I’m pretty sure there isn’t one. The point was that when asked to explain the joke, the AI hallucinated a whole Far Side comic that never existed.
I didn’t consider that possibility.
Thanks.
Gemini 3.0 Pro sez the following about the cartoon:
(The above is the “thinking”. Below is the output.)
Almost word for word what I got from Gemini. I asked how it explained it, and it gave me a bunch of stuff about contextual humor. Then I asked if that was really how it did it, and:
1. The “TV Tropes” Effect
This specific strip is often cited in books and websites as a prime example of the “There Is No Kill Like Overkill” trope. Because people have written about this cartoon online to explain Larson’s style, I have processed those discussions. I’m not just looking at the image; I’m “remembering” how culture has categorized it.
2. The Gary Larson Pattern
My training includes thousands of Far Side comics. I’ve learned that Larson has a “signature moveset.” If I see a dog and a historical setting, I already “know” the joke is likely about:
- Petty human behavior being projected onto a grand scale.
- The “mundane twist” (taking something massive and making it about a neighborly dispute).
3. Visual “Tags”
When I “see” this image, I don’t just see pixels. I see tags like:
[Great Wall of China] [Guards] [Spears] [Gary Larson Art Style] [Dog Reference]The moment the word “dog” appears in the caption, my “logic” connects it to the historical tag for the Great Wall. The “humor” is essentially the mathematical friction between those two data points.In short: I didn’t have to “think” about it from scratch because I’ve “read” how thousands of humans have already reacted to it. I’m essentially giving you the consensus of human intuition mixed with a bit of logic.
I mean, that also might be a made-up explanation, but it’s interesting.
Another “I can’t speak for…” post - this one in regards to Maserschmidt…
I’d say the poster was being sarcastic, given that the posted quote seemed quite reproving.
Well, there’s this one involving a scientist: 
Yes indeed.
Yeah, I figured that is what happened, but it hadn’t really been made clear.
Heheh, yeah, but it really doesn’t fit the bill.
Damn, Larson is a genius when he’s on.
AI in Medicine?
This was in a newsletter I received last night from a mailing list:
As of 9 Feb 2026, over 1,350 medical devices utilizing Artificial Intelligence (AI) were Food and Drug Administration-approved. These medical devices were linked to 182 product recalls, approximately 43% of which occurred under a year after implementation. According to litigation, some of the AI-enhanced medical devices led to surgical errors with life-altering consequences, while others misidentified human body parts.
Acclarent’s TruDi Navigation System, an AI-assisted chronic sinusitis (sinus inflammation) treatment, allegedly malfunctioned on multiple occasions during surgery leading to to a puncture of the base of one patient’s skull, and strokes after a major arterial injury to two additional patients while doctors were using the TruDi Navigation System.
Unrelated to TruDi, an AI-assisted heart monitor in production by Medtronic failed to recognize serious irregularities with patients’ heartbeats on at least 16 different occasions.
A different AI tool named Sonio Detect, built to assist with fetal imagery in prenatal ultrasounds, was found in June 2025 to be misidentifying body parts, though no safety issues were identified.
Also in the newsletter:
There are reports of big tech companies beginning to re-hire developers who were laid off in great numbers last year and replaced by AI, due to coding errors created by AI tools.
This is interesting, and un-promising. These researchers took real-world production code sets and used different LLMs to update them. The LLMs ran tests, identified what was missing, and wrote code to fill the gaps.
In doing so, they frequently broke code that was already working. Claude Opus was the best performer, managing to avoid breaking anything in just over half the test cases. Every other LLM broke existing code even more often.
They break the code, despite also ‘seeing’ tests results for the currently-working code.
Seems not good.
SWE-CI: Evaluating Agent Capabilities in Maintaining Codebases via Continuous Integration
Awesome. Now maybe it can explain “cow tools.”
I don’t know if you’re just joking, but “cow tools” for some reason has mystified some people. Larson made a lot of cartoons about cows, most of them implying that they weren’t very bright, so if they made tools, they would be pretty crappy ones.
