You may not have even heard of “Vibe Coding”. It is my understanding that it is a term to describe people that delegate their programming to an LLM.

Why I am not using it….yet: There are a couple reasons. While I am concerned about the legality and mortality of using AI trained from questionable sources to generate intellectual property that needs to have a clearly defined owner(Either mine or my clients depending on the situation) that is only a small part of why I have not jumped on the bandwagon quite yet.

So far I just have not been that impressed with it. I have used AI quite a bit at this point. I have tried using LLMs for things like landing pages and front end stuff. I have even tested having the LLMs write terraform to see how secure my job is. Spoiler: It’s safe… for now.

What you have to remember about an LLM is that it is a parrot. You feed it “Mary had a little…” and it knows the next word out of its mouth needs to be Lamb. But it doesn’t know why. It has little actual knowledge about the real world. It can only predict the next character in a sentence based on the inputs passed into the neural network. That is a fairly linear process. Basically it is 1 dimensional thinking.

When I design something for AWS it is not a 1d, 2d, 3d but closer to 4d. It's a living system with a /heartbeat and all. Information flows in and out in varying volume and speeds. There are millions of requests per hour getting routed in dozens of different ways to hundreds of possible pieces of hardware based on thousands of rules programmed in code.

I’ll be honest: I struggle to convey my designs to other software engineers at times much less via a 1 dimensional series of characters(a prompt) to an LLM (That is partially why I created my Animated Network Diagrams).

When I recently went nuclear with a client and started slinging code for a massive search engine project I had to take into consideration so many variables all at once. The code needed to take a very sophisticated thing and make it simple. So simple that junior devs could maintain the system once I went back to strictly advisory(no more coding) while at the same time being incredibly scalable in a cost efficient manner. There is a lot that goes into designing a system of this magnitude, more than I think current LLMs are capable of; at least the consumer grade LLMs I have access to.

Am I saying I will never use LLMs to design highly scalable infrastructure? Not at all. I am just saying they are not there today and LLMs may not be the AI technology that puts me out of a job. Perhaps a classifier or a GAN of some type, but LLM? I am just not sure they are the right tool for the job.

Also to be clear: I am not looking down my nose at people that use AI to assist them in their coding. That would be like old school programmers looking down on me for using code autocomplete back in 2016. Knock yourself out but make sure you 100% know what anything you push up to production does. You are still responsible for the code you push even if you didn’t write it.

Question: How is AI augmenting the way you spin up production infrastructure or write code?