i have trust issues with genAI text
I'm real tired and furious of reading AI text from someone with the message
hey, jakob, can you check if this is correct? *attached is 4 markdown files and 1 mermaid diagram that doesnt make sense at all*
My temper has gotten real short on that phenomenon. if you expect human attention, show that you actually gave the content human attention. if you publish slop and hand me that, you implicitly say that your time is more valueable than mine.
Ok I get it. I too have generated loads of docs with claude. It feels like flying. But flying without knowing where you'll land is an unpleasant feeling: you build up, very fast, a Comprehension Debt that'll have to be paid back some day, by someone.
I guess this is part of a natural AI-cycle I'm unvoluntary taking part of: generating loads of good looking text, then tasting it, and it tastes like shit. You can get away with pretty much form these days, it feels like. Content is plentiful, it seems.
act 1: confidence
Earlier this year, I was at a conference where Stenberg, inventor of curl, spoke. His take was that, for security related issues in curl, the AI slop is pretty much gone. Most of the reports they get now are of high quality, have proof of concept and are usually protocol implementation logic errors.
That made me think, hey, Opus could probably dream up lots of docs for me. As a programmer I've never done it properly anyways.
It looked so so nice! I got butterflies in my stomach and accepted it into the codebase. Life went on, everyone was impressed with how good the docs site looked and felt. My gut feeling was ”eh, its not perfect, but this could definitely work with some iteration. I have been the human in the loop here”.
It took maybe 3 weeks until a real person read through all the docs and concluded, it was hot garbage. Cohesiveness was off. Things were mentioned and redefined in many different places, even though I had told Opus to ”look closely for any errors”. It didn’t matter.
act 2: denial and anger
I started reading the text myself. Oh lord. It really look good on the surface. But I find my brain goes totally auto pilot. Because the slop is so, how to say, consistent in the noise to signal ratio overall. Text written by humans have some texture to it. Short sentences sometimes. Maybe some joke in it. The text from Opus reads like mixed potato soup, there is nothing to chew on. I need tree bark, something with structure and fiber. I want to feel the authors anger and feelings in my gut when I read a text.
Having been betrayed and bitten in the butt, not once, but many times, by Opus, I always keep a a skeptical eye on text that has many bullet points and content that doesn't say anything.
act 3: acceptance
On a sidenote, I have not written prose normally for almost 6 months now. I wonder why I stopped. It's so nice as a deep focus tool! I love shutting off wifi and going full offline, no distractions and just writing in a zen-like state.
I don't get to that state anywhere else really. Programming never really felt like that, because I often have to switch tabs to documentation, stackoverflow or previewing the web app I'm making.
ok so what
Eh