was for us to move faster on the things that are important to this business. Now you just want me to go in and set up a whole initiative to restructure the code base? What are we doing here?

I just feel like people should read their code. Like, I just don't think that's like a crazy. Statement.

I agree.

Say, hey, use these tools, but also know like you gotta still need to use your brain and like actually read the code before you ask someone else to review it. I just, it's not like really a crazy statement to me. I don't really, to me it's very simple.

Like it's just, I agree. I agree. I don't get it. There's some reason.

Also, you have to prompt that stuff and it asks you to accept The suggestion it's making, it's not like the AI just, if you don't know, if someone's listening, doesn't know how these tools work, like it literally is going to say, hey, do you want me to do this? And you go, no, robot, that's wrong. Next. Are they just like auto-accepting everything? Because if I auto-accepted everything, this company would've been down months ago. You know what I mean? We would've, this whole place would've been Not with my PR reviews, Jasmine. Right.

Sure.

Yeah.

Yeah.

Yeah.

We have AI coding tools also reviewing our PRs, but we still require a human being.

We do require a human being, but I find it helpful sometimes. It catches things we don't, but most of the time it's still wrong. Like it'll be like, hey, you should probably do this. And I'm like, you're getting outta here. But no thanks. Yeah. It's like tapping it on its head. It's like, it's okay. Yeah. Yeah.

Yeah.

You did enough today. Yeah, you, you get better tomorrow, but restructuring your whole code base, I don't know. I do think like maybe figuring out ways to make it more efficient is good, but I don't think that's the first step. The first step is like getting people to use these tools more effectively. That is a wild jump.

And unfortunately this is where we are ending because really this is actively happening. Wow. So Raj is basically like, I'm super burned out, my whole team is burnt out. And yeah, if we have an end note here, I'm curious, like, what would you do at this point?

First, I would get an auto-reviewer. I would put that in to combat the AI with AI. Yep, that's step one. Think that maybe I'd put— propose like a checklist on there. Also figure out how to create like a training or something. I don't know how much Raj is using these tools, but figuring out what is the best practice and find a way to influence. I would also say metrics are really helpful in this case because I think it would help storytell to directors or to senior leadership if they're not already looking at this, where you can say, hey, look at the significant amount of like cycle time, like the cycle time, overall cycle time, maybe the earlier part with the coding part has been reduced, but look at the great impact it has to the review time and look at how this has impacted overall. Like now it's taking 5 more days than maybe previously. This is why I'm excited about the work like we do, 'cause I think we're specifically trying to help people measure these things so that, 'cause it just, like I said, I think at the beginning of this is like just adding AI just isn't a strategy. It's not like it requires someone to be thoughtful and figure out how to make a successful integrated engineering organization, you need that.

Yep. Yeah, no, I love your answer and I think reading Raj's sort of impression of all of this, where I definitely disagree with him a bit, he seemed to blame a lot of this on AI. The AI did this, the AI did that, and removed the human element from it. And so I really like your emphasis on training. And actually helping people understand how to use these tools effectively. Because you give this tool to a completely different organization who has gone through training, they are more disciplined in how they approach integrating AI into their work, and they are moving faster. They're crushing code, they're meeting their outcomes, their product outcomes, like business outcomes a lot faster. And so I think that is like the key element here is it's actually, of course AI is going to make mistakes, but it's actually humans here who are doing a lot of them. Fucking up, for lack of a better term. Yes.

Yeah. And I think this is gonna continue, right? I think more tools are gonna be inspired by AI or AI, like AI is going to touch our entire tool set eventually. Yeah. So I think that this is gonna continue to evolve, but at least as it stands today, it doesn't replace us yet. I say yet, maybe it never will, maybe it will. I don't know. I tried to make it replace me. I tried. I tell you this every day. I try. I don't want, all right, code the whole thing for me, do it. And it doesn't work like that. I would do a lot less work. I'd do a lot less, more other stuff, but we're not there yet. Maybe one day, but we're gonna have to continue. I think everybody who's an engineer now and who's a leader, we're always learners, but we especially now are all students at this point. Everyone's on an equal playing field for the most part.

Yep.

And figuring out how to evolve. I love the story that you found because this is exactly like the negative stuff that I think happens when trying to introduce AI, but it's not AI that's the problem. Right. It's people. It's just like very vague intentions of what to do with AI and very like less, not teaching people what to do, what's right, what's wrong. And then people just kind of being like, well, now that robots are here, I might as well not do anything. Like, no, it doesn't work that yet.

No, we still need you to do, do your job and Please read.

Isn't that too bad?

Read the code. Darn.

Do you feel like you speak less to your, to other engineers, like to collaborate now that we have AI?

That's a good question.

Like ask for help less, things like that.

I definitely think that's a thing that's happening across the board. It's probably true a little bit for me, but I know for a fact that if I'm like looking back to when I was a junior engineer and I was like struggling between between Stack Overflow, Reddit, Googling random things, and bothering people, I know for a fact that I would have talked to way less people having AI. Even now, I mean, sometimes I'll encounter a new thing and yeah, I'm not going to bug somebody. I'm going to be like, explain this to me like I'm 1.

And then I'm like, LOL, just kidding. Can you do it like I'm 12? Because The one is like way too simple.

No, yeah, it doesn't work. But yeah, I don't know. Do you feel that way?

What you said resonates. I think it's like all the time I spent researching on, remember when we say like half of our job as an engineer is just being a good Googler? Mm-hmm. Now it's like half of our job isn't being a good Googler, it's being a good AI prompter.

Yep.

So it's like the same concept, it's just a different interface and it's a lot quicker. But we still gotta do the actual thinking and the thoughts and the putting stuff where we want, and we gotta be targeted about it.

But AI doesn't see the vision like we do, right?

5 steps ahead. It only sees what it's doing right now, the task. It's, it's a baby junior engineer. Thanks for listening to Off the VPN.

If you have any juicy tech stories or some burning questions that need answering, visit jellyfish.co/offthevpn. To share them with us.

We want to spill them on Off the VPN, and don't worry, we'll keep it on the DL.