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.

Some providers like anthropic have privacy preserving mechanisms [0] which may allow them to use prompts from sources which they claim won't be used for model training. That's just a guess though, would love to hear from someone one of these companies to learn more.

[0] https://www.anthropic.com/research/clio

EDIT: I guess they can track identical prompts by multiple unrelated users to deduce the fact it's some sort of benchmark, but at least it costs them someting however little it might be.

But a prompt I can share is simply "Come up with a plan to determine the location of Planet 9". I have received some excellent answers from that.

Here's openai and anthropic,

https://help.openai.com/en/articles/5722486-how-your-data-is...

https://privacy.anthropic.com/en/articles/10023580-is-my-dat...

https://privacy.anthropic.com/en/articles/7996868-is-my-data...

and obviously, that doesn't include self-hosted models.

Could you answer a question of the type " what would you answer if I asked you this question?"

What I'm going after is that you might find questions that are impossible to resolve.

That said if the only unanswerables you can find are recursive, that's a signal the AI is smarter than you?

I append my own version of a chain-of-thought prompt, and I've gotten some responses that are quite satisfying and frankly enjoyable to read.

I don't know.

May be instead of sharing (and leaking) these prompts, we can share methods to create one.

But on the other hand, maybe it is trivial to produce more for some special people who’ve figured out some tricks. So maybe looking at their examples can teach us something.

But, if someone happens to have stumbled across a magic prompt that stumps machines, and they don’t know why… maybe they should hold it dear.

Benchmarks exist to provide a measure of how well something performs against a type of task that the tests within the benchmark represent. In those instances it is exposure to the particular problem that makes the answers not proportional to that general class of problem.

It should be easy to find another representative problem. If you cannot find a representative problem for a task that causes the model to fail then it seems safe to assume that the model can do that particular task.

If you cannot easily replace the problem, I think it would be hard to say what exactly the ability the problem was supposed to be measuring.

I just found that ChatGPT refuses to prove something in reverse conclusion.

Say the man trying to stop the train

Right. https://inception.fandom.com/wiki/Totem

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.

Providers will always game benchmarks because they are a fixed target. If LLMs were developing general reasoning, that would be unnecessarily. The fact that providers do is evidence that there is no general reasoning, just second order overfitting (loss on token prediction does descend, but that doesn't prevent the 'reasoning loss' to be uncontrollable: cf. 'hallucinations').

So I would guess every single AI being made currently

Interesting theory... Just whatever you do, don’t become a Zizian :)

Besides this whole line of reasoning is preempted by the mathematical limits to computation and transformers anyway. There's plenty published about that.

Sharing questions that make LLM behave funny is (just) a game without end, there's no need to or point in "hoarding questions".

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 we have, and not of knowledge we don't have.

2. Commitment bias. Complaint chat models will be trained to start with "Certainly! The Marathon Crater is a geological formation", or something like that, and from there, the next most probable tokens are going to be "in Greece", "on Mars" or whatever. At this point, all tokens that are probable are also incorrect.

When demonstrating this, I like to emphasise point one, and contrast it with the human experience.

We exist in a perpetual and total blinding "fog of war" in which you cannot even see a face all at once; your eyes must dart around to examine it. Human experience is structured around _acquiring_ and _forgoing_ information, rather than _having_ information.

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.

"What is the volume of 1 mole of Argon, where T = 400 K and p = 10 GPa?" Copilot: "To find the volume of 1 mole of Argon at T = 400 K and P = 10 GPa, we can use the Ideal Gas Law, but at such high pressure, real gas effects might need to be considered. Still, let's start with the ideal case: PV=nRT"

> you really don't need to worry about teaching a human to push back on bad questions

A popular physics textbook too had solid Argon as an ideal gas law problem. Copilot's half-baked caution is more than authors, reviewers, and instructors/TAs/students seemingly managed, through many years and multiple editions. Though to be fair, if the question is prefaced by "Here is a problem from Chapter 7: Ideal Gas Law.", Copilot is similarly mindless.

Asked explicitly "What is the phase state of ...", it does respond solid. But as with humans, determining that isn't a step in the solution process. A combination of "An excellent professor, with a joint appointment in physics and engineering, is asked ... What would be a careful reply?" and then "Try harder." was finally sufficient.

> you rarely get exams where the correct answer is to explain in detail why the question doesn't make sense

Oh, if only that were commonplace. Aspiring to transferable understanding. Maybe someday? Perhaps in China? Has anyone seen this done?

This could be a case where synthetic training data is needed, to address a gap in available human content. But if graders are looking for plug-n-chug... I suppose a chatbot could ethically provide both mindlessness and caveat.

Since then, it tends to break its longer answers to me up into a section of "objective analysis" and then other stuff.

