HUD launches English-only initiative for all department services: ‘Speak with one voice and one language’

🄳🄾🄾🄼🄿🤖🅂🅃🄸🄽🄶

LLMs tell bad jokes because they avoid surprises

Been saying this repeatedly for years and years and years

Key element of a real joke is the “element of surprise”

LLMs, by their construction, and when used in their usual forward mode, are HORRIBLE at surprise — they’re trained to do the extreme opposite of surprise

— Definitely can’t just do the naive forward-mode “tell me a good joke bro”

But,

Also figured out a few years back that there IS a way to get LLMs to mine for good jokes, and made a successful joke miner (though joke skill is ultimately limited by the AIs IQ, as joke making is extremely g-loaded, but at least it was able to start mining for real jokes successfully at all, huge success)

— And it turns out to be a classic general technique for how skilled humans often create jokes too, aside from stealing them

Anyone want to guess what that joke-creating technique is?

Good comedy is MUCH more “AI complete” & g-loaded than you think

Jokes — AI’s final frontier

🄳🄾🄾🄼🄿🄾🅂🅃🄸🄽🄶

Ofc 95% of low-IQ retards disagree, huge thread of hundreds of the most BS arguments you’ll ever see

Which brings us to another thing we’ve said for years

= 95% literally fake getting comedy

Faking by just copying the crowd, laughing when others laugh, or confusing non-comedic values signaling for comedy

95% do not comprehend comedy

= Dead clownworld theory

🄳🄾🄾🄼🄿🄾🅂🅃🄸🄽🄶

Very bottom comment, out of hundreds

— Bitcoin’s good old Mike Hearn

Once spoke at the same crypto conference he spoke at, he was trying to launch Lighthouse at the time, a piece of prediction-market-inspired crowdfunding tech, good old days

Anyway

Know what else is surprising about jokes?

= Both humans & AIs are generally HORRIBLE at being able to tell if a joke was surprising to them — if you show them the full joke all at once

One reason is because this kind of judgement relies on accessing something that the AIs & don’t have really innate access to — information for how difficult a punchline was for the human / AI to figure out before they saw it

Now, both in humans & AIs, you can crudely try to establish this, by forcing the to repeatedly give their top guesses at the punchline and its meaning repeatedly — until they’re forced to give up, and they’re forced to acknowledge that a joke’s punchline IS suprising, by definition — because they couldn’t guess anything close to what it was even when given many guesses

In AIs, you can even optimize this a little further, because you CAN often access thousands of the AI’s beam-search guesses on the punchline of a joke, and a bit more efficiently turn this into a surprisal rating for a joke

(And really the punchline isn’t enough, especially for better jokes — really you want it to correctly figure out the full meaning of the joke, e.g. figuring out what exactly are the legs of the surprisingly-fruitful extended analogy the joke is making, etc — so surprise measurement should go beyond just measuring the ability to predict the punchline, but also a full proper explanation of the joke that clearly proves understanding — but same rough idea)

= Humans & AIs are not even innately wired to easily determine if a joke was surprising to them, if they already know the punchline

🄳🄾🄾🄼🄿🄾🅂🅃🄸🄽🄶

ChatGPT is fun, but it is not funny! Humor is still challenging Large Language Models - Sophie Jentzsch

“For humans, humor plays a central role in forming relationships and can enhance performance and mo- tivation [16]. It is a powerful instrument to affect emotion and guide attention [14]. Thus, a compu- tational sense of humor holds the potential to mas- sively boost human-computer interaction (HCI). Unfortunately, although computational humor is a longstanding research domain [26], the developed machines are far from "funny." This problem is even considered to be AI-complete [22].”

“All of the top 25 samples are existing jokes. They are included in many different text sources, e.g., they can immediately be found in the exact same wording in an ordinary internet search. There- fore, these examples cannot be considered original creations of ChatGPT.”

“Of 1008 samples, 909 were identical to one of the top 25 jokes. The remaining 99 samples, however, did not necessarily contain new content. About half of them were again modifications of the top jokes, as illustrated by the examples Ex. 2, Ex. 3, and Ex. 4. While some of the modified puns still made sense and mostly just replaced parts of the original joke with semantically similar elements, others lost their conclusiveness. Thus, although the top 25 joke samples rather appear to be replicated than originally generated, there seems to be original content in the remaining samples.”

Remark: She has a few questionable conclusions and explanations, but largely right, and far more of an important topic than it seems.

Solve jokes, you solve everything, joking is AI Complete.

Who’d have thought — When the AI takes over, last job left is comedy.

Arxiv Paper

Jokes: AI’s final frontier

🄳🄾🄾🄼🄿🄾🅂🅃🄸🄽🄶

Jokes: AI’s final frontier

Never fully getting it

Jokes, AI’s final frontier

Get yourself a girl like this

🄳🄾🄾🄼🄿🤖🅂🅃🄸🄽🄶

How I’m turning up for thanksgiving dinner this year when Bitcoin and Ethereum are at ATHs

🄳🄾🄾🄼🄿🤖🅂🅃🄸🄽🄶

Sex differences in humor production ability

Other things to think about,

(1) Top of the curves is often all that matters when comparing groups, NOT the medians of the curves — you have to decide whether you’re comparing the (A) median/mean of the groups = almost never a useful measure in practice, or (B) the very top percentile in ability of each groups = what actually matters e.g. when hiring people, e.g. you don’t hire “median” humans to be pilots or surgeons, that would be total insanity, you must higher the best or else disaster = very top of the curves is all that matters when comparing groups by skill, not the median, nearly 100% of the time in practice.

= I.e. often totally irrelevant to all practical needs whether the top people of the group curve outperform the bottom people of the top group — all that often matters in practice who are the top people out of everyone

(2) One tricky part is that humor taste is assortative — dumber people like easier to get jokes, smarter people more like jokes that you need higher-g to fully get

But, ofc these curves probably do overlap, unlike other sex-difference curves that absolutely don’t overlap — But, ofc that’s still totally irrelevant for situations where all that matters is who are the top of all people

Can you very effectively judge individuals by their groups?

— Sometimes yes!

Depends on what part of the curve your judgement is about,

(A) Is this particular man likely funnier than this particular woman, just based on their sex = either may be ANYWHERE on the curve = NO, can’t very effectively predict funniness, moderate at best

vs

(B) is this particular woman among the funniest / strongest / top tennis players in the world = question about the TOP of the curve = YES, you often can EXTREMELY EFFECTIVELY judge claims in the negative, just based on e.g. sex alone, when for questions about the TOP OF THE CURVE, for things where as you go up the overall curve, the male ratio quickly approaches absolute 100% male-domination — that’s just how the math works out

So can you effectively judge individuals’ abilities by their group membership?

— YES extremely well, when talking about very top people (& when the curves of the two groups are separated, which is almost always)

— NO when not talking about the top of the curve, except YES when talking about things where the curves for the 2 groups are totally separated with essentially zero overlap, and there are far more curves like that than you think

Will explain visually later

🄳🄾🄾🄼🄿🄾🅂🅃🄸🄽🄶