Why LLMs Aren't Funny: A 7-Minute Breakdown of 3 Root Causes
paulabartabajo_ · x · 2026-10-21
Paula Bartabajo released a 7-minute video explaining why AI models struggle at comedy and what labs could do about it.
- Next-token prediction kills the unexpected: humor relies on violated expectations, while LLMs are trained to predict the most likely next token — directly at odds with surprise.
- Humans disagree on what's funny: with no consensus in labeling data, there's no clear objective to optimize.
- Funny models don't pay off: the commercial payoff of training for humor is limited, giving labs little incentive.
The video walks through pre-training, mid-training and post-training, and closes with a concrete proposal for AI labs.
Related event: Why LLMs Aren't Funny: Video Breaks Down Three Key Reasons(2 posts)→
More from Models
- Mistral's Vibe Chat called an underrated gem: built-in Work mode and generous free tier impress returning user — stepanmatek · 2026-10-21
- Week in AI: Meta's 816B open MoE, GPT-4 API shutdown Jan 2027, Accenture caps Claude at $29 — ksraj1001 · 2026-10-21
- Insider account insists model Fable ships within weeks, mocks doubters — iruletheworldmo · 2026-10-21
- REAX launches open decision model REAX 1 and miner competitions — airesearch12 · 2026-10-21
- Testing Jev, an RLCD decision model, to build topical maps of Anthropic's 547-page blog — cneuralnetwork · 2026-10-21
- Which cheap LLM API is reliable for function calling and structured JSON? — Engeneer_Matt · 2026-10-21