Two years ago, people used “make the model tell a joke” to test general capability. Two years later, the same test shows virtually no improvement—even as models now provide reliable answers to math problems. That contrast is what we want to discuss.
What This Is
The post came from r/LocalLLaMA. The author shared the outputs of the same model—referred to in the post as Astra, with Light, Medium, and High tiers—on a joke-telling task: a classic joke using the structure of “a raven and a watchdog,” a joke about an AI working as an intern, and a story about a hourly worker and a fastidious owner.
The first two barely worked. The third was almost a primitive form of comedy of absurdity: the human-machine relationship was inverted, but the model did not realize it.
The author’s judgment was that humor depends on genuine human life experience and an intuitive calibration of what is funny—the ability to adjust timing, context, and exaggeration to precisely the right level. This is difficult to achieve simply by adding more parameters. He also acknowledged that some structurally weaker models had produced funnier material, so this is not purely a matter of parameter count.
Industry View
Notably, models have indeed advanced rapidly in math, coding, and retrieval—tasks with clear right-and-wrong standards—while making slow progress in humor, aesthetics, and social tact: tasks with no standard answer that depend on shared experience. This is consistent with a common view in the industry: capability growth is domain-specific, not universal.
There are risks and counterarguments here. We do not consider this a dead end. Treating the post’s scenario of cheap content being mass-produced and made addictive through engagement as pure science-fiction anxiety may underestimate the available engineering paths. Models can compensate for missing real-world experience through preference data, alignment training, and retrieval of external real-world corpora. More specifically, the structure of a joke can be annotated and imitated. The bottleneck may be time budgets and evaluation methods, rather than a ceiling on capability.
There is also an important caveat in the other direction: the post itself is a single-sample observation, and three examples are not enough to support any strong conclusion. Treating it as proof that “LLMs cannot be humorous,” or as proof that “LLMs can write puns,” would both be overinterpretations. The genuinely informative detail is the absence of change over a two-year period. At minimum, this suggests that humor is not a problem that is quickly converging.
Impact on Regular People
For enterprise IT: When procuring or selecting models, stop treating “general intelligence” as a single aggregate score. Verifiable tasks—content generation, customer-service scripts, internal documentation, and coding—are already improving. External communication that requires judging tone, boundaries, and context—public relations, sales, and brand content—still requires human review.
For individual careers: In work such as drafting copy, preparing proposals, and creating presentations, models can already save substantial time on first drafts. But “taste”—knowing which line to keep or which punchline feels forced—remains a human responsibility for now. This shifts the focus of the job from production to judgment.
For the consumer market: If models genuinely continue to struggle with jokes, fears of an “AI content glut” have at least not materialized in humor. However, mass-generated short-video scripts and product copy will appear first, and audiences will develop faster immunity to content that is correct but uninspired. The ultimate losers may be content producers whose advantage depends solely on output volume.