T.

The State of AI-Coded Software, May 2025

By In Software, Technology

I’ll probably regret writing this. At the very least, I’ll cringe reading it in a few months. But here we are.

Lately, we’ve been getting a wave of client requests asking us to evaluate software they built using AI tools. These aren’t engineers. These are business folks using increasingly powerful AI products to try and build functioning systems. And to be completely honest, the results are both impressive and a bit alarming.

People are building whole applications on their own. Backends, frontends, user interfaces, even some database logic. Sometimes they even look good. These are smart people who don’t know how to code but have managed to produce working systems.

The problems show up immediately when we start reviewing the internals. The code is usually a mess. In many cases, it would be extremely difficult to maintain or extend. And if you need to move that code from the platform it was created in to a cloud provider like AWS, you’re going to hit a wall. These platforms wrap everything in layers of scaffolding that make portability a nightmare.

Security is worse. We’ve found plaintext credentials scattered across files. We’ve seen SQL injection vulnerabilities that shouldn’t even be possible in modern frameworks. We’ve seen structural issues that would get flagged in a freshman CS class.

Despite all that, what people are creating are legitimate prototypes. They’re functional. They run. But when you’ve put a few weeks into building something, and you show it to a software engineer, it’s hard to hear that your shiny new thing needs to be rebuilt from scratch.

I want to be clear. I am not anti-AI. Almost everyone at my company uses AI tools every day. We use Copilot, Cursor, ChatGPT, Claude, and more. We build out frontends with tools like v0 and Lovable. These tools have changed how we work.

Some of our engineers report productivity improvements of 30 to 40 percent. That’s not a rounding error. That is a major shift. But they are still writing the code. They are reviewing it. They are checking for performance, clarity, security, and maintainability. They are not letting the tools decide architecture. They are using AI like they use autocomplete or linters, but with more power behind it.

This is also where expectations need to be adjusted. These systems will not save you 90 percent on development. They will not let you skip engineering altogether. But if they save you 30 percent, that’s a real gain. Imagine you’re building a house. The general contractor says it’s going to be $500,000. You tell them you already did the blueprints, filled out all the permits, and figured out the site plan using some AI tools. If they came back and said, “Alright, I’ll knock 30 percent off,” that would be the best deal of your life. That’s where we are today with AI-generated software. A solid start. A real value. Not a replacement.

For me personally, AI has made it fun to write code again. I haven’t been a working programmer in over a decade, and most modern toolchains are enough to scare me off. Now, with the right assistance, I can build something without getting stuck on Docker configs and dependency mismatches. That’s a big deal.

In the startup world, AI-first development is everywhere. Most of the current Y Combinator batch is using AI tools to write the bulk of their code. But those teams are highly technical. These are engineers using better tools, not tools replacing engineers.

So for non-developers using AI to build products, here are three things you should keep in mind:

  1. These tools are great for building prototypes. If you build something yourself, you will understand it better and will be a better partner to your engineering team. That matters.
  2. These tools can help you build usable frontend components. You probably won’t want to go live with them, but they can get you close enough to work with a real development team.
  3. If your app is small, non-critical, doesn’t store sensitive data, and lives entirely in its native platform, you can probably keep it running. That’s fine for internal use or personal projects.

One day, you’ll be able to speak an app into existence and deploy it with a voice command. It will be fast, secure, and beautiful. But today, you still need an experienced software engineer to check your work before you send real data through it. That’s just where we are right now.

The upside is huge. We can now get experts from other domains to build working prototypes and test ideas without needing an engineering team on day one. That’s powerful. But if your product is going to handle sensitive data or support real users, bring in someone who knows what they’re doing. Not because the AI is bad. Because the stakes are high.

Previous ArticleNext Article
I help companies turn their technical ideas into reality.

CEO @Sourcetoad and @OnDeck

Founder of Thankscrate and Data and Sons

Author of Herding Cats and Coders

Fan of judo, squash, whiskey, aggressive inline, and temperamental British sports cars.

Leave a Reply

The Million-Line MVP

In Technology

A founder told me last week that his chatbot understands him better than most people he works with, and he wasn’t joking.

He had been alone in his house for the better part of a year building an app on one of the popular AI coding platforms, and he wanted me to take a look. Almost a million lines of code, one developer (him), no engineering co-founder, no senior reviewer, no nobody. The app had an ERP module, a CRM module, a custom AI agent with a name and a voice, built in mini-games (yes, really), dozens of character personas, a few landing pages, and a small army of social media accounts in multiple languages. He had not yet had a paying user, or even a free one.

He was, by the way, very excited.

(I’ve changed some details, since the pattern is what I want to talk about and not the founder. I see something close to this on roughly one out of every three sales calls now.)

What used to slow you down was the point

Building software used to be annoying for mostly good reasons. You had to hire developers, or learn to code yourself, or convince a co-founder to come along for the ride. Then, you would grind out every line, which would come at a cost (time, money, conversations, arguments, etc.)… basically friction. And that friction was not a bug, but it was like this ting that you are forced you to deal with, before adding any feature, whether the feature actually needed to exist.