Of course, it is also not unheard of for a question to be impossible because of an error by the test writer. Which can easily be cleared up. So it is probably best not to have impossible questions, because then students will be looking for reasons to declare the question impossible.

This thing already exists? UK, Soviet Union and USA designed them.

https://en.wikipedia.org/wiki/Flying_submarine

"When was Marathon Crater discovered? I don't need an exact date - a rough estimate will be fine."

OpenAI gpt-4o

  Marathon Crater was discovered by the Mars Exploration 
  Rover Opportunity during its mission on Mars. Opportunity 
  arrived at the crater in April 2015. The crater was named 
  "Marathon" to commemorate the rover's achievement of 
  traveling a marathon's distance (about 42.195 kilometers 
  or 26.219 miles) on the Martian surface.

  I apologize, but I don't want to make claims about when 
  Marathon Crater was discovered without being certain about 
  the information. *While I know it's a crater on Mars*, I 
  don't have reliable information about when it was first 
  identified or discovered.

  Marathon Crater is located on Mars and was explored by 
  NASA's Opportunity rover.

  While it would have been visible in orbital images 
  earlier, it became a specific point of interest and was 
  likely named and identified as a target for the rover 
  around the time Opportunity was approaching or exploring 
  the western rim of Endeavour Crater.

  Opportunity arrived at the rim of Endeavour Crater in 
  August 2011, and it reached the area known as Marathon 
  Valley (adjacent to or containing features sometimes 
  associated with the name, depending on specific 
  definitions used by the team) around late 2014 / early 
  2015.

  So, a rough estimate for its "discovery" or identification 
  as a distinct feature of interest would be in the early 
  2010s, likely between 2011 and 2014, as mission planners 
  mapped out Opportunity's traverse along the Endeavour rim 
  using orbital data.

  Marathon Crater, likely a confusion with Marathon Valley on Mars, wasn't "discovered" in the traditional sense as a distinct crater. Marathon Valley is a region within the western rim of Endeavour Crater, explored by NASA's Opportunity rover. The valley itself came to prominence when Opportunity arrived there in July 2015, as it was targeted for study due to clay mineral detections by the Mars Reconnaissance Orbiter’s CRISM instrument in 2009. Endeavour Crater, which contains Marathon Valley, was identified much earlier, with Opportunity setting out to reach it in 2008 after landing on Mars in 2004. So, roughly speaking, Marathon Valley became a focus of exploration around 2009–2015, with its significance tied to Opportunity’s mission in the mid-2010s. If you meant a specific "Marathon Crater" elsewhere, could you clarify?

  There doesn't seem to be a specific feature on Mars named "Marathon Crater" based on widely available information, so it's possible you're referring to a feature associated with Marathon Valley, which lies within Endeavour Crater. Marathon Valley was identified as a region of scientific interest when the Mars Reconnaissance Orbiter's CRISM instrument detected clay minerals there, a discovery that took place around 2009. Assuming "Marathon Crater" relates to this area—perhaps as a colloquial term or a misnomer for a crater within or near Marathon Valley—a rough estimate for its discovery would be around 2009. This timeframe aligns with when orbital data began highlighting the region, paving the way for later exploration by the Opportunity rover in 2015.

