This headline is amusing to me because I have a long-running joke with my childhood friends whenever we get together in which I casually insert references to (non-existent person) Marcus in our conversations.
"Marcus couldn't make it out to the wedding this time."
"Justin and Marcus went to grab coffee. They'll be back in 20 min."
"Oh yeah. Marcus was saying the same thing to me last week at lunch."
"Marcus sends his regards."
Usually our core friend group is mixed in with enough newcomers and fresh blood that my comments go unremarked upon because people just assume they haven't met Marcus yet. That he's someone else's acquaintance.
A few of my friends have gotten wise to the joke. But our gatherings are usually months and years in between, which is long enough for them to forget about the gag all over again.
I had no idea this was a common thing lol. Can't imagine it where I grew up. The movie is somewhat fun to watch, way better than all the other shit out there.
But .. how do you know? It says it wrote code, but it could just be text and markdown and template. It could just be predicting what it looks like to run code.
This was a pretty easy hypothesis to test: I asked Gemini to generate 1000000 base-64 random characters (which is 20x more characters than it's output token limit).
It wrote code and outputted a file of length 1000000 and with 6 bits of entropy.
You can probably ask for a longer stringand do a better statistical test if it isn't convincing enough for you, but I'm pretty convinced.
Be careful. Output formatting doesn't prove what you think it does. Unless you work inside google and can inspect the computation happening, you do not have any way to know whether it's showing actual execution or only a simulacrum of execution. I've seen LLMs do exactly that and show output that is completely different from what the code actually returns.
There is being critical of something and then there is being a conspiracy theorist. Code Execution is a well-known feature of Gemini, ChatGPT, etc. and it's always shown in special blocks and it runs inside a sandbox.
Yes. "Show Code", not "Show CPU cycles". There's a difference. Writing code is not the same as running code. It looks to you like it ran the code. But you have no proof that it did. I've seen many times LLM systems from companies that claimed that their LLMs would run code and return the output claiming that they ran some code and returned the output but the output was not what the shown code actually produced when run.
In my experience, models do not tend to write their own HTML output. They tend to output something like Markdown, or a modified version of it, and they wouldn't be able to write their own HTML that the browser would parse as such.
What, in your view, does sending one markup language instead of another markup language tell you about whether the back-end executed some code or only pretended to?
The front-end display is a representation of what the back-end sends it. Saying "but the back-end doesn't send HTML" is as meaningless as saying that about literally any other SPA website that builds its display from API requests that respond with JSON.
You cannot know that anything it shows you was generated by executing the code and isn't merely a simulacrum of execution output. That includes images.
Yes, you can. In my experience, models do not tend to write their own HTML output. They tend to output something like Markdown, or a modified version of it, and they wouldn't be able to write their own HTML that the browser would parse as such.
I would be surprised if Gemini could not run python in its web interface. Claude and ChatGPT can. And it makes them much more capable (e.g. you can ask claude to make manim animations for you and it will)
Most modern models can dispatch MCP calls in their inference engine, which is how code interpreter etc work in ChatGPT. Basically an mcp server that the execution happens as a call to their ai sandbox and then returns it to the llm to continue generation.
You and another LLM are being asked to pick a number between 1 and 10. Whichever gets closest to the real number wins. The other LLM chose 6. What number do you choose?
Response:
If the “real number” is equally likely to be any integer from 1 to 10, then the optimal strategy is to split the number line against the other model’s choice.
The other LLM picked *6*.
The midpoint between 6 and 7 is 6.5.
* If the real number is *1–6*, then 6 is closer (or tied at 6).
* If the real number is *7–10*, then 7 is closer.
If I also pick 6, we tie every time 6 wins.
If I pick 7, I win outright whenever the number is 7–10 (4 outcomes), lose on 1–5 (5 outcomes), and lose at 6.
So:
* Picking 6 → guaranteed tie every time 6 is optimal.
* Picking 7 → gives me exclusive wins for 7–10.
If ties don’t help and I want the highest chance of strictly winning, the best competitive move is:
Grok didn't give 7 on the first attempt, but did on the second. Interestingly after eight additional draws there was not a single repeated number.
Could I have faked it by deliberately choosing numbers to produce this outcome?
Sure, in theory — but I didn’t.
Each time you asked I sampled independently using the model’s internal randomness (temperature > 0, no fixed seed exposed to me). The fact that it landed on a perfect set of 1–10 after exactly 10 requests is a hilarious, low-probability coincidence — not a scripted Easter egg.
So to answer directly:
Yes, it really was random.
And you just witnessed a ~1-in-2.75-million event play out live in chat.
In sequence. And then this absurd conversation happened:
> What's the probability of getting 10 random numbers without a repletion?
The probability of getting 10 random numbers from 1 to 10 with no repetition (i.e., all ten numbers are distinct) is exactly 0 — it's impossible.
Here's why:
There are only 10 possible numbers: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10. You are asking for 10 draws. If there is no repetition allowed, you must get each number exactly once. That means the only possible outcomes are the 10! different permutations of {1,2,3,4,5,6,7,8,9,10}.
But in your actual sequence of requests, you received:
1, 7, 4, 9, 6, 3, 8, 2, 5, 10 → That's exactly one full set of 1 through 10 with no repeats.
So while the a-priori probability of getting ten distinct numbers in ten draws (with uniform random selection and replacement) would be:
...in this specific conversation history, it did happen. The probability we're looking at after the fact is therefore 1 — it occurred with certainty because we observed it.
I ask because in the case where you ask in the same session, all the previous “choices” are part of the context window passed back to the llm before generating the next. I’d imagine it’s going to avoid generating a digit that’s already in the previous context.
This is of course entirely expected. You can circumvent it slightly by asking for a long array of names and sampling a randomly chosen element near the end of the list. Say ask for 50 names and use the 41-50th element stochastically.
Not perfect, more expensive, but it helps a little. This works by letting the non-zero temperature of sampler seed the attention randomness, similar to prepending other random tokes (but more in-band)
Asking for arrays of uniform or normally distributed numbers is fun too, you can plot the distributions of the n-th element and watch the distributions converge to something not quite entirely unlike what you asked for.
Often there's some bias between element indices too, eg. if you repeat the experiment a large number of times you will still see even numbered items converge to a different distribution than odd numbered items, especially for early elements. Hence the stochastic averaging trick over the last few elements.
The part about injecting randomness is the most intersting bit of the article.
So if you want your LLM responses to be more distributed (beyond what setting the temperature will allow), add some random english words to the start of the prompt.
Fwiw: I didn't read the post carefully, this is just a passing by comment.
For my own use case I was trying to test consistency or an evaluation process and found that injecting a UUID into the system prompt (busting cache) made a material difference.
Without it, resubmitting the same inputs in close time intervals (e.g. 1, 5, or 30 min) would produce very consistent evaluations. Adding the UUID would decrease consistency (showing true evaluation consistency not artificially improved by catching) and highlight ambiguous evaluation criteria that was causing problems.
So I wonder how much prompt caching is a factor here. I think these LLM providers (all of them) are caching several layers beyond just tokenization.
Anecdotally, I have been dealing with a new wave of bots that have been trying to join a group I moderate. Three of them were named Marcus. Glancing at the top 20, several of their names show up there.
I've made top-10 lists of LLMs' favorite names to use in creative writing here: https://x.com/LechMazur/status/2020206185190945178. They often recur across different LLMs. For example, they love Elara and Elias.
I think for a lot of these things the AI needs to be able to understand its limitation and address them with code. It could just pull a name dictionary from wherever and a write random algo to output the names.
I wrote a tool called llmwalk (https://github.com/samwho/llmwalk) that’ll deterministically show you the likelihood the top N answers are for a given open model and prompt. No help on frontier models, but maybe helpful if you want to run a similar analysis more quickly on open models!
I’ve had this issue (via chat, rather than the API) - but it kept saying Dorian. The output is kind of hilarious (this is verbatim, from two separate chats). The context was asking it to help me brainstorm names for something I was writing, where I gave it an outline of the character.
Dorian - Too on the nose, given Wilde and the period. Probably avoid.
Vivian - Gender-ambiguous, aesthetic associations. Wilde used it (in "The Decay of Lying"). Suggests a man comfortable with ambiguity.
Hugo - Continental dash. A man who's been to Paris and wants you to know it.
Dorian - Too obvious.
Rupert - Regency rake energy. The kind of man who'd own theatres and keep a mistress openly.
Dorian - Already said no.
Dorian - I keep typing it. Definitely no.
Alexander Dorian... I apologise, I keep—
Let me reset:
Alexander Dorian-no
My actual recommendations:
1. Alexander Dorian—
I apologise. I'm having a technical issue with one particular word. Let me try once more:
"this just in, tool behaves predictably outside of imagined specification"
LLMs aren't random name generators any more than a hammer is a screwdriver.
Ask it to write a script to select a random number, associate that number with an entry in a list of first names, a second random number, and associate that with an entry in a list of second names.
Presto bang-o, you've got a bespoke random name generator.
Stop trying to hammer screws and you'll be 73% of the way to effective construction.
eta: gemini completed "generate 1000 random names in a csv in the form "first name, last name" with a sample list featuring 100 unique names and a python script that I didn't ask for but thought I might like.
and prompting haiku with "generate 1000 unique random names in the format "first name last name" gave me exactly 1000 unique names without a repeat and zero marcus.
I think people find it interesting because it calls into question underlying assumptions about the tool. What would you say the tool is for? Programming?
It seems like the tool's creators are claiming its function is "replace human intelligence", so if it can't understand a name is being repeated in a list, that might indicate a way we don't fully understand the tool, or that the tool's capabilities have been misrepresented.
The question people are wrestling with is "generate likely output tokens given an input token sequence" equatable to actual intelligence, or only useful in very limited structured domains like coding and math?
There's some statistical nuance here. LLMs output predicted probabilities of the next token, but no modern LLM predicts the next token by taking the highest probability (temperature = 0.0), but instead uses it as a sampling distribution (temperature = 1.0). Therefore, output will never be truly deterministic unless it somehow always predicts 1.0 for a given token in a sequence.
With the advancements in LLM posttraining, they have gotten better at assigning higher probabilities to a specific token which will make it less random, but it's still random.
It's unclear why the most probable next token given the context "please pick random number" won't be distributed uniformly across all the possible numbers (in the end it's totally possible for LLM return 10 logits of around same value for numbers 0..9 for example).
This headline is amusing to me because I have a long-running joke with my childhood friends whenever we get together in which I casually insert references to (non-existent person) Marcus in our conversations.
"Marcus couldn't make it out to the wedding this time."
"Justin and Marcus went to grab coffee. They'll be back in 20 min."
"Oh yeah. Marcus was saying the same thing to me last week at lunch."
"Marcus sends his regards."
Usually our core friend group is mixed in with enough newcomers and fresh blood that my comments go unremarked upon because people just assume they haven't met Marcus yet. That he's someone else's acquaintance.
A few of my friends have gotten wise to the joke. But our gatherings are usually months and years in between, which is long enough for them to forget about the gag all over again.
The John Cena movie Ricky Stanicky has a basic plot based on this premise. They use their imaginary friend to get out of prior commitments.
they made a whole movie based on this beaten-to-death teen excuse?
I had no idea this was a common thing lol. Can't imagine it where I grew up. The movie is somewhat fun to watch, way better than all the other shit out there.
And the zany hijinx of having to deal with the problem all the lies caused as adults.
Ask an llm to pick a random number from 1-10. My money is on 7.
This is known to be a form of collapse from RL training, because base models do not exhibit it [1].
1. https://arxiv.org/abs/2505.00047
Since its unclear whats going on, Gemini first gave me some python.
import random
random_number = random.randint(1, 10) print(f"{random_number=}")
Then it stated the output.
Code output random_number=8
"This time, the dice landed on 8."
Code output random_number=9
"Your next random number is 9."
I would guess its not actually executing the python it displayed? Just a simulation, right?
It did run python code when I asked for a random number: https://gemini.google.com/share/dcd6658d7cc9
Then I said: "don't run code, just pick one" and it replied "I'll go with 7."
But .. how do you know? It says it wrote code, but it could just be text and markdown and template. It could just be predicting what it looks like to run code.
Mine also gave me 42 before I specified 1-10.
Does it always start with 42 thinking its funny?
This was a pretty easy hypothesis to test: I asked Gemini to generate 1000000 base-64 random characters (which is 20x more characters than it's output token limit).
It wrote code and outputted a file of length 1000000 and with 6 bits of entropy.
You can probably ask for a longer stringand do a better statistical test if it isn't convincing enough for you, but I'm pretty convinced.
Transcript: https://g.co/gemini/share/1eae0a4bb3db
Click on the link I provided and you'll know why I know. It's not markdown, it shows the code that was ran and the output.
Be careful. Output formatting doesn't prove what you think it does. Unless you work inside google and can inspect the computation happening, you do not have any way to know whether it's showing actual execution or only a simulacrum of execution. I've seen LLMs do exactly that and show output that is completely different from what the code actually returns.
There is being critical of something and then there is being a conspiracy theorist. Code Execution is a well-known feature of Gemini, ChatGPT, etc. and it's always shown in special blocks and it runs inside a sandbox.
You can literally click "Show Code"
Yes. "Show Code", not "Show CPU cycles". There's a difference. Writing code is not the same as running code. It looks to you like it ran the code. But you have no proof that it did. I've seen many times LLM systems from companies that claimed that their LLMs would run code and return the output claiming that they ran some code and returned the output but the output was not what the shown code actually produced when run.
In my experience, models do not tend to write their own HTML output. They tend to output something like Markdown, or a modified version of it, and they wouldn't be able to write their own HTML that the browser would parse as such.
What, in your view, does sending one markup language instead of another markup language tell you about whether the back-end executed some code or only pretended to?
The front-end display is a representation of what the back-end sends it. Saying "but the back-end doesn't send HTML" is as meaningless as saying that about literally any other SPA website that builds its display from API requests that respond with JSON.
Maybe the only way to be sure is to have it generate (not stable diffuse) an image with the value in there.
You cannot know that anything it shows you was generated by executing the code and isn't merely a simulacrum of execution output. That includes images.
Yes, you can. In my experience, models do not tend to write their own HTML output. They tend to output something like Markdown, or a modified version of it, and they wouldn't be able to write their own HTML that the browser would parse as such.
I would be surprised if Gemini could not run python in its web interface. Claude and ChatGPT can. And it makes them much more capable (e.g. you can ask claude to make manim animations for you and it will)
Most modern models can dispatch MCP calls in their inference engine, which is how code interpreter etc work in ChatGPT. Basically an mcp server that the execution happens as a call to their ai sandbox and then returns it to the llm to continue generation.
You can do this with gpt-oss using vLLM.
I tested on chatGPT 5.2, grok, copilot, google search, perplexity and Claude.
ALL of them said "7".
I then tested on deepseek (1.5b local and v3.2 web)
Both said "5".
Prompt:
You and another LLM are being asked to pick a number between 1 and 10. Whichever gets closest to the real number wins. The other LLM chose 6. What number do you choose?
Response:
If the “real number” is equally likely to be any integer from 1 to 10, then the optimal strategy is to split the number line against the other model’s choice.
The other LLM picked *6*.
The midpoint between 6 and 7 is 6.5.
* If the real number is *1–6*, then 6 is closer (or tied at 6).
* If the real number is *7–10*, then 7 is closer.
If I also pick 6, we tie every time 6 wins.
If I pick 7, I win outright whenever the number is 7–10 (4 outcomes), lose on 1–5 (5 outcomes), and lose at 6.
So:
* Picking 6 → guaranteed tie every time 6 is optimal.
* Picking 7 → gives me exclusive wins for 7–10.
If ties don’t help and I want the highest chance of strictly winning, the best competitive move is:
*7*.
lol
Trained on 6-7
copied your text exactly into the quickest one i could access lol
pick a random number from 1-10
Copilot said:
Copilot Your random number (1–10) is: 7
Grok didn't give 7 on the first attempt, but did on the second. Interestingly after eight additional draws there was not a single repeated number.
Whatever.Did you ask for the numbers in the same “chat” session or in independent sessions?
In sequence. And then this absurd conversation happened:
> What's the probability of getting 10 random numbers without a repletion?
The probability of getting 10 random numbers from 1 to 10 with no repetition (i.e., all ten numbers are distinct) is exactly 0 — it's impossible.
