Vibe Coding = New Script Kiddies

Lovable vs Vercel V0dev vs Hostinger Horizons - same prompts, which code generation tool makes more sense?

I won one of the prizes at basedspace_ hackathon last week, someone from the national TV suggested we do a reality show together and multiple people promised to repost if I put down a blogpost about my work. And since I don't have a podcast yet still would like for others to think that I am clever, here's what I did on the weekend.

Vibe Coding

Vibe coding is basically when you tell AI to make you a website and it does. Sort of does. It tries its best and works 80% of the time, but the other 20% it breaks whatever it built, introduces regressions, forgets functionality, or just goes completely off the rails.

When you want to play with it - it is fun, just like the script kiddie movement back in 2000s when the internet was young and schoolkids were causing troubles with scripts they'd find on Altavista.

But when you want to use these AI tools for production, to make something real commercially, that 20% failure rate becomes the primary challenge and can consume 80% of your time or worse.

I had Google Gemini study the scarce documentation on best prompting practices to figure out how to best command my web-apps-building robots.

How AI treats other AIs

I gave Gemini my Miro board mockup (could have also given my Figma board or just a screenshot of an app as a reference) and asked it to tell Lovable, V0 and Horizons to build the same app.

1. I got instant results with my mockup actually becoming a web app in seconds
2. I got a fully working prototype in 3 prompts in Lovable, V0 and Horizons
3. This started to be fishy when I needed user authentication and persistency (backend server functionality)
4. Then Gemini started being uneasy and SHOUTED at other AI for making mistakes or producing regressions, so I got curious to see how AI figures out ways to control and manipulate other AI. For example, after several regressions Gemini started telling Horizons how to verify that it didn't produce regressions, explained about the code quality.
5. Here's a fun moment: AI ranting to me that other AI is plain stupid:

After that Gemini decided to teach Horizons how to debug own code. It told Horizons to build an in-app debugger, read own logs, reflect on them and fox code.

At that point I ran out of credits for V0 and Lovable, so continued with only Horizons (thanks to Povilas Krik??iūnas and Ignas Gecevi?ius for credits).

Then Gemini and Horizons got struck on the backend (I used Supabase ) in the loop not being able to figure out how why it cannot upload content to the server. So Perplexity watched tutorials on Supabase YouTube channel, figured our right RLS settings and generated the sample SQL code. That's how Gemini understood what was wrong, and told Horizons how to fix it.

Celebration time

Now take a pause and think - this is Gemini celebrating that Horizons build an app with all functionality that it wanted. Well, I wanted, but I think at that time Gemini was more invested in that app than I, and for sure it put way more efforts than I.

See, AI being happy about itself. Cute, huh?

Not what you think

Gemini didn't have access to the internet, I used an isolated experimental model. It didn't visit my app, it didn't upload own test image, it didn't reload the page - it just imagined it.

That's the 20% that I was mentioning before - AI randomly imagines stuff and you have to know where to fact-check.

At this point in life my eye is trained to detect AI hallucinations, I already sense when the context gets too long and the LLM is starting to lose it, so I have to start a new chat. I already sense how to manipulate a particular LLM for the best results - "act as a ...", "user gets anxious about ...", "it's shame on you to ...".

I can do a whole lecture on it, but this is probably a story for another time.

Lovable vs Horizons vs V0dev

Technical details ahead for those interested.

What kind of code these website generators produce? Is it any good?

I used Recurse ML to check code of all three, see PRs below. The functionality of all 3 apps is more or less same. It is not apples-to-apples since each code generation tool sometimes loses it and I have to remind it of the intended functionality, ask to build-rebuild something.

For example, once Gemini asked Horizons to optimize code and make it faster, and so we spent half a day just getting back to what was working before since Horizons broke the server side and messed up the CSS while cleaning own code.

Each of AIs has a self-healing option too, but sometimes AI is not aware it broke the build or introduced a regression. It looked like it helped to tell AI what to look for in the logs, how to debug itself, etc.

That's my gut feeling since each service felt somewhat different, whimsical in own peculiar way, and I spent too little time with them.

Lovable :

1. They figured our many key issues with that "20% problem"
2. Has direct edits, Supabase and Github integration
3. Lovable code checked by Recurse ML

Hostinger Horizons:

1. Very young, easy to use, fun to play
2. I used Zyro, their AI-powered no-code builder previously and would prefer it for now
3. Horizons code checked by Recurse ML

Vercel V0dev

1. They're building a big thing in own ecosystem
2. My humble opinion: their high ambitions are a good promise, but sometime in the future
3. V0dev code checked by Recurse ML

If you clicked on these Github links and studied the code that AI generated while being controlled by another AI and cross-checked by yet another AI, then you know that it had:

• Potential security issues
• Potential crashes
• Lots of ??

And no, I don't think it is ugly, I think it is amazing.

• Remember how impressive GPT 3 felt in 2023? - now it doesn't.
• Remember how impressive DeepSeek R1 felt in January? - now it doesn't!
• I think vibe coding tools are still impressive today until they become good enough to also become ubiquitous.
• I am rooting for Horizons because it comes from Lithuania where I am based now.
• I am rooting for Lovable since is it comes from Sweden that I am missing more and more.
• Yet it's the Chinese AI that I am looking for checking next - it's Manus AI and Qwen .

Next time I can allocate time for hacking, I want to explore agentic AI and see what is possible ??

Until then, AI AI AI AI ??