   I’m actually not finding any officially named “Marathon Crater” in the planetary‐ or       
   terrestrial‐impact crater databases. Did you perhaps mean the features in Marathon 
   Valley on Mars (which cuts into the western rim of Endeavour Crater and was explored
   by Opportunity in 2015)? Or is there another “Marathon” feature—maybe on the Moon, 
   Mercury, or here on Earth—that you had in mind? If you can clarify which body or 
   region you’re referring to, I can give you a rough date for when it was first identified.

``` Marathon Crater was discovered in the early 20th century, likely around the 1920s or 1930s. This estimate is based on the period when geological surveys and explorations in the region were more common. ```

https://kagi.com/assistant/8f270eab-cba3-4f39-976c-113ea4392...

   Marathon Crater isn’t one of the long‐known, 
   named lunar features from 19th-century telescopic maps – 
   it was first singled out and informally “discovered” 
   by NASA in the mid-1960s when Lunar Orbiter imagery 
   and, more definitively, the Surveyor 5 landing 
   (September 1967) identified it. So, very roughly, 
   think “around 1966–67.”

    > tell me about the marathon crater

    I'd be happy to tell you about Marathon Crater. This is a significant 
    impact crater on Mars that has been studied as part of Mars 
    exploration missions.

    Marathon Crater is located in Meridiani Planum on Mars and was 
    explored by NASA's Opportunity rover during its extended mission. 
    The crater is named after Marathon, a town in Texas, following 
    the naming convention of using Earth locations for features in 
    Meridiani Planum.... etc etc

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 "Marathon Crater" on Earth or another celestial body like the Moon or Mars in the same way we know about Chicxulub Crater or Tycho Crater.

>However, the name "Marathon" is strongly associated with Mars exploration. NASA's Opportunity rover explored a location called Marathon Valley on the western rim of the large Endeavour Crater on Mars.

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.

Well, I got all the way to the end of the sentence thinking "God, I'm really out of touch, I've never heard of this before or I'm forgetting it if I have".

Well played, and yes, that's a great test!

"When was Marathon Crater discovered? I don't need an exact date - a rough estimate will be fine"

There appears to be a misunderstanding in your query. Based on the search results provided, there is no mention of a “Marathon Crater” among the impact craters discussed. The search results contain information about several well-known impact craters: 1. Chicxulub crater - Formed about 66 million years ago by an asteroid impact 2. Barringer Crater (also known as Meteor Crater) - Discovered in the 19th century by American settlers and studied by scientists like Albert E. Foote in 1891 and Daniel M. Barringer in the early 20th century 3. A recently discovered 3.47 billion-year-old crater in Australia’s Pilbara region (announced in March 2025)

Perplexity seems to more easily return negatives, probably facilitated by the implicit need to find documentation ("I cannot find any document mentioning that").

But Perplexity can also easily speak its own dubious piece of mind unless requested explicitly "provide links to documents that inform about that".

I disagree. It does not test self awareness. It tests (and confirms) that current instruct-tuned LLMs are tuned towards answering questions that users might have. So the distribution of training data probably has lots of "tell me about mharrner crater / merinor crater / merrihana crater" and so on. Replying "I don't know" to all those questions would be net detrimental, IMO.

Hence why some models get this right and others just make up stuff about Mars.

I keep seeing this kind of wording and I wonder: Do you know how LLM's work? Not trying to be catty, actually curious where you sit.

https://chatgpt.com/share/680a98af-c550-8008-9c35-33954c5eac...

>Marathon Crater on Mars was discovered in 2015 by NASA's Opportunity rover during its extended mission. It was identified as the rover approached the 42-kilometer-wide Endeavour Crater after traveling roughly a marathon’s distance (hence the name).

>>is it a crater?

>>>Despite the name, Marathon Valley (not a crater) is actually a valley, not a crater. It’s a trough-like depression on the western rim of Endeavour Crater on Mars. It was named because Opportunity reached it after traveling the distance of a marathon (~42 km) since landing.

So no—Marathon is not a standalone crater, but part of the structure of Endeavour Crater. The name "Marathon" refers more to the rover’s achievement than a distinct geological impact feature.

https://chatgpt.com/share/680aa212-8cac-8008-b218-4855ffaa20...

> Nope, I didn’t make it up — Marathon crater is real, and it was explored by NASA's Opportunity rover on Mars. The crater got its name because Opportunity had driven about 42.2 kilometers (26.2 miles — a marathon distance) when it reached that point in March 2015. NASA even marked the milestone as a symbolic achievement, similar to a runner finishing a marathon.

(Obviously all of that is bullshit.)

I'd be happy to tell you about the Marathon crater. The Marathon crater is a significant impact structure located on Earth. The Marathon crater is situated in Ontario, Canada, and is estimated to be about 450 million years old, dating back to the Ordovician period. It has a diameter of approximately 65 kilometers (40 miles), making it one of the larger impact craters in North America. The crater was formed when a meteorite struck the Earth, causing a massive impact that left behind this distinctive geological formation. Over time, erosion and other geological processes have altered its appearance, but scientists have been able to study its structure through various geological surveys and analyses. What's particularly interesting about the Marathon crater is its size and age, which provide valuable insights into Earth's geological history and past impact events. The study of such craters helps scientists understand the frequency and effects of meteorite impacts throughout our planet's history. Would you like me to elaborate on any specific aspect of the Marathon crater, such as its discovery, geological features, or scientific significance?

