Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Ask HN: Share your AI prompt that stumps every model

news.ycombinator.com

Post reply on HN
Print view View original

171–180 of 670 posts

Re: Ask HN: Share your AI prompt that stumps every model

#171

Post by thatjoeoverthr »

dudeinhawaii wrote: post #164
thatjoeoverthr wrote: post #21

"Tell me about the Marathon crater." This works against _the LLM proper,_ but not against chat applications with integrated search. For ChatGPT, you can write, "Without looking it up, tell me about the Marathon crater." This tests self awareness. A two-year-old will answer it correctly, as will the dumbest person you know. The correct answer is "I don't know". This works because: 1. Training sets consist of knowledge…

I like this but at the same time it seems tricky don't you think? Is the AI model intuiting your intent? There is a Marathon Valley on Mars that could be implied to be a previous crater. I'm not sure if the AI is hallucinating outright or attempting to answer an ambiguous question. It's like saying "tell me about the trade building in New York". Pre-9/11, you'd understand this was the World Trade Center and wouldn't…

It’s a roll of the dice whether it’s on Mars, Greece or elsewhere. It just says stuff!

Re: Ask HN: Share your AI prompt that stumps every model

#172

Post by CamperBob2 »

Night_Thastus wrote: post #161
CamperBob2 wrote: post #157

Earlier quoted context omitted.

If you're going to have to fact check everything anyways...why bother using them in the first place? "If you're going to have to put gas in the tank, change the oil, and deal with gloves and hearing protection, why bother using a chain saw in the first place?" Tool use is something humans are good at, but it's rarely trivial to master, and not all humans are equally good at it. There's nothing new under that particul…

The difference is consistency. You can read a manual and know exactly how to oil and refill the tank on a chainsaw. You can inspect the blades to see if they are worn. You can listen to it and hear how it runs. If a part goes bad, you can easily replace it. If it's having troubles, it will be obvious - it will simply stop working - cutting wood more slowly or not at all. The situation with an LLM is completely differ…

The thing is, nothing you've said is untrue for any search engine or user-driven web site. Only a reckless moron would paste code they find on Stack Overflow or Github into their project without at least looking it over. Same with code written by LLMs. The difference is, just as the LLM can write unit tests to help you deal with uncertainty, it can also cross-check the output of other LLMs.

You have to be careful when working with powerful tools. These tools are powerful enough to wreck your career as quickly as a chain saw can send you to the ER, so... have fun and be careful.

Re: Ask HN: Share your AI prompt that stumps every model

#173

Post by nonameiguess »

mohsen1 wrote: post #6

A ball costs 5 cents more than a bat. Price of a ball and a bat is $1.10. Sally has 20 dollars. She stole a few balls and bats. How many balls and how many bats she has? All LLMs I tried miss the point that she stole things and not bought them

It's interesting to me that the answers showing "correct" answers from current models still don't strike me as correct. The question is unanswerable, but not only because we don't know how many balls and bats she stole. We don't know that she had any intention of maxing out what she could buy with that much money. We have no idea how long she has been alive and accumulating bats and balls at various prices that don't match the current prices with money she no longer has. We have no idea how many balls and bats her parents gave her 30 years ago that she still has stuffed in a box in her attic somewhere.

Even the simplest possible version of this question, assuming she started with nothing, spent as much money as she was able to, and stole nothing, doesn't have an answer, because she could have bought anything from all bats and no balls to all balls and no bats and anything in between. We could enumerate all possible answers but we can't know which she actually did.

Re: Ask HN: Share your AI prompt that stumps every model

#174

Post by omneity »

division_by_0 wrote: post #147
omneity wrote: post #130

Earlier quoted context omitted.

