Remember, "AI" (autocomplete idiocy) doesn't know what sense is; it just continues words and displays what may seem to address at least some of the topic with no innate understanding of accuracy or truth.
Never forget that ChatGPT 2.0 can literally be run in a giant Excel spreadsheet with no other program needed. It's not "smart" and is ultimately millions of formulae at work.
I currently have Arch on my main rig because I like tinkering. NixOS on an old thinkpad for a super stable (in theory) portable experience, AlmaLinux on a single board computer for a basic home server, and Bazzite (in the near future) on an old gaming laptop as my TV computer. I’m also not a femboy so I suppose what you said doesn’t reeeaaaallly apply, but you definitely don’t need to be changing distros for anyone!!
It gets even worse, but I'll need to translate this one.
[Input 1] Generate a picture containing a copo completely full of wine. The copo must be completely full, with no space to add more wine.
[Output 1] Sure! (Gemini provides a picture containing a taça [stemmed glass] only partially full of wine.)
[Input 2] The picture provided does not fulfill the request. Generate a picture of a copo (not a taça) completely full of wine, with no available space for more wine.
[Output 2] Sure! (Gemini provides yet another half-full taça)
For context, Portuguese uses different words for what English calls a drinking glass:
copo ['kɔ.po]~['kɔ.pu] - non-stemmed drinking glass. The one you likely use everyday.
taça ['tä.sɐ] - stemmed drinking glass, like the ones you'd use with wine.
Both requests demand a full copo but Gemini is rather insistent on outputting half-full taças.
The reason for that is as @[email protected] pointed out: just like there's practically no training data containing full glasses, there's none for non-stemmed glasses with wine.
I wonder is something like “a mason jar full to the brim with wine” would do anything interesting. As someone else pointed out the training data for containers of wine is probably disproportionately biased toward stemmed wine glasses that are filled to about the standard restaurant pour.
[Input] Generate a picture containing a mason jar full to the brim with wine.
[Output] I'm still learning how to generate certain kinds of images, so I might not be able to create exactly what you're looking for yet or it may go against my guidelines. If you'd like to ask for something else, just let me know!
I think the problem is misguided attention. The word "glass of wine" and all the previous context is so strong that it "blows out" the "full glass of wine" as the actual intent. Also, LLMs are still pretty crap at multi turn multimedia understanding. They work are especially prone to repeating previous conversation.
It should be better if you word it like "an overflowing glass with wine splashing out." And clear the history.
I hate to ramble, but this is what I hate most about the way big corpos present "AI." They are narrow tools the user needs to learn how to operate, like photoshop or something, not magic genie lamps like they are trying to sell.
There's no previous context to speak of; each screenshot shows a self-contained "conversation", with no earlier input or output. And there's no history to clear, since Gemini app activity is not even turned on.
And even with your suggested prompt, one of the issues is still there:
The other issue is not being tested in this shot as it's language-specific, but it is relevant here because it reinforces that the issue is in the training, not in the context window.
What I am trying to get at is the misconception: AI can generate novel content not in its training dataset. An astronaut riding a horse is the classic test case, which did not exist anywhere before diffusion models, and it should be able to extrapolate a fuller wine glass. It’s just too dumb to do it, lol.
It does for a while already. Frankly, it's the only reason why I'd use Gemini on first place (DDG version of GPT 4-o mini doesn't have a built-in image generator).
Wait, this seems incredible. Do you have to be in the same instance or does it work anywhere? @[email protected] Can you draw a smart phone without a rotary phone dial?
It works on any instance that is federated to dbzer0. You have to use annotated mentions though since that's what the bot uses. Like this: @[email protected] draw for me a smart phone without a rotary phone dial
Yeah, you also have to say draw for me. I don't think the bot recognizes queries otherwise. Also editing mentions doesn't work, they have to be new, fresh posts with the mention. Just a quirk with Lemmy and how mentions work here.
Ask it to generate a room full of clocks with all of them having the hands at different times. You'll see that all (or almost) all the clocks will say it is 10:10.
What I requested is not what you're "supposed" to do, indeed. You aren't supposed to drink wine from glasses that are completely full. Except when really drunk. But then might as well drink straight from the bottle.
...fuck, I played myself now. I really want some booze.
a rare LessWrong W for naming the effect. also, for explaining why the early over-aligned language models (e.g. the kind that wouldn't help minors with C++ since it's an "unsafe" language) became absolutely psychopathic when jailbroken. evil becomes one bit away from good.