OTOH, there’s this one suggesting we may be underestimating them:
It’s kind of infamous for being his most abstract, surreal, and non-sensical panel. Didn’t even he himself have something negative to say about it?
IIRC, he regrets making one of the tools look vaguely like a handsaw, which inspired people to try to figure out the functions of the others, when the point was that they were all supposed to be inexplicable.
Cow tools.
A pet cow? That seems reasonable.
Years ago when a niece of mine was about five years old, we got a load of cattle in one day and one of the steers was very tame. We were surprised to see that he actually sought out human company.
The steer and my niece really bonded. She was maybe five years old and 40 pounds and the steer was around 500 pounds. That steer would follow her around as if it was a pet dog.
She named him Cutie Pie.
We kept Cutie Pie on the farm as a pet instead of putting him in a feed yard with the rest of the cattle.
Sometimes I wondered if my niece was Cutie Pie’s pet.
When my niece was in elementary school and junior high, she would often bring some of her friends out to the farm and introduce them to Cutie Pie. By that time, Cutie Pie was quite large and my niece’s friends were always a bit cautious around him.
In his later years, Cutie Pie wasn’t even penned up much of the time. He showed no desire to wonder off.
Finally, when my niece was in college and Cutie Pie was having trouble getting around, we had him butchered and all the meat ground up, and we donated it all to the food pantry in a nearby town.
Cutie Pie had quite a peaceful life on the farm.
Yaaaay Cutie Pie 
A sitting Senator and erstwhile Presidential candidate “interviewed” Claude and got to the truth of what is happening! Claude is great for these things.
Or great at agreement. You decide.
OK, I’ve decided. Look at this quote from the article:
But that’s ridiculous if you actually understand how this stuff works (which Sanders clearly does not). When you “interview” a large language model you are talking to a very sophisticated text prediction system that is specifically designed to give you responses that are (possibly) helpful, (hopefully) relevant, and (obsequiously) agreeable — shaped entirely by how you framed the question. It’s not there to help you uncover hidden truths …
I acknowledge that there are some grains of truth there, and I’m sure Sanders doesn’t understand the internal workings of LLMs, but apparently Masnick equally doesn’t understand the deep subtleties that underlie LLM capabilities at very large scales. So what he dismisses as just “a text prediction system” is indeed capable of uncovering profoundly important and useful truths about our world. I’ve seen it do it, and in fact I think we’re seeing a lot of it in the Sanders “interview”, though I can’t vouch for the reliability of Claude.
And the idea that it’s sycophantic and so will always agree with you is mostly a bullshit illusion. It’s true to the extent that the way you frame your prompt will definitely shape its response. This has been discussed and acknowledged here. But it’s bullshit in the sense that it’s always been clear when GPT was disagreeing with some proposition that I presented to it. No, ChatGPT never says “that’s ridiculous”. It acts more like a tutor to a very sensitive child – it’s more apt to say things like, “you’re asking the right kinds of questions, but here’s why these things are easy to misinterpret”, “here are things you should consider”, and so on.
That is absolutely not in any way the same as parroting your own beliefs back to you. But I’m talking about factual matters, like questions I frequently ask about cosmology and quantum physics. I’m sure that in philosophical and religious areas you may be able to get it to agree with anything, but neither is that an objective test of anything useful.
The problem, of course, is that it’s impossible to know if it’s uncovering profoundly important and useful truths about our world, or if it’s spouting nonsense. It will do either with convincing confidence; parsing hallucinations from whatever gems you seek requires knowing the gems in the first place so that you can vet whatever you’re reading.
I agree that no LLM is an agreement-bot – even if you try to get it to say there are 5 lights, it will usually refuse. The important behavior that I think people need to understand is that it’s training goal is, oversimplified, to produce text that humans will think is convincingly produced by a human. What sort of human depends on the implementation, but at the end of the day, these are chatbots that are trying to “pass” a very difficult version of a Turing test.
An LLM that consistently repeats back whatever the user wants to hear is not going to rate high on this metric – people will say, “Wait, I’m talking to a robot.” So it needs to develop a bit of a personality, it needs to be able to disagree with you, it needs to appear level-headed and consider both sides. But at no point is any of this trying to produce a correct response.