Do you wake up at night in a cold sweat concerned that you will be replaced by a bot?

Don’t worry you are not the only one of us tech professionals with that concern.

Yesterday I got tagged in a post by my esteemed colleague, Rodolfo, about a new AI tool for spinning up AI infrastructure, infra.new. I think that is the 5th tool like that I have encountered in the past 6 months.

The sales page said it could do some pretty impressive stuff, and I have no reason to doubt it lives up to expectations… with a few small caveats.

I found it interesting that under the FAQ question “How production-ready is the generated code?” this was their answer:

“All generated code is derived from the latest docs and follows DevOps best practices with proper IAM, security groups, and cost optimization. However, we still recommend having a human review the code before deployment.”

So we still have some time but there will likely be a day they change the answer to that question.

I’m not the type to shame people for using an LLM to code unless you have no idea what it did and you send it to production.

It is a tool like anything else. If you don’t understand it then it is difficult to use it properly. You don’t need to know the specifics of how an internal combustion engine works to drive a car but you need to know that if you don’t put gas in it then it will eventually stop working.

At a high level let’s break down these new AI tools.

Strengths:

LLMs by nature predict sequences. They know that if you say “Marry had a little…” there is a 99.999% chance the next word would be “Lamb”. They know that for every { there eventually must be a }.

When it comes to coding, in our case Infrastructure As Code, it is no different. The LLMs are just parroting the tens of thousands of IoC repositories they have been fed.

Sure some of these companies are getting clever and adding in a validation layer that checks the output from the LLMs against a schema to make sure there are no hallucinations (which is a really good idea by the way) but that doesn’t make them good problem solvers. It makes them a parrot, a very fast well read parrot but a parrot all the same.

Weaknesses:

While it may be good at copying IoC and spitting that out, how do LLMs fare with existing infrastructure already running? Can it diagnose problems with a production system?

I have not seen them excel at quite yet is non-sequential thinking or 4D Thinking(The fourth dimension in this case being time).

Cloud infrastructure has many different pieces all operating in symphony spread out across a vast virtual landscape. When it comes to problem solving a large production system there thousands of branching paths where the problem could have originated this gets more difficult.

So when you ask an LLM to generate a story for you there is not really a wrong answer. If it takes the main character on a trip to Africa instead of Australia, great! When finding the root cause of a massive server outage there is likely only one or two very specific causes and often do direct way to pin point the issue.

You have to trace a lot of non linear branching paths to find the issue and, unless your team is really bad at its job, the same issue doesn’t pop up twice. You fix it and move on.

Being as an LLM needs patterns I am sure it can solve a problem that you have solved a million times before but it would struggle with new ones.

With that said I will give it credit for solving common problems. One of the most common beginners issues I see is mismatched security group rules(A sort of firewall for you non AWS people). I would be eager to see how easy it is to solve that one. Also I would be curious what an LLM would use as inputs. Would sequentially feeding an LLM all the resources and their current states give it a great picture of the system? The system exists across multiple dimensions so I am unsure. The fact that I can give an LLM one of my 2D comic and I can comprehend it makes me think it is possible.

Other Types Of AI:

Now I will say it is more likely a Classifier AI could do some of this problem solving. “Hotdog” vs “Not hotdog”. “Properly functioning infrastructure” vs “Not properly functioning infrastructure”. I would wager this would be way too easy to over train.

Let’s talk about innovation for a second. Now LLMs though they may appear to be creative are not. They can just restructure and regurgitate ideas at an insanely rapid pace and it looks like new ideas but it is not. But what if we trained a GAN, the thing that generates AI art? It is possible that we could use GANs to generate new infrastructure designs at a pace that no human could keep up with. The question there is quality control. How do you score the best designs so the model that figures it out gets promoted to the next round? Sure you could write some rules to do that but we all know that real life production cloud environments get all sorts of crazy stuff thrown at it.

So unless you want an experimental AI model that is currently in training controlling your production infrastructure you don’t have a great way to validate and score the AIs in a real world environment.

Also unless you were cool with your infrastructure changes at an insanely rapid pace then it would take forever to train.

Actually that would be a great experiment to run with the domain I bought to blow up for the Cloud War Games since no one expects that to be working for extended periods of time.

What this all means to us Cloud Professional humans:

After reading all this you may think that I don’t think AI is coming for my job. That is not the case. Just like the elevator operator jobs of the early 1900s were made obsolete, so will mine in its current state. I will just have to evolve or retire and so will you.

In the meantime though things will get interesting. The lowest level jobs are being replaced and the mid level is starting to feel the effects of this as well. It is only a matter of time until it makes its way all the way to the most senior level positions.

Will they be eliminated? No, just augmented.

How do I stand out in a stack of AI generated resumes as jobs are being eaten up by AI?

I cannot tell you for certain but I can tell you what I am doing. It’s pretty obvious I am building up my presence online, my network, my mailing list, and my content backlog so if the day ever came I needed a “Job” I would stand out like a stack of resumes like a sore thumb.

Does that sound like a giant pain… sometimes it is but most of the time it is a lot of fun. Do you need a little help doing that yourself? That is partially why I created CloudWarGame.com(I also just like blowing stuff up virtually speaking). Not only to help Cloud Professionals level up their skills but also to showcase their abilities during a live War Game event that they can then include a recording of with every resume and cover letter they send out.

Conclusion:

LLMs, Gans, Classifiers, and a bunch of people spamming employers with AI generated resumes and cover letters are competing with you for your next job opportunity.

How will you evolve, adapt, and thrive? (Let me know in the comments or via email).

PS: Wow, this whole post went from a short product review of an AI tool and on to a whole thing about AI taking our jobs.

Let me know if you want me to do a deeper dive on these new AI tools. I could see it being a fun video to stress test these things by sending them through a few Cloud War Games scenarios.