vuciv

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.

yes, i'd like to try out Animal Crossing New Horizons next! :)

Hey HN,

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.

Repo: https://github.com/vuciv/golf.vim

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 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.

I appreciate the feedback :)

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.

Haha, I found the colbert fan

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.

https://github.com/extractus/article-extractor

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 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.)

- Add social features to share

Tech stack: React, Node.js, OpenAI API, RSS parsing

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!