Clojurians Log v2
ai-skeptics 2026-07-25
This is what I call debugging: https://youtu.be/g_NmJHCLEVc?t=3652 Presses gas, hears the squeel, Presses gas more, Squeel increases, turns around, wheel is spinning, turns around again, turns around again, water pump is not spinning, stops the motor immediately. The bolts are too long, did that before. 1:00:52 - 1:01:06; 14 seconds.
This happens all the time, everywhere. When it's fast enough, you don't even register it. Hard to realize a sequence of events were actually a process of debugging when that process took a few steps that were all done by your spinal cord. :D
I've had this post title in my head for a little while and finally wrote it up last night: https://obryant.dev/p/2x-not-10x/
My experience is between 3x and 0.25x.
And I never am able to predict when it'll be one or the other, which is the thing that frustrates me.
For example, recently I was working on something and I had a very, very simple task - remove two fields. That was it - just that. The fields where in the DB, in an API, and at an extraction layer.
I was about to start my Spanish class so I asked for a LLM to do this simple task. It asked for a bunch of permissions, including making a temp Python code... and suddenly I found it asking me to access /proc
I aborted it, rolled back all of its work, found that it deleted my dependencies folder, and uninstalled my programming language (I was using Mise).
I'm sorry, but this is hilarious. :D I have also seen LLMs trying to "understand" Clojure code by writing a bunch of throwaway Python, then struggling even with that. Using all the wrong tools for the job.
woah great reason to run it in a sandbox 😬
Claude likes using Python scripts to count and balance parens in Clojure
Pretty sure I've seen it with GPT and Gemini as well.
The issue for me is how unpredictable it is. That task was remove two DB fields from a Ruby project. It struggled a lot and did the wrong thing Another task was to reimplent the VSCode API so that extensions work in Pulsar, my fork of Atom. LLM complained that it wasn't going to be able to do it, but I forced its hand and it did it anyway - and it worked quite well so... 🤷
+1, just now I had Opus 5 write code that has a bug that was a simple boolean flip. if AGI means that AI writes the same bugs as humans, I don't want it
> remove two fields To be fair, one interpretation of this instruction is "format /dev/disk0". "Life isn't just about dropping a couple of columns; its about the other files we remove along the way."
@foo Your post got featured in the TL;DR newsletter today!
hehe nice! I saw that it got to top 4 or 5 I think on HN yesterday.