As some of you may know, I'm a developer. I do this for a ridiculously long amount of time, spreading across almost every area of coding, from Cobol to Swift, from mobile apps to blockchain dapps.

Recently, I got into what's called "vibe coding". It's a conversational way of building software, in which you chat with an LLM (you may know this LLM as ChatGPT, or Claude, or Gemini, but they're all just Large Language Models, or LLMs) and then iterate based on the model output.

The results are, for the coding novice, astonishing.

For me, not so much.

Let me explain.

Introducing Flippando

Flippando is a Solidity game I wrote over the last 2 years. It's deployed on a few EVM compatible chains, and there's even a GNOVM version of it.

It's a simple memory game, in which you try to uncover the matching tiles on an empty matrix. So far, nothing unusual.

Once you finished a board, you can mint it as an NFT. When you do that, a fungible token, $FLIPND, is "locked" inside the minted NFT. Meaning that the token is there, but it cannot be used until, well, here's where it gets interesting, OTHER PLAYERS use your NFTs, to assemble them on bigger canvases, essentially creating on-chain art. When your NFT has been used, the fungible token is unlocked and sent to your wallet, as a reward for Proof of Attention, and the NFT is burned. It's no longer needed, as its image is now part of a bigger artwork.

If you're curious about the game, you can just pay a visit to https://flippando.xyz for all the nitty-gritty details.

But for the purpose of this article, let's stick with the key part, which is OTHER PLAYERS.

"This Is Wrong"

When I asked Gemini to give me feedback on the game implementation, to my surprise, it identified a "major vulnerability" in my code. It said: "because you do this and that in your code, users are giving control over their NFTs indiscriminately, so other players can do whatever they want with them".

Doh. That's the idea, Gemini. That's how it should be.

To make a long (and potentially boring) story short, a fairly advanced AI model, Gemini, was not capable to understand the logic of the game just by reading its code. I had to explain in natural language what is going on there and why this is actually a feature of the game, not a bug.

My point?

AI can only help you with - and validate - something that has already been done. That's how these models are trained. They are hugely knowledgeable, in the sense that they "know" what was done so far, but they don't have built in features like courage, making assumptions, etc, which will let them take risks. That's what creativity is though, doing something that no one did before, but which can still be recognized as valuable by a majority.

Creativity is dangerous, because it challenges the status quo. LLMs cannot afford to be dangerous. They're just playing it cool, offering you what's already tested and validated.

In the short term, AI will become extremely restrictive. It will limit drastically what we can build, simply because people using it will cease to be creative. They will just reshape something that other people did.

If they try something that hasn't been done before, AI will tell them they're "wrong".

In the short to medium term, we will need some sort of revolution - after an initial phase of enthusiastic adoption - against AI, learning to think again with our own heads and to own and fix our mistakes.

And that's why vibe coding kinda sucks - it can only help build you something that has already been built.