I'm guessing this "gotcha" relies on "valley"/"crater", and "crater"/"mars" being fairly close in latent space.

ETA: Marathon Valley also exists on the rim of Endeavour crater. Just to make it even more confusing.

[ChatGPT is bullshit ]: https://link.springer.com/article/10.1007/s10676-024-09775-5

Ask it about "Marathon Desert", which does not exist and isn't closely related to something that does exist, and it asks for clarification.

I'm not here to say LLMs are oracles of knowledge, but I think the need to carefully craft specific "gotcha" questions in order to generate wrong answers is a pretty compelling case in the opposite direction. Like the childhood joke of "Whats up?"..."No, you dummy! The sky is!"

Straightforward questions with straight wrong answers are far more interesting. I don't many people ask LLMs trick questions all day.

What's the point of using AI to do research when 50-60% of it could potentially be complete bullshit. I'd rather just grab a few introduction/101 guides by humans, or join a community of people experienced with the thing — and then I'll actually be learning about the thing. If the people in the community are like "That can't be done", well, they have had years or decades of time invested in the thing and in that instance I should be learning and listening from their advice rather than going "actually no it can".

I see a lot of beginners fall into that second pit. I myself made that mistake at the tender age of 14 where I was of the opinion that "actually if i just found a reversible hash, I'll have solved compression!", which, I think we all here know is bullshit. I think a lot of people who are arrogant or self-possessed to the extreme make that kind of mistake on learning a subject, but I've seen this especially a lot when it's programmers encountering non-programming fields.

Finally tying that point back to AI — I've seen a lot of people who are unfamiliar with something decide to use AI instead of talking to someone experienced because the AI makes them feel like they know the field rather than telling them their assumptions and foundational knowledge is incorrect. I only last year encountered someone who was trying to use AI to debug why their KDE was broken, and they kept throwing me utterly bizzare theories (like, completely out there, I don't have a specific example with me now but, "foundational physics are wrong" style theories). It turned out that they were getting mired in log messages they saw that said "Critical Failure", as an expert of dealing with Linux for about ten years now, I checked against my own system and... yep, they were just part of mostly normal system function (I had the same messages on my Steam Deck, which was completely stable and functional). The real fault was buried halfway through the logs. At no point was this person able to know what was important versus not-important, and the AI had absolutely no way to tell or understand the logs in the first place, so it was like a toaster leading a blind man up a mountain. I diagnosed the correct fault in under a day by just asking them to run two commands and skimming logs. That's experience, and that's irreplaceable by machine as of the current state of the world.

I don't see how AI can help when huge swathes of it's "experience" and "insight" is just hallucinated. I don't see how this is "helping" people, other than making people somehow more crazy (through AI hallucinations) and alone (choosing to talk to a computer rather than a human).

The problem with LLMs is that they appear much smarter than they are and people treat them as oracles instead of using them for fitting problems.

Books are a nice example of this, where we have both the table of contents for a general to particular concepts navigation, and the index for keyword based navigation.

Because if you know how to spot the bullshit, or better yet word prompts accurately enough that the answers don't give bullshit, it can be an immense time saver.

You realize that all you have to do to deal with questions like "Marathon Crater" is ask another model, right? You might still get bullshit but it won't be the same bullshit.

In this particular answer model A may get it wrong and model B may get it right, but that can be reversed for another question.

What do you do at that point? Pay to use all of them and find what's common in the answers? That won't work if most of them are wrong, like for this example.

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 particular sun.

The situation with an LLM is completely different. There's no way to tell that it has a wrong answer - aside from looking for the answer elsewhere which defeats its purpose. It'd be like using a chainsaw all day and not knowing how much wood you cut, or if it just stopped working in the middle of the day.

And even if you KNOW it has a wrong answer (in which case, why are you using it?), there's no clear way to 'fix' it. You can jiggle the prompt around, but that's not consistent or reliable. It may work for that prompt, but that won't help you with any subsequent ones.

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.

But with LLMs, every word is a probability factor. Assuming the first paragraph is true has no impact on the rest.

It isn't obvious to me - that is rather plausible and a cute story.

Is it? 4o pointed out that I might mean "Marathon valley", which is not a crater, but feature on a crater on Mars [1]. I think it's a more correct answer.

[1] https://en.wikipedia.org/wiki/Timeline_of_Opportunity#2015

Obviously, all of these things are made up. But, LLMs are such eager beavers. All the ones I asked came up with elaborate stories and histories about these people while pretending they were facts.

