|
|
|
ChatGPT can't bullshit me
Monday, October 27 2025
setting: rural Hurley Township, Ulster County, New York
Charlotte was extremely excited when I got home from work today, doing that greeting howl she does and then punctuating it with some barks for good measure as she leapt over the back of the couch at me. I saw she was wearing the new Tractive tracker I'd recently bought for her to replace the dinky cat tracker whose batteries typically lasted only a couple days. This meant Gretchen had managed to attach it to her collar. It's bigger and heavier, though supposedly it can run for weeks.
When I took the dogs for a walk, both came, and we headed some distance down the Gullies Trail before I crossed over to the Stick Trail to return home. I don't have a lot of time for dog walking on evenings when I am expected to cook dinner, so the best possible thing that can happen is that Charlotte and Neville go have their own adventure while I am returning home. Some version of that happened today, though it was not a big adventure.
Back at the house, I proceeded to make my low-effort chili (which is only enough to fill a large frying pan), which is tastier (I think) than my usual big pot of chili. I think this is because I use a lot more tomato.
After dinner, I took a nice solar-heated bath. One of the shower heads still drips more than I would prefer, though cold drops of water landed on me only three or four times.
Recently someone in San Francisco got in contact with me about my Arduino sketch that can read SolArk inverter data. Like me, he has an off-grid cabin somewhere and wants to gather data for internal use. He offered to pay me for consulting with him and asked if I could flash him a NodeMCU with my SolArk-reading sketch. I said sure, and was even thinking I would do it for free just to have someone else to bounce off-grid remote access ideas off of. In the end I said I'd do it for $20, and he Venmo'd me $25. Only later did I realize that I need a fair amount of information from him in order to flash him a custom NodeMCU, including the access credentials for his WiFi hotspot.
Today when using ChatGPT, I ran into a "cognitive issue" (for lack of a better term). I asked it to do something I have found useful to calculate on occasion: the beginning and ending of the warmest half of the year for a given location, in this case Kingston, NY. Initially it came back with the beginning of the warm half of the year as being May 28 and the end as being November 30. But then I pointed out that the average temperature on the beginning and ending days of the warm half should be about the same. It tried to explain why it didn't calculate it this way by saying something that was clearly bullshit:
That asymmetry comes from seasonal lag, which means the air and ground continue warming or cooling for weeks after the solstices or equinoxes because the Earth's surface stores heat. The effect is especially pronounced in temperate climates like Kingston, NY.
I then asked it to actually align the beginning and end of the warm half with days that are of roughly equal average temperatures, and it came back with what is close to the correct answer (that the warm half begins on April 20 and ends on October 22). I then asked it to give me the difference in length between the warm "half" (since it isn't truely a half) and the cold "half," and it correctly said that the warm half is 7 days longer. I then asked it to explain why this might be, and again it retreated to bullshit explanations, one of which it had already used explain why the beginning and ending days of the warm half had different average temperatures:
Why the imbalance exists
Seasonal lag — The Earth's surface warms more slowly than it cools, especially in summer when the sun's energy continues to build up even after the solstice.
Continental vs oceanic influences — Kingston is inland but still within reach of Atlantic air masses, which delays autumn cooling more than spring warming.
Temperature curve asymmetry — The rise from cold to warm in spring is steeper, but the decline in autumn is more gradual, so the area above the mean (warm half) covers more time.
I then asked:
don't you think the best explanation for this discrepancy is the fact that the sun is further away from the earth during the warm half and thus our planet is moving more slowly?
It then did the math and "realized" that the position of Earth with respect to the Sun in summer versus winter completely explained the difference in length between the two "halves."
This is all to say that one must be wary of ChatGPT bullshitting explanations for phenomena when getting to the correct answer requires a sound model of how the Universe works. As good as ChatGPT is, it lacks that model, though perhaps some day one can be integrated into it somehow.
The Chamomile Wall as it looked today from the south with Neville walking through its one aperature: for the Stick Trail. Click to enlarge.
Looking east from a third of a mile from home on the Gullies Trail. This is in the area where a tornado went through, so many trees that would've blocked the view were knocked down. The ridge to the east is across Dug Hill Road a quarter mile away and is lit up by the low late October sun. Click to enlarge.
A wider view. Click to enlarge.
Looking northward on the Gullies Trail about a third of a mile from home. Click to enlarge.
For linking purposes this article's URL is:
http://asecular.com/blog.php?251027 feedback
previous | next |