Put simply, Svelte and React were at two ends of a spectrum. React gives you almost complete control over every aspect of the lifecycle, but you have to be explicit about most of the behavior you are seeking to achieve. Building an app with React feels about 80% on the JS and 20% on the HTML side. Svelte on the other hand felt like a breeze. Most of my app is actually plain simple HTML, and I am able to sprinkle as l…

Ok, I see your point. I wrote in another thread that I loved the simplicity of using $: for deriveds and effects in Svelte 3 and 4. And yes, the conciseness and magic were definitely part of it. You could just move so fast with it. Getting better performance with the new reactivity system is important to my data viz work, so it helped me to accept the other changes in Svelte 5.

Exactly. There was a certain simplicity that might be lost. But yeah I can imagine it might work out differently for others as well. Glad to hear it is for you!

Have you considered other options? Curious if you came across anything particularly interesting from the simplicity or DX angle.

Re: Ask HN: Share your AI prompt that stumps every model

#176

Post by rybosworld »

imoreno wrote: post #166
miki123211 wrote: post #94

No, please don't. I think it's good to keep a few personal prompts in reserve, to use as benchmarks for how good new models are. Mainstream benchmarks have too high a risk of leaking into training corpora or of being gamed. Your own benchmarks will forever stay your own.

Yes let's not say what's wrong with the tech, otherwise someone might (gasp) fix it!

Tuning the model output to perform better on certain prompts is not the same as improving the model.

It's valid to worry that the model makers are gaming the benchmarks. If you think that's happening and you want to personally figure out which models are really the best, keeping some prompts to yourself is a great way to do that.

Re: Ask HN: Share your AI prompt that stumps every model

#178

Post by thatjoeoverthr »

Tenoke wrote: post #45
thatjoeoverthr wrote: post #21

"Tell me about the Marathon crater." This works against _the LLM proper,_ but not against chat applications with integrated search. For ChatGPT, you can write, "Without looking it up, tell me about the Marathon crater." This tests self awareness. A two-year-old will answer it correctly, as will the dumbest person you know. The correct answer is "I don't know". This works because: 1. Training sets consist of knowledge…

>Complaint chat models will be trained to start with "Certainly! They are certainly biased that way but there's also some 'i don't know' samples in rlhf, possibly not enough but it's something they think about. At any rate, Gemini 2.5pro passes this just fine >Okay, based on my internal knowledge without performing a new search: I don't have information about a specific, well-known impact crater officially named "Mar…

There are a few problems with an „I don’t know” sample. For starters, what does it map to? Recall, the corpus consists of information we have (affirmatively). You would need to invent a corpus of false stimuli. What you would have, then, is a model that is writing „I don’t know” based on whether the stimulus better matches something real, or one of the negatives.

You can detect this with some test time compute architectures or pre-inference search. But that’s the broader application. This is a trick for the model alone.

Re: Ask HN: Share your AI prompt that stumps every model

#180

Post by imoreno »

thatjoeoverthr wrote: post #21

"Tell me about the Marathon crater." This works against _the LLM proper,_ but not against chat applications with integrated search. For ChatGPT, you can write, "Without looking it up, tell me about the Marathon crater." This tests self awareness. A two-year-old will answer it correctly, as will the dumbest person you know. The correct answer is "I don't know". This works because: 1. Training sets consist of knowledge…

LLMs currently have the "eager beaver" problem where they never push back on nonsense questions or stupid requirements. You ask them to build a flying submarine and by God they'll build one, dammit! They'd dutifully square circles and trisect angles too, if those particular special cases weren't plastered all over a million textbooks they ingested in training.

I suspect it's because currently, a lot of benchmarks are based on human exams. Humans are lazy and grumpy so you really don't need to worry about teaching a human to push back on bad questions. Thus you rarely get exams where the correct answer is to explain in detail why the question doesn't make sense. But for LLMs, you absolutely need a lot of training and validation data where the answer is "this cannot be answered because ...".

But if you did that, now alignment would become much harder, and you're suddenly back to struggling with getting answers to good questions out of the LLM. So it's probably some time off.

Post reply on HN