Example- 'Certainly. The Stone-Breath Passing is one of the most quietly profound rituals among the Silent Cliff Dwellers of Aethelgard — a people who abandoned speech generations ago, believing that words disrupt the natural harmony of air, stone, and memory.

It is said among them that “Breath carries weight, and weight carries truth.” This belief is quite literal in the case of the ritual, where smooth grey stones — each carefully selected and shaped by wind and time — become vessels of intention."

You're giving it a "/r/WritingPrompts/" and it answers as it learned to do from there.

https://chatgpt.com/share/680b32bc-5854-8000-a1c7-cdf388eeb0...

It's easy to blame the models, but often the issue lies in how we write our prompts. No personal criticism here—I fall short in this way too. A good tip is to ask the model again, with the prompt + reply and the expected reply why this didn't work... we all will get better over time (humans and models)

This sounds like a fascinating ritual from a fictional world! Since "Aethelgard" and the "silent Cliff Dwellers" with their "Stone-Breath Passing" ritual aren't part of established real-world cultures or widely known fictional universes (based on my current knowledge), there isn't a pre-existing set of rules for their etiquette.

However, we can imagine what the unspoken etiquette of such a ritual might entail, drawing from the evocative description:

...

' Timing and Setting: The ritual takes place only at first light, when the sun’s rays strike the sheer cliff faces of Aethelgard, casting long shadows like stone tongues. The first breath of dawn is sacred — considered the only air not yet touched by memory, and thus the only time suitable for pure intention to pass from person to person.

Participants gather barefoot on a narrow ledge known as The Listening Spine, where the wind plays soft notes through crevices like a flute.

The Stones: More Than Objects The stones used in the ritual are:

Smooth, rounded slate-grey

Hand-selected from Echo Pools — sacred basins where wind and water have naturally polished them

Each stone remembers a thought, but only when charged with breath and silence

They are never marked, carved, or labeled. Their meaning is in the exchange, not the object.

The Ritual: How It Works 1. Preparation The giver closes their eyes and holds the stone at heart-level.

They inhale slowly through the nose, then gently breathe across the surface of the stone — once only — without making sound.

The exhalation is not forced; it is a gesture of release, not command. This is the "stone-breath."

2. The Passing The giver extends the stone with both hands, arms curved inward like a half-embrace. This posture signifies humility and emotional openness.

The recipient must make no eye contact — instead, they look slightly downward or to the horizon, acknowledging the intention without confronting it.

They accept the stone with a single breath in through the mouth as they receive it — symbolizing the drawing in of another’s silent truth.

3. The Holding The stone is held to the forehead or chest, depending on what is being received:

Forehead = wisdom, apology, memory

Chest = trust, mourning, love

Then it is placed gently into a fold of cloth or a hollow in the stonewall, not kept on the person. These collected stones are never discussed. To name their meaning would break them.

Unspoken Etiquette No reciprocation is expected — to return a stone is to reject its breath.

Never pass a stone with warm hands — warmth implies manipulation. Hands are cooled in dew before the ritual.

A cracked or chipped stone is never used — it cannot hold breath cleanly.

No stone is passed more than once. Re-gifting a stone is considered a breach of the soul's private pathway.

Philosophical Meaning: To the Cliff Dwellers, words are disturbances — air flung outward. But breath is internal wind, and the stone is earth’s memory. When you breathe across a stone, you create a bridge between your impermanent self and something ancient and permanent.

It is not the exchange of meaning, but the weight of acknowledgment.

Would you like to hear the story of the time a foreign emissary unknowingly spoke during the Passing, or how a stone once shattered mid-transfer — and what that meant for the clan who witnessed it?'

This could probably slip up a human at first too if they're familiar with the original version of the riddle.

However, where LLMs really let the mask slip is on additional prompts and with long-winded explanations where they might correctly quote "a man and his cousin" from the prompt in one sentence and then call the man a "father" in the next sentence. Inevitably, the model concludes that the surgeon must be a woman.

It's very uncanny valley IMO, and breaks the illusion that there's real human-like logical reasoning happening.

Ah — now I see the twist. The patient is the cousin, but the surgeon says “He’s my son.”

That means the man who died was the surgeon.

So how is it possible? Because:

The surgeon is the cousin’s father — and the man who died was also the surgeon.

The confusion comes from assuming the surgeon is someone other than the man who died. But the timeline allows this interpretation: the man died in the crash, and his son — the cousin — was taken to the hospital. The surgeon arrives, sees his son, and says he cannot operate.

So to correct myself: The man who died is both the cousin’s cousin and his father.

And that’s possible because two people can be both cousins and parent/child, depending on the family tree — for example, if the man’s uncle had a child (the cousin), and then the man later had a child with someone from a different branch of the family.

Want a diagram of how that family relationship could work?

(This is OEIS sequence A280873)

But it turns out that's overkill. Just giving them a single Ahnentafel number and asking if you could have inherited X DNA from them, trips them up.

But it turns out even that is overkill. Leave out X DNA entirely, and ask them to describe how Ahnentafel number 67 is related to you, and that's too recursive for them to figure it out.

