Author here. I've always been fascinated by the Sorites Paradox (at what point does a pile of sand become a heap?), so I decided to run an experiment to see how different LLMs handle vague predicates.
I didn't just want a text answer, so I measured the probability logits for "Yes/No" tokens across pile sizes ranging from 1 to 100M grains.
Key takeaways:
1. Prompting "Is this a heap?" directly is useless (the model just agrees with your framing).
2. Few-shot prompting creates a fascinating sigmoid "heapness curve" for most models (Mistral, DeepSeek).
3. Llama-3-8B was the outlier—it remained perpetually uncertain (probs ~0.35-0.55) across almost the entire range. I argue this is actually the most "philosophically honest" reflection of how humans use the word.
I have a feeling that there is an optimal prompt for this type of experiment, but struggle to find it, or even know if I have found it. The charts in the post are rendered in-browser using the data points I collected. Curious to hear your thoughts :)
I would love to work on this. I have worked on this a ton and have really strong ideas for some.
1. I think people assume you have one LLM per character, but I think if you had specialized ones for each quest, item, etc.., this would actually work quite well.
3. I actually think if you cached responses under certain conditions, costs can be saved significantly. This would require quite a robust context, though, to still feel dynamic.
I actually think now that I've gone through the process, memory scanning and writing will be enough... Except, they probably have different control codes that I'd need to reverse engineer.
You should be able to run Cheat Engine on your emul*tor of choice to tweak New Leaf "and newer" titles.
And if you're a stickler for pissing Nintendo off in very specific ways, LayeredFS + Atmosphere opens up some modding opportunities right on the console itself. Not sure how easy it would be to pull something like this off though...
I think my 'old man shakes fist at clouds' thing is this. The social media platforms that censor you do it to make your content easier to sell ads against. It's actual corporate badthink correction that is rebuilding the English language. STOP VOLUNTARILY DOING IT WHEN YOU DONT HAVE TO. You should not sacrifice your free thought on the altar of quarterly results. Say the whole fucking word.
I doubt a star will make Nintendo lawyers go "ow nose, they didn't spell out emulator in full, we can't attack them! Damn those star armors!". I don't think it changes anything technically.
The only thing this kind of censoring does is countering basic censor bots I think, and somehow making swear words publishable in the US.
Look; if you want to risk your chipper little lifestyle explaining the various ways to run New Horizons, be my guest. I've insinuated enough already, anyone who cares about the discussion on-grounds wouldn't have anything else to ask.
My point is that you don't take any more risk if you write emulator instead of emul*tor. You already took the risk explaining what you explained. One character swap doesn't change anything to this.
I'm insisting because if you care about not being sued, the stars are not an adequate defense despite what you seem to believe it is, and false sense of security is dangerous.
Not that I think that what you wrote here is remotely likely to cause you troubles, but it won't protect you the day you actually document something illegal.
To rub it in:
> I'm covering my ass
No, not at all, and it's important that you realize this.
Low-level emulators can be legally identical to a virtual machine, but often isn't. Most modern consoles can't be emulated that way, and most require you to dump a bootrom from your own console hardware, alongside game keys and other dubious digital paraphernalia.
I’m excited to share golf.vim, a Vim plugin that delivers daily editing challenges right inside your editor—like a crossover between VimGolf and Wordle. Each day (or on-demand by difficulty), you get a short puzzle where you aim to match a target text in the fewest keystrokes possible. Once you succeed, your strokes and time are submitted to a leaderboard, where you can compare your scores (Eagle, Birdie, Bogey, etc.) with others.
Key Highlights:
- Daily Challenges: Automatically fetch a fresh puzzle each day, or select challenges by difficulty, date, or tag.
- Leaderboards: After completing each puzzle, you will see the top 5 shortest keylogs from start to puzzle completion.
- Scoring: Each puzzle has a par; your final result could be an Eagle, Birdie, Par, Bogey, or worse.
- Built in Public: The community suggested extra commands like :Golf easy and more flexible scoreboard displays, and helped shape the MVP.
Check out the README for setup instructions, screenshots, and details. I’d love your feedback—especially if you have challenge ideas or feature suggestions. Enjoy, and happy golfing!
I got tired of feeling down every time I read the news, so I built a site that takes real news articles and rewrites them in the style of The Onion using GPT. It pulls from various RSS feeds (BBC, Fox News, TechCrunch, etc.) and uses AI to generate satirical headlines and content in the style of Jon Stewart.
How it works:
- Fetches news from multiple RSS feeds across different categories
- Uses GPT to generate witty headlines (it first generates 5 options, then picks the funniest one. I found this actually does well in choosing a funnier option)
- Rewrites articles in a satirical style while (TRYING) to preserve the core facts
Future plans:
- Let users add their own RSS feeds for personalized satire
- Add local news integration so you can laugh at what's happening in your city
- Implement different comedic "personalities" (think John Oliver, Stephen Colbert, different political affiliations, sassy gossip, etc.)
I built this because doomscrolling was affecting my mental health, and I figured injecting some humor into the news cycle might help others too. Would love feedback from the HN community!
Interesting idea, fun use for an LLM. I think a common theme is that when people ask LLMs to be funny or satirical, they overuse simile and analogy. e.g. "screaming and calling each other names like it was the most epic rap battle of the century" in the article about Elon's subpoena. I'm not sure if it's the quantity or the lack of connection to reality, but somehow, these jokes never land for me.
