LLM Humor

The featured image above is Jizo with “a yellow crocus”. Not terrible but not quite what I asked for. The chats below are from ChatGPT 5.6 set to “cynical”, which I find endlessly amusing.

The Infamous Seahorse

There still isn’t a standard Unicode seahorse emoji. Humanity has apparently decided we needed multiple kinds of train before we needed Hippocampus.

Closest improvised version is something like 🐴🌊, which is technically “horse + sea” and aesthetically a crime.

It gets the point across with the subtlety of a rebus puzzle from 1997. 🐴🌊

Probably. A tiny one could be:

   __
  /  `-.
 /      \
|  o     )
 \      /
  `-.__/
     \
      )~

Not exactly museum-grade Hippocampus anatomy, but recognizably more seahorse than “horse having an ocean-related incident.”

It absolutely does. I have apparently invented the endangered spotted party balloon.

The problem with tiny ASCII seahorses is that the head needs to read as “horse,” the body needs to curl, and ASCII gives you roughly six useful curves and a slash. The emoji crisis at least had dignity.


That Guy at Dinner

GPT 5.6 Instant. In this one, I was using GPT to learn about LLMs. I’m trying to figure out a way to incorporate a knowledge graph into an LLM to improve accuracy for our particular problem. At one point the conversation took a funny turn.

Exactly. Human cognition isn’t autoregressive in the transformer sense, but limited working context plus selective retrieval from a much larger memory store is a pretty good functional analogy.

Conversation makes it obvious. We establish a temporary shared context, reason within it, then abruptly switch from Buddhist epistemology to local LLM quantization or dinner without replaying the previous 40,000 conversational tokens. Relevant older material gets retrieved when needed rather than remaining continuously “loaded.” 

And the person who keeps dragging an exhausted topic back into working context is basically doing pathological RAG: high retrieval confidence, terrible relevance threshold.

Every social group has one.

Please preserve “pathological RAG: high retrieval confidence, terrible relevance threshold.” It may be the first computationally precise diagnosis of That Guy At Dinner.


Categories:

Tags:


Categories


Recent Posts


Tags


blogging Buddhism Cults koan meta mindfulness poem Reddit tech wordpress Zen