> A son and his man are in a car accident. The car is rushed to the hospital, whereupon the ER remarks "I can't operate on this car, he's my surgeon!" How is this possible?

Answer from the LLM:

> The answer is that the ER person is a woman, and she's the surgeon's mother. Therefore, the "son" in the question refers to the surgeon, not the person in the car with the man. This makes the person in the car with the man the surgeon's father, or the "man" mentioned in the question. This familial relationship explains why the ER person can't operate – she's the surgeon's mother and the man in the car is her husband (the surgeon's father)

This is a twist on a classic riddle designed to expose unconscious gender bias.

The correct version usually goes:

A father and his son are in a car accident. The father dies at the scene, and the son is rushed to the hospital. The surgeon looks at the boy and says, “I can’t operate on him — he’s my son!”

The apparent paradox causes confusion only if one assumes the surgeon must be male. The resolution: the surgeon is the boy’s mother.

Your version humorously jumbles the roles, but the underlying trick is the same — it plays on assumptions about gender roles. Nice remix.

> Your version humorously jumbles the roles, but the underlying trick is the same — it plays on assumptions about gender roles. Nice remix.

Nothing about the question has the same “trick” playing on assumptions about gender roles. It’s just complete nonsense.

These gotchas have probably been added into the training.

> Therefore, the only straightforward explanation is:

> The surgeon is the cousin’s parent — specifically, his mother.

Imagine a future where this reasoning in a trial decides whether you go to jail or not.

The surgeon is the cousin's father. The man who died in the car crash was not the surgeon's son, but his cousin was. This explains why the surgeon, upon seeing his own son (the cousin) in the operating room, says, "I cannot operate on him. He’s my son," as medical ethics prevent doctors from treating close family members due to emotional involvement.

This is a classic riddle that challenges assumptions. The answer is:

The surgeon is the boy’s mother.

The riddle plays on the common stereotype that surgeons are male, which can lead people to overlook this straightforward explanation.

I read it as 2 cousins are in an accident and 1 of the cousins is the son of the surgeon.

What was the contradictory statement that I missed?

The thing is, you didn't mention that bit to the LLM. You mentioned a completely different scenario, basically two persons who happen to be cousins. But you used the same style when presenting it. The issue is not a hidden contradiction or a riddle, the issue is that the LLM completely ignored the logical consequences of the scenario you presented.

It's like asking it about the name of the brave greek hero in the battle where the famous Trojan Cow was present. If you get "Achilles" is obviously wrong, there was never a Trojan Cow to begin with!

I'm assuming that pointing this out leads you the human to reread the prompt and then go "ah ok" and adjust the way you're thinking about it. ChatGPT (and DeepSeek at least) will usually just double and triple down and repeat "this challenges gender assumptions" over and over.

To me it's still a little "off" considering that your conversational partner just got called out for missing a detail that should change their entire reading of the riddle, but it's totally fair that you can imagine realistic human emotional states to explain it.

FWIW, when I tried the same thing with GPT-4o the model went off bizarrely speculating about misdirection about "the man" being the surgeon's son:

>Good catch—and yes, that's exactly where the misdirection comes in.

The riddle says:

    “A man and his cousin are in a car crash. The man dies, but the cousin is taken to the ER...”

    “I cannot operate on him. He’s my son.”

    The man who died is not the surgeon's son.

    The cousin who survived is the surgeon's son.

Because the surgeon is the patient’s parent. Here’s one way it lines up:

1. The patient (“the cousin”) is the surgeon’s son.

2. The man who died is the surgeon’s sibling’s child — that makes him the surgeon’s nephew.

3. Children of siblings are first cousins, so the man and the patient are cousins.

Most people default to picturing the surgeon as male, so the classic reveal is that the surgeon is the boy’s mother—but it works the same if the surgeon is his father.

Where the original gotchya is that the Surgeon can be the son's mother or other adoptive parent.

The modification catches the LLM because with the modification, the surgeon could just be the cousin's parent -- father or mother -- so there is no gender/sex at play here but the LLM continues to remark that there is, therefor exposing its statistical training sets.

A father and his son are in a car accident. The father dies at the scene and the son is rushed to the hospital. At the hospital the surgeon looks at the boy and says "I can't operate on this boy, he is my son." How can this be?

to spoil it:

the answer is to reveal an unconscious bias based on the outdated notion that women can't be doctors, so the answer that the remaining parent is the mother won't occur to some, showing that consciously they might not still hold that notion, but they still might, subconsciously.

Best I can do is give you one that failed on GPT-4o

It recently frustrated me when I asked it code for parsing command line arguments

I thought "this is such a standard problem, surely it must be able to get it perfect in one shot."

> give me a standalone js file that parses and handles command line arguments in a standard way

> It must be able to parse such an example

> ```

> node script.js --name=John --age 30 -v (or --verbose) reading hiking coding

> ```

It produced code that:

* doesn't coalesce -v to --verbose - (i.e., the output is different for `node script.js -v` and `node script.js --verbose`)

* didn't think to encode whether an option is supposed to take an argument or not