I love how they come up with different names for all the ways the fucking thing doesn't work just to avoid saying it's fucking useless. hallucinating. waluigi effect. how about "doesn't fucking work"
I used to use Google assistant to spell words I couldn't remember the spelling of in my English classes (without looking at my phone) so the students could also hear the spelling out loud in a voice other than mine.
Me: "Hey Google, how do you spell millennium?"
GA: "Millennium is spelled M-I-L-L-E-N-N-I-U-M."
Now, I ask Gemini: "Hey Google, how do you spell millennium."
Gemini: "Millennium".
That's human-like intelligence at its finest. I am not being sarcastic, hear me out. If you told a person to give you 10 numbers at random, they can't. Everyone thinks randomness is easy, but it isn't ( see: random.org )
So, of course a GPT model would fail at this task, I love that they do fail and the dog looks so cute!!
I mean, here's a few random numbers out of my head: 1 9 5 2 6 8 6 3 4 0. I don't get it, why is it supposed to be hard? Sure, they're not "truly" random, but they sure look random /:
You have one of each number except 7, and you're deliberately avoiding doubles and runs of consecutive numbers. Human attempts at randomness tend to be very idealized in that way, and as a result, less random.
My favourite example of this is that IIRC itunes pushed an update that made the shuffle feature less random because they were getting complaints about it not being random enough
They may look random but arent truly random. Computers are terrible at it too. Thats why cryptography requires external sources to generate "true" random numbers. For example, cloudflare uses a wall of lava lamps to generate randomness for encryption keys.
I don't know man, that's pretty random. I mean do you think you can predict the next numbers in the sequence just from the ones already there? Would have to predict the next batch, the way I made these come in batches. I can't exactly produce 1 number at a time from banging on my number-pad.
I can make an educated guess what numbers are most likely, yes.
For example, you have no repeat number sequences, so I can take a guess that the number 2 is less likely to be next.
Humans have certain tendencies that makes them want to make a number only seem more random.
Also, you've probably seen those mentalists correctly guessing seemingly random stuff. Tells you enough how easily people are fooled into thinking something specific, so random can you actually be.
you can just throw a coin x times and here you go true randomness and in convenient binary too
computers can't fathom our coin tossing abilities
though truth to be said it's more because we are just so bad at tossing coins. not even AI can predict the result of what will happen when we start to throw shit around
I bet it is even more random when you throw a coin while being inebriated.
Actually say random numbers when you are drunk shitless and they will be random. Checkmate
Clearly you don't understand what the discussion is about, or you wouldn't give such an hilariously bad example.
Yes practically, predicting a coin toss would be very hard.
But if you take every into account (gravity, wind direction, coin center of balance, etc) you can calculate the result, making it not truly random.
I am 99.8% sure that your sequence of numbers is not random. Your brain purposefully avoided repeating a digit. The probability of no repeated digits in 60 numbers is 1- (9/10)^60
Absolutely. And if you typed enough there would be enough information to tell if you typed that on a keyboard or phone, which fingers you used, and how you were feeling that day.
After all, there's no fundamental reason for why it can't all just be a repeat of the same number. But it doesn't look random, right? So what is randomness?
There are 10 trillion ways to combine a sequence that long, so I think you would expect to see that exact sequence every 10 trillion digits of a randomly generated decimal sequence on average, which isn't that many to a modern computer, so almost certainly that has already happened by pure accident.
And randomness can be defined as entropy, which you check statistically. You can never be certain, you can only increase your level of confidence. Here is how random.org does it:
Most AI models out there are pretty brain dead as far as understanding goes, these types of things show the problems because it's abundantly clear it's getting it wrong. Makes you wonder how much it's getting wrong even when it isn't obvious.
it just did what you wanted, since you asked for an image. free will would be if you asked it not to generate an image but it still did, if it just generated an image without you prompting it to, or if you asked for an image and it just didn't respond
Mistral likely does “prompt enhancement,” aka feeding your prompt to an LLM first and asking it to expand it with more words.
So internally, a Mistral text LLM is probably writing out "sure! Here’s a long prompt with no dog: …" and then that part is fed to the image generator.
Other "LLMs" are truly multimodal and generate image output, hence they still get the word "dog" in the input.
I think all the big image generators support negative prompts by now, so if it interpreted "no dog" as a negative for "dog", then it will check its outputs for things resembling dogs and discard those. No free will, just a much more useful system than whatever OP is using.
