The blog post that got me to put my words down.

Last year I decided to give AI a shot. Previous to that, I had dabbled with the chat interfaces of Claude and ChatGPT, but had resisted really diving in.

Just so no one doubts my ‘allegiance’, or to brand myself into the camp of water-boilers, over that time span I have:

  • Fully vibe-coded a web-rtc baby monitor (flutter app and Go backend)
    • I’ve vibe-coded or AI assisted many other projects, but that was the most substantial project that works.
  • Paid for various subscription services (Claude) for said vibe coding
  • Used tons of chatting with AI to plan projects, make recipes, research things on my mind
  • At work, used Opencode to complete many of my work tickets, typically iterating in Plan mode, then executing said plan
    • No Pr’s were ever merged without some minor intervention as far as I recall, there was always a little something here or there that needed tweaking, but it definitely felt very productive.
  • See-saw’d every week between “this is awesome” and “I’m pretty sure I’m getting dumber and need to stop”

I feel I have fully submerged myself in the slop swamp. What have I learned?

LLM’s are Excellent

Large Language Models are really cool piece of technology. I hope one day there is a fairly and freely (as in libre) trained model to replace search. It is so nice to “describe” what you are searching when you only have the vagaries in your mind. Turning English text prompts into random statistical output is a wildly cool idea.

Unfortunately

It is so very problematic.

LLM’s are Evil

I think my experiment with going all-in with LLM’s has taught me just how insidious this technology can be. I believe it has wreaked havoc in many peoples’ lives (Wikipedia). I feel that the price a software engineer pays to have a prediction engine produce the code for them is steep. It erodes personal expertise, cheapens the craft and love of the craft, sucks the enjoyment out of work, and steals the ability to make any large maintenance decisions with confidence.

I feel that my original aversion to these programs was right. The technology inserts itself (by default) in the loop, creating dependence.

LLM’s do not produce Quality, they cannot. So, I think I am done with AI code assistance. Rubber ducking and search still seem very helpful to me, so I’m going to try and move to self-hosted models for those use-cases.

I want to continue to improve my craft. I want to love this craft. It’s why I started down this road nearly ten years ago in college.