How funny are frontier AI models? Full methodology and the winning joke
paraschopra · x · 2026-10-06
Paras Chopra published the full methodology and results of his humor experiment testing whether frontier models are improving at jokes.
- Motivation: GPT-3-era models mostly regurgitated old internet jokes ("Why can't you trust atoms?"), so he required original jokes
- 62 respondents blind-rated 48 jokes from six models; Astra won with 50%+ chuckle/laughter rate
- The crowd's funniest joke: "The probability of seeing a rainbow increases after rain. Unfortunately, so does the probability that I left my laundry outside."
- Core question: do gains in verifiable domains like math and coding transfer to taste-dependent, hard-to-verify skills
Same experiment as the earlier tweet; this post adds methodology and the winning joke.
Related event: Blind Test of 48 LLM Jokes Shows AI Humor Is Improving(2 posts)→
More from Fun
- Joke goes viral: humanity is just a bootloader for the next big model — ryunuck · 2026-10-06
- AI slop movies are now showing up on in-flight streaming as 'genuine' franchises — moyix · 2026-10-06
- Claude Opus 5.5 worked autonomously for 1h37m to flex its own capabilities — mcraddock · 2026-10-06
- Doomers are pouring big money into a sudden wave of high-quality doomer YouTube content — basedjensen · 2026-10-06
- A game where you beat up your problems, built in 30 minutes with prompts — Tegadesigns · 2026-10-06
- Claude Opus 5.5 lists the 20 jobs it will replace — according to itself — minchoi · 2026-10-06