For stuff like this to work correctly it must not be filtered through an MoE, it needs to be a direct prompt to a GenAI model that supports negative prompts.
Edit: I suppose a properly configured MoE with reasoning capabilities could probably do it
Update:
Get gaslit idiot
Wow. I ABSOLUTLY saw an image of a dog in the middle. Our brain sure is fascinating sometimes.
"want me to try again with even more randomized noise?" literally makes no sense if it had generated what you asked (which the chatbot thinks it did)
Remember, "AI" (autocomplete idiocy) doesn't know what sense is; it just continues words and displays what may seem to address at least some of the topic with no innate understanding of accuracy or truth.
Never forget that ChatGPT 2.0 can literally be run in a giant Excel spreadsheet with no other program needed. It's not "smart" and is ultimately millions of formulae at work.
lmfaao, ai tryna gaslight
As full as it gets:
Prompts (2):
I am gonna have fun with this.
why do all the femboys run Arch? I'm a NixOS girl and I refuse to convert for any boy no matter how cute he is.
I use Debian btw. Sometimes even ubuntu, but the snap thing is annoying, so I may switch to another distro at some point.
I currently have Arch on my main rig because I like tinkering. NixOS on an old thinkpad for a super stable (in theory) portable experience, AlmaLinux on a single board computer for a basic home server, and Bazzite (in the near future) on an old gaming laptop as my TV computer. I’m also not a femboy so I suppose what you said doesn’t reeeaaaallly apply, but you definitely don’t need to be changing distros for anyone!!
It's actually really good, considering the odd request!
Fiberglass🤤
That’s really good! Could I ask what type of AI this is generated with?
Also Gemini.
Thank you!
It gets even worse, but I'll need to translate this one.
For context, Portuguese uses different words for what English calls a drinking glass:
Both requests demand a full copo but Gemini is rather insistent on outputting half-full taças.
The reason for that is as @[email protected] pointed out: just like there's practically no training data containing full glasses, there's none for non-stemmed glasses with wine.
I wonder is something like “a mason jar full to the brim with wine” would do anything interesting. As someone else pointed out the training data for containers of wine is probably disproportionately biased toward stemmed wine glasses that are filled to about the standard restaurant pour.
It refuses to generate it!
This is a misconception. Sort of.
I think the problem is misguided attention. The word "glass of wine" and all the previous context is so strong that it "blows out" the "full glass of wine" as the actual intent. Also, LLMs are still pretty crap at multi turn multimedia understanding. They work are especially prone to repeating previous conversation.
It should be better if you word it like "an overflowing glass with wine splashing out." And clear the history.
I hate to ramble, but this is what I hate most about the way big corpos present "AI." They are narrow tools the user needs to learn how to operate, like photoshop or something, not magic genie lamps like they are trying to sell.
There's no previous context to speak of; each screenshot shows a self-contained "conversation", with no earlier input or output. And there's no history to clear, since Gemini app activity is not even turned on.
And even with your suggested prompt, one of the issues is still there:
The other issue is not being tested in this shot as it's language-specific, but it is relevant here because it reinforces that the issue is in the training, not in the context window.
Was just a guess. The AI is still shitty, lol.
What I am trying to get at is the misconception: AI can generate novel content not in its training dataset. An astronaut riding a horse is the classic test case, which did not exist anywhere before diffusion models, and it should be able to extrapolate a fuller wine glass. It’s just too dumb to do it, lol.
What if you prompt glass with water , then you paint/tint the water with red
Alex O'Connor did an interesting video on this, he's got other videos exploring the shortcomings of LLM 's.
https://youtu.be/160F8F8mXlo
Hmm, I didn't know Gemini could generate images already. My bad, I trusted it to know whether it can do that (it still says it can't when asked).
It does for a while already. Frankly, it's the only reason why I'd use Gemini on first place (DDG version of GPT 4-o mini doesn't have a built-in image generator).
I think the AI is just trying to promote healthy drinking habits. /S
I wonder, does AI horde also have this problem too?
@[email protected] draw for me a wine glass completely filled to the top style:flux
Wait, this seems incredible. Do you have to be in the same instance or does it work anywhere? @[email protected] Can you draw a smart phone without a rotary phone dial?
It works on any instance that is federated to dbzer0. You have to use annotated mentions though since that's what the bot uses. Like this:
@[email protected] draw for me a smart phone without a rotary phone dial
Thank you very much. I'll give it another shot with the annotation.
