Ai Pilled

I drank deep of the AI koolaid. I’ve changed a bit since. (but not much)

I wrote my post last week after a day of marvelling at the wisdom and wonder of an AI coding agent writing a bunch of code and magicking my vision from thin air into reality.

This post was written a day later, when I spent more time with the agent and spent downtime browsing twitter. I’ll leave you to guess which one was the biggest mistake, though it’ll probably become very clear, very soon.

Let’s start with the agent:

After basically one-shotting (albeit with frequent tweaking and guidance) my vision for the app, I was in awe. This was glorious! I never needed to learn to code after all! The main limiting factor was my terrible way of explaining what I wanted in a manner that didn’t sound like barely educated gibberish.

The day after, once it had built me an authentication solution, expanded on my objectives elsewhere, and assured me that the solution was super duper secure and ready to go live, I braced myself for a casual 30 mins of prompting and told me what I wanted it to do - take that nextjs app that now lives in a github repo, and build a way to package and deploy it on a remote server using docker.

“Don’t worry about proxying or anything” I said, “I’ll be using a cloudflare tunnel to do that. Easy. Just stick a Cloudflare Tunnel docker config in the docker compose, and anything crucial needs to go in a .env file so I can secure it”.

Like the cloudflare tunnel key, and my mailgun SMTP credentials.

“Sure” said the intelligent AI “tell me all the details, and I’ll put them in the .env file for you”.

NOPE.

After a bit of back and forth, it finally sucumbed to my much better idea of just creating the file, and letting me put in the details. With that hurdle over, we moved to the next point.

Four hours later…

Apparently building a docker container from the image was a lot harder than it thought. It kept tripping up on an error, then it started to get dumb… telling me that it had made changes to the app, and I just needed to “git pull” on the server. “You need to push the changes first”, says I.

We went down that route many, many times. Commands I had to run on the server would mysteriously change. It would hit the error over and over again. I got tired of copy/pasting the errors, but I persisted, and we’d solve one problem, hit another, and four hours and about £15 later it was done, and it worked.

That app took about 60 million tokens to build, I think. Not bad. Waaay faster than building it myself. The kicker was that if I didn’t understand how docker worked, or server commands, or being able to set up a deploy an ssh key on my server to do a secure pull from an ssh protected repo, or a ton of other things that i’ve picked up by being an active partner in this, then I would have been lost.

I need to be clear - agents of today are way better than the ones I was trying to use a couple of years ago. Those ones? They were dogshit.

But if you listen to the twitter AI bros, we’re all finished.

Our jobs are done.

The people whose livelihoods depend on us burning dem tokens are telling us we need to burn more.

What the actual fuck?

I’ve worked with these tools now - I’m by no means an expert, but I’ve done enough to understand where they work and where they don’t - and I’m comfortable in saying that it’s going to be a long time before I have to worry about my work being done by a robot. Bleep bloop bloop and all that shit.

My job is essentially selling through writing, and constant innovation and creativity comes with the job. Repeating what worked the last time don’t mean shit. In my sector, all the AI bro, VC backed companies are selling AI tools but the bit they deliberately miss out on is the one where the magical AI needs to slurp up all the good content - the stuff written by humans for decades - in order to be any more useful than a wet fart in a spacesuit.

But because they are all VC backed, and burning through cash as fast as I’m burning through tokens, they have to sell those big expensive subscriptions.

I thought “fuck that”, and I’m building my own tools, instead. Not the ones that do the writing, but the ones that cope with my dumb brain when it comes to meaningless admin.

I’m spending my time using my AI assistant to do all the trawling through decades of content, and find the stuff I want it to find - that is ultimately curated and assembled by me, to form the pitch I want to form, in a way that I know will meet the needs of my clients.

I specialise in the human element.

If you do anything more than just assemble tiny chunks of meaningless code, then you are going to be fine for a long, long, time. Us humans are more creative than the robots give us credit for, and the AI experts have forgotten what it’s like to be human and embrace that humanity in millions of different ways, in millions of different jobs.

What you do need to learn is how to harness those tools to make your job easier. That will be easy, and I’m doing it right now.

But our AI Overlords? They are a long way away at the moment.