Vibe Coding Is So In 🎶 But It's Not All Hunky-Dory

Roughly two months ago, Andrej Karpathy floated out the idea of a new kind of coding that has cropped up in existence — “vibe coding.”

Vibe coders embrace giving up control to the AI; share their product vision, give out guidelines in English, watch the results come in, and “accept all” the code written by the AI for them.

Now, there are competitions being organized with millions of dollars in prizes up for grabs for best vibe coders out there.

Setting the Stage

Most expert developers have been of the opinion that AI just can’t write expert-level code (although more and more are changing this view with time).

So, this approach coming from Karpathy may have been a surprise to some — him being one of the most respected engineers in the world.

And if you ask Karpathy, who co-founded OpenAI and for a long time led Tesla’s AI and self-driving initiatives, LLMs’ ability to code will only get better from here.

The thing with vibe-coding is it opens doors for people who have neither any experience with coding nor any interest in learning it.

The second bit is important because it highlights the difference between vibe-coding and generally AI-assisted coding.

Coders generally use AI assistance to quickly write repetitive or mundane stuff.

Say when coding in Django, I need to frequently write view functions to render a template with some data from module — and I just use AI to write these now, and save myself 10-15 minutes each time.

My Snakes game and Artificially Boosted landing page are all vibe-coded, while I would never ever leave the driver’s seat with my main projects Dzambhala or TushitaAI, even when using a lot of assistance from AI.

The Hottest Programming Language To Learn In 2025 Is English

I think it is understandable why top engineers don’t think AI is particularly helpful for coding in practice. Same would apply for a top journalist thinking AI can’t write a good column or a top tax lawyer think it isn’t helpful for legal research.

Yet, I think all of these critics put emphasis on the wrong point.

Take me, for example.

I never let AI write my words. It is a craft I have honed over a lifetime and it just can’t compete with me or even better me when it comes to conveying my thoughts in way that I can, with emotions and a sense of emotional quotient (although I love GPT-4.5’s attempt). But it does serve as a good aide in helping me find the right direction, or help tackle writer’s block.

As a writer, I don’t need chatgpt to “vibe write” for me — it would be darn offensive.

Yet, ChatGPT is why my family members don’t ask me to write their emails any more, and I am very thankful to it for that.

At the same time, I don’t have any issues with AI generating images in entirety for me, or vibe-coding an entire landing page. Those are not areas I am expert in or have any emotional attachment to — so it saves me a lot of time, effort and money to have it do those tasks.

So, in a similar vein, I think while those top developers might not think much of vibe-coding, but it would have very real repercussions for entry-level and lower-skilled devs, who are simply not needed anymore. AI can code the mundane, repetitive and at times, even very creative stuff just fine.

Why Some People Are Struggling With Vibe-coding And Giving Up

The biggest challenges non-coders face when starting with vibe-coding is:

1) Codebase getting too large that AI becomes ineffective at keep going.

2) security of their deployed app

There are a couple of example floating around on Twitter, with people who took entire vibe-coded projects live, getting hammered by cyberattacks.

The first one happens due to limited context windows in models. There is only so much information an LLM can remember at a time.

While most LLMs are seeing this context window rise, that doesn’t necessarily bring solution.

There are a lot of internal mechanics involved and the more “context” an LLM has when responding to a query, the more ineffective or imprecise it can be, as a result.

The solution here is to keep code fragmented, and only vibe-coding small projects to begin with.

So, it is completely fine to vibe-code a static website or game, but don’t keep hopes of being able to code an entire social media app or a finance terminal through this process just yet.

And learning git control can be an important step here, so you have a better ability to “reverse” the code, if something gets ducked up.

Security concerns are even more tricky to deal with — and can happen to anyone, vibe-coding or not.

I am for example going through a subscription form spam attack for Dzambhala right now, and working through adding more security layers to bring about a fix.

Typically, security issues happens on front of you telling the LLM the API key, accidentally exposing it directly in the code, not having enough spam detection mechanisms or database security.

If you are coding static websites, this is something you have to worry about very little. Entry for cyberattacks typically come from javascript, dynamically-rendered html or API keys.

From Around The Web

Levelsio is organizing a competition to vibe-code the best game with AI. Judges include Karpathy, the man himself, and a couple of other leading game developers

Give it a shot if you ended up vibe-coding a game!

Bolt.new is also organizing some hackathons for vibe-coding, with cash prices!

I am continually working on ways to improving the value this newsletter brings to audiences of different calibre when it comes to using AI! If you have any feedback, please share anytime!