thinking about why llms aren't considered funny.

if we look at how kantian humor is driven by intentional incongruity/laughter arising from recognizing a mismatch between expectation and reality

+ how LLMs work

it's a little clear why llms struggle to be universally funny👇

everything that can work, will work | unpacking papers, musings on mathematics, and meditations on craft from an anon founder (@yhprums_law on X)

this intentionality means humor depends on context (𝑐), expectation (𝐸), reality (𝑅), and shared knowledge (𝑆). formally, humor 𝐻 ≈ 𝑓(𝑐,𝐸,𝑅,𝑆).

humor is subjective, mapping differently onto each observer’s internal mental model 𝑀. each model varies in prior experiences (𝑃), culture (𝐶), context (𝑐), so 𝑀(𝑖) = 𝑔(𝑃,𝐶,𝑐).

large language models (llms) optimize for likelihood: maximizing probability of next token given prior context. formally, maximize Σ log 𝑝(tokenₙ|contextₙ₋₁,...,context₀).

llms thus seek average predictability (high probability next tokens), conflicting fundamentally with humor’s goal: intentional unpredictability (low probability next tokens).

kantian humor explicitly subverts expected outcomes (low-probability event). llms inherently struggle: their optimization biases output toward predictable rather than surprising tokens.

further, intention is absent in llms: no true mental model 𝑀(𝑖), merely statistical patterns learned during training. hence, they simulate but never intentionally craft incongruity.

in short, kantian humor requires intention (conscious incongruity), context sensitivity, and tailored mental models. llms lack all three… limiting their comedic potential significantly.

some of the ideas to fix this also are hard, for example explicitly model surprise: define a reward function 𝑅 inversely proportional to token predictability, i.e., reward ∝ 1/𝑝(token|context).

this fix fails practically because inverse predictability alone (1/𝑝) encourages randomness—not coherent incongruity. humor requires carefully structured surprise, not noise.

maybe introduce an intention mechanism: condition model outputs explicitly on humor-intent prompts, shifting from purely predictive to generative incongruity.

intent mechanisms remain simulated, not genuine: llms mimic intention without true semantic grounding, limiting their ability to intentionally craft nuanced incongruity.

train specialized modules to learn individualized humor preferences (𝑀ₖ for each user 𝑘), ensuring personalization: humor optimized for 𝑀ₖ rather than global optimum.

maybe mathematically, humor becomes optimization of ∑ humor(𝑀ₖ,joke), ∀𝑘 users, instead of a non-existent global frontier joke set.

but like... individualized models (𝑀ₖ) suffer sparsity: insufficient humor-specific training data per user 𝑘. fine-tuning each 𝑀ₖ accurately at scale is unrealistic computationally.

and you can keep going: rlhf for humor is bottlenecked by subjective human variability: inconsistent feedback results in noisy reward signals (𝑅), causing suboptimal or bland humor outputs.

statistical optimization of token likelihood (max log𝑝) inherently conflicts with intentional low-probability joke construction... it's a core tension unresolved.

context-aware embeddings (𝑐ₜ) help coherence but struggle capturing subtle context shifts like irony, sarcasm, or social cues without explicit symbolic understanding.

you could say, well maybe they're not smart (i think it means we haven't figured out this type of intelligence), or perhaps llms represent a distinct intelligence altogether: capable, yet alien, highlighted vividly by their struggle with humor, our most human cognitive frontier.

a critical challenge: no universal set 𝐻* of jokes exists that maximizes humor for all observers. mathematically, ∄𝐻* s.t. ∀ individuals 𝑖, humor(𝐻*,𝑖) ≥ humor(𝐻,𝑖), ∀𝐻.