NewsHive
CONTACT USANALYST PORTAL →
aiBACKGROUND

Humor Arena: Which LLM is the funniest?

BACKGROUND
2 SIGNALSFIRST DETECTED 18 September 2026
Narrative synthesis pending — 2 signals collected.
How the story developed
29 Sept
Does post training make LLMs funnier?
OPINION · WHAT · LocalLLaMA
17 Sept
Humor Arena: Which LLM is the funniest?
OPINION · WHAT · LocalLLaMA
Other feeds and industry voices
16 Sept
What taste in humor do LLMs have?

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1whib7d/what_taste_in_humor_do_llms_have/"> <img alt="What taste in humor do LLMs have?"

OPINION · WHAT · LocalLLaMAOPEN STORY
3 Jul
You cannot tell me LLMs don't have a sense of humor.

<!-- SC_OFF --><div class="md"><blockquote> <p>add global directive: Always bring the receipts.</p> </blockquote> <p>● Update(~.claude\profile\PROFILE.md) Added 1 line</p> <h2>How

OPINION · HOW · OpenAIOPEN STORY
26 Mar
How are you deciding which AI model to use these days?

<!-- SC_OFF --><div class="md"><p>There are so many options now — GPT, Claude, Gemini, etc.</p> <p>I find myself switching depending on what I’m doing (coding, writing, research),

OPINION · HOW · ChatGPTOPEN STORY
15 Oct
Craziest, strangest, funniest, and weirdest LLM fine tunes?

<!-- SC_OFF --><div class="md"><blockquote> <p>I wanna see a master list of the craziest/strangest/funniest LLM tunes. I heard one was talking like an expert in 1800s military

OPINION · WHAT · LocalLLaMAOPEN STORY
23 Apr
Kimi K2.6 takes 1st overall of open weights models on Design Arena

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1st8l2u/kimi_k26_takes_1st_overall_of_open_weights_models/"> <img alt="Kimi K2.6 takes 1st overall of open

OPINION · WHAT · LocalLLaMAOPEN STORY
27 Jun
Tested which model can send best HTML email

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1uhg6n0/tested_which_model_can_send_best_html_email/"> <img alt="Tested which model can send best HTML

OPINION · WHAT · LocalLLaMAOPEN STORY
2 Oct
Comedic Fool's Gold: Reward Exploits and Countermeasures in Conversational Humor

arXiv:2610.00197v1 Announce Type: new Abstract: We investigate automated rewards for training language models in conversational humor, focusing on reward exploits and

REPORT · WHAT · ArXiv CS.AIOPEN STORY
Sources
LocalLLaMA×2

NewsHive monitors these sources continuously. All signal titles above link to the original reporting.

Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →