Had a lot of fun with my stats students today.

[poleguy@mastodon.social]

@futurebird The trouble is that people can accept that "factual" output from an LLM may be statistically generated until they hit words that are generated that sound like "reasoning." Then even the most aware humans can get lulled into thinking that the words can be trusted.

[gkrnours@mastodon.gamedev.place]

@futurebird I assume from this post someone already mentioned statistics from the python standard library?

[seachaint@masto.hackers.town]

@futurebird there was a study that found that if you give an LLM some prompting to push it into a particular sampling-space (say, "bleeding heart leftie") and then ask it for some random numbers, you can then feed those numbers into another fresh instance and it'll drift towards the same sampling space.

In other words, even the numerical distributions they sample from can be connected to the broader "noosphere" they're trained on, and that relation is a fucked sort of bijection

[seachaint@masto.hackers.town]

@futurebird if you prompt it into "stats prof" or "crypto nerd" sampling space does it improve the quality of the fake RNG output?

[david_chisnall@infosec.exchange]

@futurebird @Bumblefish

It’s a trick question. Neither list is random because 7 is the most random number and does not appear in either list. A six-sided die is not able to produce a 7 and cannot therefore produce a random number.

- ChatGPT, probably.

[tschfflr@fediscience.org]

@futurebird @Bumblefish I vote for listB: I counted the times that two subsequent numbers are equal (1,1 or 4,4). In listA this occurs ~23 times so almost 1/4 of times, which seems too many (should be around 1/6). In listB it is ~9 times unless I missed some. Seems fewer than expected but anyway. If I’d spend more time I’d go for higher order ngrams

[cstross@wandering.shop]

@okohll @futurebird I was about to suggest Benford's Law too!

[meuwese@mastodon.social]

@ai6yr @ohmu @futurebird wait so... is that the ultimate question? "What number will an LLM always include when generating random numbers?"

[life_is@no-pony.farm]

@burnitdown@beige.party @futurebird@sauropods.win raNDOm. A play on words.

[okohll@hachyderm.io]

@cstross @futurebird God does play dice, but there’s a big lead weight in one side

[thisalex@hachyderm.io]

@futurebird
> what are we doing?

I think that the best description is, that we take part in a play. LLM makes its best effort to write how this dialogue could continue to look plausible for the reader. Choose your own adventure.

[mildouze@mamot.fr]

@futurebird @Bumblefish
B
(Random answer)

[lamecarlate@pouet.it]

@futurebird @Bumblefish I'm no stats student, so maybe I haven't the bases (for lack of a better term, English is not my main language), but I think listA is the random one. The fact that in the listB there is nearly no triplets seems too good to be true.

[ingalovinde@embracing.space]

@AbyssalRook @futurebird I see two mistakes in your reasoning.
One is technical: events "numbers with position N, N+1 and N+2 are the same" for different values of N are _not_ independent of each other. (For example, if we know that this statement is true for N=10, then there likelihood of it being true for N=11 is 1/6, not 1/36.)
Another symbolizes a deeper problem with a lot of modern research that relies heavily on p-values: consider how many statements of this kind, containing the same amount of information, could you make? Unless you commit to a specific statement beforehand, before seeing the data: "this statement would only be true in 8% of cases for truly random data" does not really mean anything if it's just one out of 20 equally "interesting" statements one could make about the data (e.g. "how many triplets of incrementing numbers (modulo six) are there", "how many decrementing triplets are there", etc), each only 8% likely. Because of course it is expected that for most random sequences, a few of these individually not very likely statements will be true.

[futurebird@sauropods.win]

@lamecarlate @Bumblefish

I've got some bad news. I've posted the solution with a CW on the original thread.

[futurebird@sauropods.win]

@IngaLovinde @AbyssalRook

It's been really helpful for me to see how many people focused on the order of the numbers in the list, which I didn't think very important since the list is so short that that type of analysis might not be that useful.

I used the random list to scramble the fake numbers twice. I should have scrambled them more.

[abyssalrook@mstdn.social]

@IngaLovinde I'm not following the first problem in the logic. The situation you're describing might be important if we're looking at more and more instances of it happening, but looking at it happening at least once (~94%) doesn't change at all, and it happening ONLY once might jiggle the ~8% estimate I had, but not significantly move it.

[flockofcats@famichiki.jp]

@Bumblefish @futurebird
That was an interesting thread. Our brains are wired to think certain things are “random” when they’re not, so when people try to create something that looks random, they often avoid repeated numbers, even though there’d be repeats, if truly random, with some expected frequency. Also, odd numbers are often overrepresented cuz they feel more random, e.g., 5973 vs 6084. This “ looks random, but isn’t” often comes up when people fabricate scientific data

[abyssalrook@mstdn.social]

@IngaLovinde As for the latter, that is entirely true from a research perspective, but I picked the 3-of-a-kind pattern because I assumed the non-random list was entirely human constructed, and that particular pattern is one that sticks out to us the most. Someone making a list by hand is more likely to see "6-6-6" as less random than "6-1-2" or "3-4-5".

I did not clock 'Which is random?' as one being a dice roll and the other being a shuffled deck of prescribed cards.

[fsologureng@chilemasto.casa]

@futurebird listA has the subsequence 1,1,1,6,1,4 repeated twice at very short distance between them, which is, while plausible, extremely improbable. That's the way I found it's crafted.