Era of AI slop cleanup has begun

From hype to hard lessons in AI engineering

Hey everyone,

It’s been a while since my last update in July. Between work, writing, and life, I wanted to make sure I had something genuinely worth sharing before showing up in your inbox again.

I’ll keep this short and focused and thanks for sticking around.

Well I was recently talking to couple of my friends/collegaues/founders about AI and one thing is very common ‘era of AI slop cleanup has begun’

one of them is in EU and work as a freelance guy and he said I’m a freelance software engineer with about 10 years of experience work with startups. All of a sudden, he have a pretty steady flow of referrals. he don’t take every project on and not every one works out, but enough do that he can do it more than full time.

Lately, he have noticed a large increase in projects where startups paid a ton of money for an AI software and it does not work well at all.

Tons of errors, unreasonably slow, inefficient and taking up a lot of resources, and large security flaws.

It might be a few years before we start to see this on an enterprise scale, but he noticing this becoming a serious problem for small businesses and startups, especially when founders / people are in charge aren’t technical enough to identify this ahead of time.

Another guy worked in SF was like ‘Oh yes 100% my company just did this. Some designer who knows CSS but not much more created a full on React app using cursor and code was a mess so they hired a freelancer React guy to fix it up and he just removed 90+ files out of 100’

I’m not surprised at all. AI is trained on all the publicly available code.

So take all of that code and get the average and that’s what AI is using to generate code. As a professional developer into my second decade of coding I can safely say that most of the code I see is bad to mediocre and less than 10% is good and a smaller percentage is excellent.

It’s absolutely no surprise that AI produces almost all bad to mediocre code in large volumes.

Personally I see multiple times when I ask coding agents to fix a bug, they will delete the method in question especially if it is not referred anywhere else in the code. The concept of public API is a bit unknown and this is with prompts saying don’t delete code.

And one of the funninest this is the number of times ChatGPT has suggested using “setTimeout” unironically to fix a legitimate issue makes me terrified of what I’m going to find in codebases over the coming years.

We’re at a point where devs are feeling pressure to use AI to deliver, however may not yet have learned how to use it effectively to maintain code quality (or when to use it and when not to use it).

Often people find it easier to write code than to read it. It’s one reason why developers hate legacy code and want to rewrite into the way they prefer it.

If someone has been the one to write the code, it is easier to debug, or evaluate additions to the code. They know it more intimately. Or they can ask questions of another person who wrote the code.

However trying to make sense of buggy code written by AI will be more challenging.

• I trust ai to explain code well.

• I trust ai to read docs and find stuff for me.

• I trust ai write boiler plate scaffolding code.

BUT I never trust it to write core functionality.

And until we teach it to distinguish good code from mediocre code, I don’t really see it getting better anytime soon.

I see the same issues with all LLMs.

Indeed reviewing code is often more difficult and less rewarding than writing it yourself. If you have to fix AI-generated code, this is a shitty job because the “creative” part was done by AI, while the tedious part is on you.

It’s the exact opposite of what we were promised.

LLMs make it easy to write code but aren’t as good at refactoring and maintaining a cohesive architecture also apart from general maintainability constraints this will hurt the use of AI tools in the longterm because more repetitive code with unclear organization will also trash the LLM’s context window.

I think code reviews are a good place to push back against AI excesses.

• A function is too long?

• duplicate code appears in three places?

• code lacks clear architecture?

We should always go back and do refactoring.

And you know one of the problem here is that a lot of the design work is moved from the developer to the reviewer and a dev with a Cursor subscription can overwhelm the team’s capacity for reviews (especially as LLM generated code needs more review effort).

But there are some techniques I recommend to all devs, whether AI tools are involved or not:

• do self code reviews before requesting peer code review or after raising a PR.

• use automated tools to check for common problems. This is highly ecosystem specific, but linters, type checkers, and compiler warnings are already automated reviews.

• be sceptical if modified code is not covered by tests.

• try to strictly separate changes that are refactoring from changes that change behavior.

Or as Kent Beck quote goes: ‘first make the change easy, then make the easy change’.

This must drastically reduces your review effort and helps maintain a cohesive architecture in long term play.

tldr: I pre assumed that you already have a healthy code review culture because ‘code reviews’ are a good place to push back against AI excesses.

My current AI coding loop:

• Claude code/Codex opens a PR

• CodeRabbit reviews and fails if it sees problems

• Claude/Codex or I push fixes I repeat until the check turns green and merge.

You can go very fast with AI, but you accrue technical debt at a much higher speed too. Few people bother with technical debt but it’s not an abstract concept.

I have been telling people from the start of this whole AI thing, that I’m happy keeping my development skills sharp. There will be a huge demand for people who can clean up the mess and I’ll be happy to help with it for a good price.

I’m curious to hear what AI tools you’re using these days and what your experience has been so far.

By the way, I recently started Creator Partnerships to help startups connect with technical creators YouTubers, newsletter writers, or anyone with a strong distribution channel. If you think you’d be a good fit, just reply to this email or send me a quick DM with your audience details and expectations. I’d be happy to explore how I can sponsor you or connect you with dev tools companies looking to collaborate with authentic creators in the community.