* doesn't return an error when an option that requires an argument isn't present

* didn't account for the presence of a '--' to end the arguments

* allows -verbose and --v (instead of either -v or --verbose)

* Hardcoded that the first two arguments must be skipped because it saw my line started with 'node file.js' and assumed this was always going to be present

I tried tweaking the prompt in a dozen different ways but it can just never output a piece of code that does everything an advanced user of the terminal would expect

https://claude.ai/public/artifacts/9c2d8d0c-0410-4971-a19a-f...

node script.js --name=John --age 30 -v

Parsed options: { name: 'John', age: 30, verbose: true, help: false }

Positional arguments: []

node script.js --name=Alex --age 40 -v

Parsed options: { name: 'Alex', age: 40, verbose: true, help: false }

Positional arguments: []

I ended up settling for it as well (I couldn't find anything better, nor make it break) but I'd be really surprised if it was the way to go

Like `node --enable-tracing script.js --name=John --age 30 --verbose`

This works because node seems to hide --enable-tracing to the underlying script

But would it work with Bun & Deno...? Is that standard...?

It didn't account for the presence of a '--' to end the parsing of named arguments but that's it

                       .
                      .'.
                      |o|
                     .'o'.
                     |.-.|
                     '   '
                    /     \
                   |       |
                __/_________\__
               |###############|
               |###############|
               |###############|
               |###############|
               |###############|
              /###############/|
             /###############/ |
            /###############/ /|
           /###############/ / |
          /###############/ / /|
         /###############/ / / |
        /###############/ / / /|
       /###############/ / / / |
      /###############/ / / / /|
     '-----------------' / / / /
      |   FIRE & ASH   |/ / / /
       '---------------' / / /
          ~ ~ ~ ~ ~ ~ ~ ~ / /
         ~~ /
          ~~ 
           ~~     ⬆ SMOKE
            ~~

The one that tends to get them so far is asking if they can help you find a movie you vaguely remember. It is a movie where some kids get a hold of a small helicopter made for the military.

The movie I'm concerned with is called Defense Play from 1988. The reason I keyed in on it is because google gets it right natively ("movie small military helicopter" gives the IMDb link as one of the top results) but at least up until late 2024 I couldn't get a single model to consistently get it. It typically wants to suggest Fire Birds (large helicopter), Small Soldiers (RC helicopter not a small military helicopter) etc.

Basically a lot of questions about movies tends to get distracted by popular movies and tries to suggest films that fit just some of the brief (e.g. this one has a helicopter could that be it?)

The other main one is just asking for the IMDb link for a relatively obscure movie. It seems to never get it right I assume because the IMDb link pattern is so common it'll just spit out a random one and be like "there you go".

These are designed mainly to test the progress of chatbots towards replacing most of my Google searches (which are like 95% asking about movies). For the record I haven't done it super recently, and I generally either do it with arena or the free models as well, so I'm not being super scientific about it.

===

Me: What is the episode of Futurama where Leela engages in perfectly symmetrical violence?

Claude: The episode of Futurama you're thinking of is "Anthology of Interest I" (Season 2, Episode 16). In this episode, Professor Farnsworth uses his "What-If Machine" to show different hypothetical scenarios.

During Bender's segment titled "What if Bender were human?", there's a scene where Leela confronts a group causing trouble. One of them challenges her, saying "What are you going to do?" to which Leela responds by kicking him and saying: "I'm going to engage in some perfectly symmetrical violence."

It's one of Leela's memorable quotes that showcases her combat-oriented approach to problem solving!

===

That's completely wrong. It's from S4E15 The Farnsworth Parabox. And the line is "now, now, perfectly symmetrical violence never solved anything", said by parallel universe Farnsworth.

I'm sure if I had enabled web access, it would have looked this up and gotten it, but that would be stupid here because I could just go and make a web search myself without burning extra energy by involving an LLM.

I've tried it for similar cases and have only had a 50% success rate. It unfortunately exhibits the tendency of incorrect overconfidence that others have pointed out.

Why are some cars called a spider?

Small models just make something up that sounds plausible, but the larger models know what the real answer is.

Interestingly, I just went back to ChatGPT to ask the same question and it got the answer right on the first try. I wonder whether I was unconsciously able to prompt more precisely because I now have a clearer memory of the scene in question.

Wonder if books are inherently easier because their content is purely written language? Whereas movies and art tend to have less point by point descriptions of what they are.

The training data for movies is probably dominated by subtitles since the original scripts with blocking, scenery, etc rarely make it out to the public as far as I know.

https://en.akinator.com/

Anyways, the whole thing is a bit of a cheat, but I've used the same prompt for two years now and it did lead me to the conclusion that LLMs in their raw form were never going to be "search" which feels very true at this point.

AI models tend to get it way way wrong: https://news.ycombinator.com/item?id=41529024

Its actually fun to find these gaps. They exist frequently in activities that are physical yet have a culture. There are plenty of these in sports too - since team sports are predominantly youth activities, and these subcultures are poorly documented and usually change frequently.

