|
|
|
avoiding LLM doom loops
Friday, July 18 2025
location: 940 feet west of Woodworth Lake, Fulton County, NY
Fairly early this morning, Gretchen wanted to take action against all the weeds and other plant life growing so lushly near our house. In particular, she wanted to clear the pathways of all the tall sweet clover and queen anne's lace. So I fetched the weed eater, and she tried that for a while. But she kept losing the nylon string when attempting to attack plants that were just too thick. So I suggested she use a pair of shears instead. Gretchen wanted to know what happened to the nylon string when it disappeared, apparently thinking it somehow gets teleported into another dimension. "It goes into the environment and becomes microplastics," I said. She's been worked up about microplastics since learning that the typical American has a plastic fork worth of microplastics in his or her brain. But if one wants to eat weeds using a weedeater, one must generate microplastics; whipping around a metal wire instead would probably be too dangerous.
Eventually our walkways were tidy enough, so Gretchen headed off to the lake, leaving me to start my workday.
Today would be the first time I would be attempting to work remotely for my new employer from the cabin. After managing to get access to the VPN, I attempted a rather tricky set of procedures that would really test the quality of my internet access (which, from the cabin, has been a little less reliable of late). I needed to use Remote Desktop to connect to an Oracle database server and then jump through the hoops Oracle places in front of someone to retrieve a .dmp file, copy it (via my terrible internet connection) to my local laptop, and then struggle more with Oracle to turn that .dmp file into a live database. As always, I had ChatGPT to help me at every stage, though I did my best to avoid getting stuck in LLM doom loops (now that I am more fully aware of them), short-circuiting suggestions that I already knew led to such loops. I worked much much more effectively this way and realized that it was an important strategy when using LLMs to explore unknown (but testable) fields of interest. (It's important to note that it's LLM responses to failed tests that leads to the loops, since, without testability, it's easy to take ChatGPT at its word, failing to see its frequent propensity to provide well-informed bullshit).
After multiple attempts to download a nearly on gigabyte .zip file containing a database dump, I asked ChatGPT how to segment .zip files. It helpfully suggested 7Zip, which worked like a charm.
In amongst all this, I went down to the lake on my lunch break to bring down a large ceramic water dispenser and check in on how Gretchen was doing. The day was rather cool, with temperatures in the 60s, but the dock was sunny, so it was a nice place to be. Also, the lake was fairly warm, so Gretchen had gone swimming multiple times already.
When my day ended at 3:30pm, I grabbed a beer and headed for the lake. [REDACTED] I eventually set out in a kayak, stalking three or four belted kingfishers (A family? A thruple with a kid?) in hopes of getting a photo, which I'd never successfully gotten before. I managed to get photo, though it wasn't great.
I ended up having a rather low-energy evening, though I did gather a couple buckets of topsoil from a good source of such material to further improve the soil along the west end of the south foundation wall.
The not-great photo I got of the belted kingfisher today. My photos of Sri Lankan kingfishers were much better! Click to enlarge.
For linking purposes this article's URL is:
http://asecular.com/blog.php?250718 feedback
previous | next |
