Is "Vibe Coding" Working?

#ai #machinelearning #openai #discuss

We just launched an AI backend code generator. We had to because there's simply no existing LLM out there that currently understands Hyperlambda, so if we want to have AI-assistance in our platform, we had to create our own model by fine tuning an existing.

We created it by fine tuning OpenAI's GPT-40-mini model. It took us a month to create enough training snippets to have the thing do the right thing 80% of the time, but I'm fairly confident in that I can bring it up to 99% over the next couple of months by simply incrementally adding to its training material. You can test it here, just please remember to start your prompts with "Generate Hyperlambda ..."

Below is a screenshot of me testing it ...

For the record, it did a perfect job for my prompt above.

Will "View Coding" replace software development?

I am supposed to be running around screaming out of the top of my lungs that vibe coding is the new thing. According to what I've seen related to AI and FOMO, I'm supposed to lie through my teeth, and convince a million gullible people to buy my product by over selling it like crazy. Unfortunately I'm not created that way, so I'll just tell it to you like it is instead ...

... and yes, vibe coding is working, but only for software developers! I am 100% confident in that people without prior software development experience can create amazing things using "vibe coding", but creating something that "works" and creating something that's secure, scales, and stable, are two different things.

Alex, one of our community members just sent me a prompt he had tried, and I'm not going to repeat it, since it's very long and only relevant for his database schema, but basically the prompt went like this.

Create an API endpoint allowing users to register, storing users into my database, and send them an email afterwards. Make sure you hash their passwords and that their passwords are at least 12 characters in length.

I studied the code, and it was about 90% correct. It's still hallucinating some functions and constructs, but the number of changes that had to be applied were not really that many.

This was the result we accomplished with 3,500 training snippets fine tuning an OpenAI LLM. These days we're adding about 100 additional snippets per day. When we get it up to 5,000, the theory is that at that point it will recognise the programming language completely by itself.

So my verdict currently is that it makes an existing software developer about 1,000x more productive, but it simply can't be done by non-software developers - At least not yet! And I suspect you'll always need to know some of the basics, regardless of how good it becomes.

The reason is that for Hyperlambda, at least today, you still have to know what an API is, what an HTTP endpoint is, the basics of how HTTP and HTML works, etc, etc, etc. Today it is simply not possible for a non-software developer to even ask the correct questions - Let alone sanity check the code the LLM produces.

Is "Vibe Coding" a threat to your job?

That's a different question, and my answer is "it depends". If you embrace it, and you're already a seasoned software developer, vibe coding will make you capable of delivering 1,000 times faster. Obviously this will make you "more valuable", since you can outcompete others competing for the same jobs. This implies vibe coding would make you more valuable in the market place, resulting in more job opportunities and offers.

If you have no knowledge about software development, you might be able to create an MVP that's working "most of the time" - But to believe you no longer need dev heads because you've got AI is at best delusional. And I suspect that regardless of how good vibe coding becomes, there will always be a need for people understanding what the generated code does, and improve upon it manually.

Will it steal your job in the future?

That's another question again, and an honest answer here would have to be "I have no idea". However, we've got a plan at AINIRO for a process that implies creating a RAG-based database of all system Hyperlambda files, for then to wrap these into AI functions, allowing us to have an orchestrator LLM we can feed entire specifications to, which then automatically starts assembling complete products.

Once we are there, we can literally turn on AI summaries when we're having sales meetings, make sure we ask the right questions to the client, ensuring the AI summary contains all crucial points from their requirements.

This results in a 2 to 5 pages long summary, which we then can break down into individual tasks. These tasks can then be fed into our "orchestrator" LLM, which starts assembling projects entirely autonomously by creating files, and generating the required code.

Now we're basically at the point where we can have sales meetings with clients, turn on AI summaries, and once the meeting is over, we feed these AI summaries to our orchestrator. An hour later, and the entire solution is assembled and production ready after one of us having sanity checked its code.

I can have about 10 sales meetings per day. If we pull through the above, we'll have an automated AI-based pipeline, that basically delivers entire projects from A to Z. These are projects other outsourcing companies will charge you 10,000 to 100,000 for may I add, while for us it took us zero minutes of our own time to actually deliver it.

However, I'm still not sure it will actually steal your job. First of all, the size of the market will inevitably explode. Secondly, there will always be a need for a human being hand holding the AI I suspect, to sanity check its result.

Should you embrace vibe coding?

Yes, yes, and yes! If you're an existing software developer vibe coding will make you literally 1,000 times more productive. If you've never done any software development previously, then vibe coding can be an easy introduction to software development for you, since you can also have the AI explain its own code, making it easier to understand concepts and ideas.

Vibe coding works is my conclusion, but it's not a threat to your job, unless you ignore it - At which point that's on you ...