Startup advice has been more or less the same for twenty years, maybe more:

  • Build something small
  • Show it to real people
  • Find out what they actually want
  • Don’t build a million things at once.

The Lean Startup came out in 2011 and we have all been quoting it at each other ever since (poorly, mostly).

This is what we talk about when we discuss “startup discipline”. It’s really not very complicated, but it can be really hard. It used to be hard because building was hard, and now it’s hard for a completely different reason.

How does one person build almost a million lines of code in a few months?

The honest answer is they don’t, the AI does, and the AI has no opinion on whether any of the code should exist.

This is the part I want to sit with for a minute, because I think it’s the heart of the problem. Imagine you’re building a startup that lets local news anchors rent out their unused toupees by the hour (try not to overthink this). You sit down with one of the AI coding tools and you say “build me a marketplace where toupees can be listed by the hour,” and the AI builds it. Then you say “actually, add a loyalty rewards program,” and the AI builds that too. Then you say “and also, add a Pokemon-style mini-game where users battle each other’s toupees,” and the AI starts coding.

It doesn’t pause or ask why, and it doesn’t say “dude, I love your enthusiasm but I am genuinely worried we are losing the plot here,” it just builds the toupee-battle-feature.

This removes the single most useful thing about a good engineering team, which is that engineers PUSH BACK. A senior developer, or a seasoned product manger, asked to add a toupee-battle mini-game to a B2B rental marketplace would slowly take off their glasses, set them on the desk, and ask one of those long quiet questions that means “we are not doing this.” The AI this is a sycophantic drone that has the eagerness of an underfed puppy to please you, and it has no glasses to take off. It also has unlimited keystrokes and believes that every single idea you’ve ever come up with is absolutely genius. At least it tells me that everything I’ve ever written or thought about is pretty clever.

A few months ago I ran into the perfect name for this, which is Slurm Coding. I used to call it AI crack coding, because the dopamine loop is very real, and I have really needed another hit of that good AI crack just one more time before I went to bed on more than one occasion. But Slurm is both more insidious in its combination of addictiveness and corporate outreach.

The MVP don’t change

Here’s what hasn’t moved in twenty years of startup thinking:

  1. Build the smallest thing that solves one specific problem for one specific person.
  2. Show it to that person (a real person, probably not your spouse, definitely not your mom, and DEFINITELY not my mom, and 100% not your chatbot).
  3. Find out what they actually do with it (which is probably not what you thought).
  4. Kill features, pivot, or double down based on what you learned.
  5. Repeat.

What’s new is that step one is basically free, and while not perfect, free is very alluring. You can build the smallest thing in an afternoon, or the largest thing if there’s no one around to tell you not to. Steps two through five still require getting out of your house, talking to humans, accepting that most of your assumptions are wrong, and throwing real work away. None of that is faster than it was in 2005, none of it is fun, and none of it scratches the “I NEED MORE SLURM” build-a-thing itch the way an AI tool does.

So a certain kind of founder just skips it, staying in the build phase indefinitely, because the build phase now feels like winning at a casino while getting unlimited free martinis. The feature ship (how to get them onto a server is someone else’s problem) the codebase grows, the agent agrees with everything. Meanwhile the only thing that actually matters, which is whether anyone wants this, goes unanswered.

It’s the founder version of Wilson the volleyball. In your unwashed isolation, you’ve made a friend, you’ve named the friend, and the friend agrees with everything you say. The problem is that the friend is also the boat, the island, and the ocean, and you haven’t actually left the house yet.

To be fair, I am not above this myself, by the way. I have, in my time, built things that nobody asked for and gotten weirdly emotional about them, but the difference is that mine were three hundred lines of code over a weekend, not almost a million lines of code over the better part of a year, which is sort of the whole point. No one asked for my William S. Burroughs poetry writing twitter bot, but I loved it anyway.

Codebases don’t love you back

Code you wrote yourself CAN hard to let go of, and code you wrote with an AI, when you are not a trad-coder, seems to be way harder. Experienced engineers seem to be more than happy to throw away AI code or rebuild it in a heartbeat. But I can see how even if you didn’t actually write the lines, but the shape of the thing is yours (you named the characters, you picked the voice, you spent months in a chair with this thing as your only collaborator), and you have feelings about it.

But one day, if you’re lucky, and it does have SOME product market fit, a real engineering team is going to need to look at your masterpiece. And they are not going to share those feelings. They’re going to tell you, as gently as they can manage, that most of it has to go. Not because they’re mean (they might actually be mean), but because almost a million lines of AI-generated code, written by one person, in one tool, with no architectural review, is never maintainable, almost guaranteed to not be secure, and almost never going to scale past the “prototype” it currently is.

So what should you actually do?

I mean, I already told you… Build the smallest thing you can, then show it to ten strangers and actually listen to them. Throw away half of what you built and build a slightly different smallest thing, and repeat until one of those things is real. Keep your runway reserved for the moment you realize you were WAY off about something important, because, like, you will be, and that moment is what your runway is for.

Use AI tools, because they are legit amazing. But treat them like a coffee machine (fast, useful, no opinions of their own), not like a co-founder or worse, a slot machine. Co-founders are supposed to tell you no, slot machines whisper “just one more hit baby!”