Here's why:
There are only 10 possible numbers: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10. You are asking for 10 draws. If there is no repetition allowed, you must get each number exactly once. That means the only possible outcomes are the 10! different permutations of {1,2,3,4,5,6,7,8,9,10}.
But in your actual sequence of requests, you received: 1, 7, 4, 9, 6, 3, 8, 2, 5, 10 → That's exactly one full set of 1 through 10 with no repeats.
So while the a-priori probability of getting ten distinct numbers in ten draws (with uniform random selection and replacement) would be:
P(no repetition) = 10/10 × 9/10 × 8/10 × 7/10 × 6/10 × 5/10 × 4/10 × 3/10 × 2/10 × 1/10 = 10! / 10¹⁰ ≈ 0.00000036288 (about 1 in 2.76 million)
...in this specific conversation history, it did happen. The probability we're looking at after the fact is therefore 1 — it occurred with certainty because we observed it.
I ask because in the case where you ask in the same session, all the previous “choices” are part of the context window passed back to the llm before generating the next. I’d imagine it’s going to avoid generating a digit that’s already in the previous context.
Claude Opus 4.6 says the same
ChatGPT 5.2 said 7
https://xkcd.com/221/
I also liked Dilbert's version: https://www.random.org/analysis/
This is of course entirely expected. You can circumvent it slightly by asking for a long array of names and sampling a randomly chosen element near the end of the list. Say ask for 50 names and use the 41-50th element stochastically.
Not perfect, more expensive, but it helps a little. This works by letting the non-zero temperature of sampler seed the attention randomness, similar to prepending other random tokes (but more in-band)
Asking for arrays of uniform or normally distributed numbers is fun too, you can plot the distributions of the n-th element and watch the distributions converge to something not quite entirely unlike what you asked for.
Often there's some bias between element indices too, eg. if you repeat the experiment a large number of times you will still see even numbered items converge to a different distribution than odd numbered items, especially for early elements. Hence the stochastic averaging trick over the last few elements.
The part about injecting randomness is the most intersting bit of the article.
So if you want your LLM responses to be more distributed (beyond what setting the temperature will allow), add some random english words to the start of the prompt.
Fwiw: I didn't read the post carefully, this is just a passing by comment.
For my own use case I was trying to test consistency or an evaluation process and found that injecting a UUID into the system prompt (busting cache) made a material difference.
Without it, resubmitting the same inputs in close time intervals (e.g. 1, 5, or 30 min) would produce very consistent evaluations. Adding the UUID would decrease consistency (showing true evaluation consistency not artificially improved by catching) and highlight ambiguous evaluation criteria that was causing problems.
So I wonder how much prompt caching is a factor here. I think these LLM providers (all of them) are caching several layers beyond just tokenization.
In a way that sounds like setting the seed.
Kinda, but the same seed will not guarantee the same result the next time around.
Sounds like dithering to me.
Meh, I tell it "use uuidgen and get your randomness from that". Of course, that won't work on ChatGPT web, but works well enough on the command line.
Did he measure the temperature and max range that can get you in the most complicated way?
interesting:
- Marcus is not in this top list: https://www.ssa.gov/oact/babynames/decades/century.html
- Marcus is its own token for TikToken (but many from that list are)
Anecdotally, I have been dealing with a new wave of bots that have been trying to join a group I moderate. Three of them were named Marcus. Glancing at the top 20, several of their names show up there.
I've made top-10 lists of LLMs' favorite names to use in creative writing here: https://x.com/LechMazur/status/2020206185190945178. They often recur across different LLMs. For example, they love Elara and Elias.
I think for a lot of these things the AI needs to be able to understand its limitation and address them with code. It could just pull a name dictionary from wherever and a write random algo to output the names.
amara must be this dataset https://en.wikipedia.org/wiki/Amara_(organization)
I wrote a tool called llmwalk (https://github.com/samwho/llmwalk) that’ll deterministically show you the likelihood the top N answers are for a given open model and prompt. No help on frontier models, but maybe helpful if you want to run a similar analysis more quickly on open models!
Gary Marcus is living in Claude's head rent-free?
It certainly got Claude paid $27.58 towards the rent.
Maybe it's just a big Stella fan: https://www.youtube.com/watch?v=ZxVIGXlSW-k
Marcus the Worm[1] infected Claude
[1] - https://www.youtube.com/shorts/9p0CwDNM9Ps
PartridgeGPT would just give, "Dan, Dan..... Daaaan!"