On the other hand, I actually liked "Greece Declares State of Emergency on Santorini, Residents Quickly Realize They Have Nowhere Else to Vacation" because it's funny to me that the entire country of Greece has a single place where they go for vacation. However, after I realized there were actual earthquakes, the joke felt a little insensitive, but I suppose that's to be expected if you're doing irreverent news.
"Aliens Prefer UK Over US for Sightseeing Trips, Survey Says" was also a pretty clever way to reword the original article, but it misunderstood the statistics. The article says that belief in UFOs is rising in the UK, but it is still less than in the US (20% across the pond vs 42% in the US of A), making the AI headline inaccurate.
Anyway, this is clearly meant to be a fun project, but I wanted to leave some feedback in case it helps find ways to make AI funnier. I think it's a good tool for comedy, but so far, the best AI comedy that I've seen has made use of its absurdity, like the Netflix special (https://youtu.be/nH_bEtbfB9U?si=TRF7v_Zu7sNVaeWD) or DougDoug's videos (https://youtu.be/W3id8E34cRQ?si=ucL-c4gOXRgor1LB). I think that there are probably certain prompting techniques that could make an AI more naturally funny, but I'd bet you probably have to get into CoT/reasoning-type stuff to get it to really make well-though-out jokes, because ultimately the best jokes are the ones that have layers to them.
I think you're right. I actually have seen exactly what you're talking about with similes. I am hoping to fine-tune on stuff that actually make me laugh out loud over time. I'm hoping that doesn't just make it always repeat the same joke, though. Maybe if I add a like/dislike feature, I can also see what made people laugh the most?
I'm also using 3.5-turbo right now, i wonder if the ones that people use for stuff like character ai would be better?
Also, I think the absurdity angle would be really fun. I would love to have personalities that come out with a weekly digest or something. Or maybe people can submit questions to "the love doctor" or something, lol. There's a lot to play with.
Would you mind sharing any prompting or tuning in terms of 'funniness'? Aside from commanding it to 'write like the Onion', did you do anything else to judge the relative hilarity of the headlines and articles?
Thanks for asking! The headline generation is actually a two-step process that I found works better than a single prompt. I think this is my real "secret sauce" that I found success with:
1. First, I ask GPT to generate 5 different headlines using specific Onion-style techniques I found in satire writing guidelines:
- Highlighting an unspoken truth
- Expressing raw honesty of a character
- Treating a grand event in a mundane manner (or vice versa)
- Delivering critique through absurdity
Here's the actual prompt I use:
"Craft 5 satirical, humorous headlines for the given article, employing techniques such as highlighting an unspoken truth, expressing raw honesty of a character, treating a grand event in a mundane manner (or vice versa), or delivering a critique, inspired by The Onion's distinctive style. If the article allows for it, maintain a positive and light-hearted tone."
2. Then I have a second prompt that acts as a "comedy judge" to pick the best one:
"Pretend that you are a world class comedian. Choose the funniest satirical article headline out of these headlines."
It's hit or miss - some headlines are pretty bland, but occasionally it produces gems that make me laugh out loud. I'm collecting the ones that work really well with the plan to fine-tune a model specifically for my type of humor.
For the article content itself, I use Jon Stewart as the style guide because I found his approach of "truth through comedy" tends to work well with keeping the actual news facts intact while adding humor.
I'm really interested in exploring different comedic styles and personalities. I think having different "voices" (like a John Oliver version vs a Stephen Colbert version) could be cool. Would love to hear if you have any other ideas for improving the humor generation!
I appreciate the detailed response. TBH I've really only been using LLMs for mundane corporate tasks so I don't have a ton of experience on the "less-definitive" creative side of genAI.
Using the writing guidelines seems like a good way to get into the 'comedic' prediction space of LLMs that more basic prompts like 'tell me a joke' might not open up. Neat.
Mm, I actually mostly use it for my local news. Sadly, the onion doesn't write specific to Austin, Texas. Also, on my local setup, I have it for gaming/films/other niche interests that are just for me. Also, I like that it's realtime, and that I can click on it to get the real headline/info.
Ah that makes sense. Bleak comment on what's happened to local journalism though, time was you could rely on finding a good writer or broadcaster to deliver that.
Actually, I kept it simple - I use the original images from the news articles! When I fetch an article through RSS and extract its content using the @extractus/article-extractor library, it pulls the main image along with the content.
I didn't just want a text answer, so I measured the probability logits for "Yes/No" tokens across pile sizes ranging from 1 to 100M grains.
Key takeaways: 1. Prompting "Is this a heap?" directly is useless (the model just agrees with your framing). 2. Few-shot prompting creates a fascinating sigmoid "heapness curve" for most models (Mistral, DeepSeek). 3. Llama-3-8B was the outlier—it remained perpetually uncertain (probs ~0.35-0.55) across almost the entire range. I argue this is actually the most "philosophically honest" reflection of how humans use the word.
I have a feeling that there is an optimal prompt for this type of experiment, but struggle to find it, or even know if I have found it. The charts in the post are rendered in-browser using the data points I collected. Curious to hear your thoughts :)