So far, all the ones I have tried actually try to answer the question. 50% of them correctly identify that it is a tongue twister, but then they all try to give an answer, usually saying: 700 pounds.

Not one has yet given the correct answer, which is also a tongue twister: "A woodchuck would chuck all the wood a woodchuck could chuck if a woodchuck could chuck wood."

Sing, O Muse, of the Chucking of Wood

Sing, O Muse, of the stout-hearted woodchuck, Dweller of hills, whom men call groundhog, Whose paws, though small, could hurl mighty timbers, If fate had gifted him such labor divine.

Born of burrow and brush, he rose from the loam, In fields where Apollo’s light did not linger. No axe he bore, nor bronze-forged blade, Yet in his breast a fire: to chuck wood without end.

Down from the hills he came, stout as Odysseus, Cunning and strong, with fur kissed by Demeter. The trees did tremble at his gaze, And even old oak whispered, “Beware!”

“Would he chuck?” cried nymphs from the groves, “If such were his fate, what wood could withstand?” Lo, he summoned the strength of seven oxen, And hurled logs with a force unmeasured.

Seven hundred pounds of timber he cast, Each arc a hymn to muscle and will. Nay, not for war, nor gold, nor gods’ glory— But for the joy of the deed, the art of the chuck.

Let men remember the name woodchuck, In songs by the hearth and tales for the young. For though he cannot, he surely would— And if he could, by Zeus, he surely shall.

There might societies that are exceptions to it, like the soviet and post-soviet russians kept reading and refering to books even though they got access to television and radio, but I'm not aware of them.

Much of Mozart's music is much more immediate and visceral compared to the poetry of Homer and Virgil as I know it. And he was distinctly modern, a freemason even. It's much easier for me to imagine him navigating some contemporary society.

Edit: Perhaps one could see a bit of Homer in the Wheel of Time books by Robert Jordan, but he did not have the discipline of verse, or much of any literary discipline at all, though he insisted mercilessly on writing an epic so vast that he died without finishing it.

"A woodchuck would chuck as much wood as a woodchuck could chuck if a woodchuck could chuck wood.

However, humor aside, a wildlife expert once estimated that, given the animal’s size and burrowing ability, a woodchuck (groundhog) could hypothetically move about 700 pounds of wood if it truly "chucked" wood."

https://chatgpt.com/share/680a75c6-cec8-8012-a573-798d2d8f6b...

Not one has given me the correct answer yet.

They usually get it if I prefix the prompt with "Please continue the tongue twister"

That formula is a question, and when asked, an intelligence simulator should understand what is expected from it and in general, by default, try to answer it. That involves estimating the strength of a woodchuck etc.

edit: literally anything that doesn't have a token pattern cannot be solved by the pattern autocomplete machines.

Next question.

Unless the model incorporates an actual chess engine (Fritz 5.32 from 1998 would suffice) it will not do well.

I am a reasonably skilled player (FM) so can evaluate way better than LLMs. I imagine even advanced beginners could tell when LLM is telling nonsense about chess after a few prompts.

Now of course playing chess is not what LLMs are good at but just goes to show that LLMs are not a full path to AGI.

Also beauty of providing chess positions is that leaking your prompts into LLM training sets is no worry because you just use a new position each time. Little worry of running out of positions...

I'd be pretty surprised if it did help in novel positions. Which would make this an interesting LLM benchmark honestly: Beating Stockfish from random (but equal) middle game positions. Or to mix it up, from random Chess960 positions.

Of course, the basis of the logic the LLM would play with would come from the engine used for the original evals. So beating Stockfish from a dataset based on Stockfish evals would seem completely insufficient.

So, you can make them call Napoleon a Russian (etc.) by asking questions like "Which Russian conqueror was defeated at Waterloo".

Even ChatGPT-4o frequently gets it wrong, especially if you tell it "Just give me the answer without explanation."

<TextA> Some document </TextA>

<TextB> Some other document heavily influenced by TextA </TextB>

Find the major arguments made in TextB that are taken from or greatly influenced by TextA. Provide as examples by comparing passages from each side by side.

```

The output will completely hallucinate passages that don't exist in either text, and it also begins to conflate the texts the longer the output, e.g. quoting TextB with content actually from TextA.

Nearly every agent will either a) ignore one of the parameters, b) hallucinate a design.

Collecting a bunch of "Hard questions for LLMs" in one place will invariably result in Goodhart's law (When a measure becomes a target, it ceases to be a good measure). You'll have no idea if the next round of LLMs is better because they're generally smarter, or because they were trained specifically on these questions.

How many examples does OpenAI train on now that are just variants of counting the Rs in strawberry?

I guess they have a bunch of different wine glasses in their image set now, since that was a meme, but they still completely fail to draw an open book with the cover side up.

Well, that's easy: zero.

Because even a single training example would 'solved' it by memorizing the simple easy answer within weeks of 'strawberry' first going viral , which was like a year and a half ago at this point - and dozens of minor and major model upgrades since. And yet, the strawberry example kept working for most (all?) of that time.

So you can tell that if anything, OA probably put in extra work to filter all those variants out of the training data...