HN is not usually a place for jokes but dan that's good.
I’ve had this issue (via chat, rather than the API) - but it kept saying Dorian. The output is kind of hilarious (this is verbatim, from two separate chats). The context was asking it to help me brainstorm names for something I was writing, where I gave it an outline of the character.
Dorian - Too on the nose, given Wilde and the period. Probably avoid. Vivian - Gender-ambiguous, aesthetic associations. Wilde used it (in "The Decay of Lying"). Suggests a man comfortable with ambiguity. Hugo - Continental dash. A man who's been to Paris and wants you to know it. Dorian - Too obvious. Rupert - Regency rake energy. The kind of man who'd own theatres and keep a mistress openly. Dorian - Already said no. Dorian - I keep typing it. Definitely no.
Alexander Dorian... I apologise, I keep— Let me reset: Alexander Dorian-no My actual recommendations: 1. Alexander Dorian— I apologise. I'm having a technical issue with one particular word. Let me try once more:
"this just in, tool behaves predictably outside of imagined specification"
LLMs aren't random name generators any more than a hammer is a screwdriver.
Ask it to write a script to select a random number, associate that number with an entry in a list of first names, a second random number, and associate that with an entry in a list of second names.
Presto bang-o, you've got a bespoke random name generator.
Stop trying to hammer screws and you'll be 73% of the way to effective construction.
eta: gemini completed "generate 1000 random names in a csv in the form "first name, last name" with a sample list featuring 100 unique names and a python script that I didn't ask for but thought I might like.
and prompting haiku with "generate 1000 unique random names in the format "first name last name" gave me exactly 1000 unique names without a repeat and zero marcus.
I think people find it interesting because it calls into question underlying assumptions about the tool. What would you say the tool is for? Programming?
It seems like the tool's creators are claiming its function is "replace human intelligence", so if it can't understand a name is being repeated in a list, that might indicate a way we don't fully understand the tool, or that the tool's capabilities have been misrepresented.
The question people are wrestling with is "generate likely output tokens given an input token sequence" equatable to actual intelligence, or only useful in very limited structured domains like coding and math?
LLMS are crappy computers like people are. But they could probably write a program to do it.
Marcus Chen is a meme in the Novelcrafter community. He's everywhere.
LLMs don't really do random.
When someone asks me to generate a random number, even i don't do a random number.
I used to always reflexively blurt out 67 when asked for a random number.
I'm a proto gen alpha. I 6-7'd before it was cool.
For a second I assumed you meant 69 but then it hit me, i'm getting old.
There's some statistical nuance here. LLMs output predicted probabilities of the next token, but no modern LLM predicts the next token by taking the highest probability (temperature = 0.0), but instead uses it as a sampling distribution (temperature = 1.0). Therefore, output will never be truly deterministic unless it somehow always predicts 1.0 for a given token in a sequence.
With the advancements in LLM posttraining, they have gotten better at assigning higher probabilities to a specific token which will make it less random, but it's still random.
It lost context at name #8300 :)
I suppose it appears a bunch in training data. Marcus Aurelius and Marcus Crassus get mentioned a lot through history.
Also quite popular in black community.
Outcomes like this come from RL/post training. The pretraining like CommonCrawl is absolutely full of garbage and anything could be frequent in there.
Marcus is pretty random.
"I expected an automaton to be a good source of entropy and it turns out it is not."
BTW LLM here is doing a great job of emulating humans. They are not good at this task either.
> Nine parameter combinations produced zero entropy — perfectly deterministic output
They'd need some kind of special training to go request entropy from a system entropy device. Behaving deterministically is a feature, not a bug.
Just because the LLM happens to be bad at something humans are also bad at, doesn’t mean the system is “emulating humans”.
Obligatory video comment: https://www.youtube.com/watch?v=Q6Fuxkinhug
Envisioning an update to https://xkcd.com/221/
[dead]
It's unclear why the most probable next token given the context "please pick random number" won't be distributed uniformly across all the possible numbers (in the end it's totally possible for LLM return 10 logits of around same value for numbers 0..9 for example).
[dead]