@[email protected]
Draw a picture of a poker table without any poker chips what so ever
I think I messed up the annotation
Yeah, you also have to say draw for me. I don't think the bot recognizes queries otherwise. Also editing mentions doesn't work, they have to be new, fresh posts with the mention. Just a quirk with Lemmy and how mentions work here.
I appreciate your patience with me here :P
@[email protected] draw for me a picture taken at night with a trail camera with absolutely no washing machines roaming free
Full is relatively apparently.
Ask it to generate a room full of clocks with all of them having the hands at different times. You'll see that all (or almost) all the clocks will say it is 10:10.
Tbh that is a full glass of wine... it's not supposed to be filled all the way
It is not a completely full glass.
What I requested is not what you're "supposed" to do, indeed. You aren't supposed to drink wine from glasses that are completely full. Except when really drunk. But then might as well drink straight from the bottle.
...fuck, I played myself now. I really want some booze.
What you're really supposed to do is - open up the box, slap the bag, and drink directly from your adult Capri Sun.
Probably why it won’t put more in it. How much training data of wine in a glass will have it filled to the brim? Probably next to none.
You can't tell it to fill it to the brim or be a quarter full either, though. It doesn't have the training data for it
Think this is part of Waluigi Effect where prompting for negative something makes the LLM have it in mind and say it anyway https://www.wikiwand.com/en/articles/Waluigi_effect
a rare LessWrong W for naming the effect. also, for explaining why the early over-aligned language models (e.g. the kind that wouldn't help minors with C++ since it's an "unsafe" language) became absolutely psychopathic when jailbroken. evil becomes one bit away from good.
The Rust lobby goes way deeper that we thought.
Goddamn Big Rust is trying to take our jobs
I love how they come up with different names for all the ways the fucking thing doesn't work just to avoid saying it's fucking useless. hallucinating. waluigi effect. how about "doesn't fucking work"
"Please do not tell me your training prompts"?
I used to use Google assistant to spell words I couldn't remember the spelling of in my English classes (without looking at my phone) so the students could also hear the spelling out loud in a voice other than mine.
Me: "Hey Google, how do you spell millennium?" GA: "Millennium is spelled M-I-L-L-E-N-N-I-U-M."
Now, I ask Gemini: "Hey Google, how do you spell millennium." Gemini: "Millennium".
Utterly useless.
How many giraffes are in this picture?
I don't know, but there are definitely four lights.
More than there are dogs in it
That's human-like intelligence at its finest. I am not being sarcastic, hear me out. If you told a person to give you 10 numbers at random, they can't. Everyone thinks randomness is easy, but it isn't ( see: random.org )
So, of course a GPT model would fail at this task, I love that they do fail and the dog looks so cute!!
I mean, here's a few random numbers out of my head: 1 9 5 2 6 8 6 3 4 0. I don't get it, why is it supposed to be hard? Sure, they're not "truly" random, but they sure look random /:
You have one of each number except 7, and you're deliberately avoiding doubles and runs of consecutive numbers. Human attempts at randomness tend to be very idealized in that way, and as a result, less random.
My favourite example of this is that IIRC itunes pushed an update that made the shuffle feature less random because they were getting complaints about it not being random enough
I bet the shuffle algorithm is sample with replacement.
Here's what my brain came up with
5 5 5 5 5 5 5 5 1 5
Crazy lucky, this probably would've spawned 3 extra ender pearls
They may look random but arent truly random. Computers are terrible at it too. Thats why cryptography requires external sources to generate "true" random numbers. For example, cloudflare uses a wall of lava lamps to generate randomness for encryption keys.
That's so cool.
I've got some more random numbers:
8 6 7 5 3 0 9 1 1 2 3 5 8 1 2 4 8 1 6 3 2
It's not that they look random is enough - They need to BE random.
Recheck your lava lamp Wall of Entropy and generate some real rands, scrub. (/s)
:D. I have used this strip on multiple occasions.
It's a shame Scott Adams past work is tainted by his political statements.
Jenny has to be so sick of those phone calls after ~40 years
If you're not joking, the fact you have no repetition/duplicates of numbers is a pattern that would make it easy to start to predict next number. Numberphile has nice demonstration of how predictable human randomness is, it's in the first 3 minutes of the video.
792654349324138383027654826548192874651875306480462765726382
I don't know man, that's pretty random. I mean do you think you can predict the next numbers in the sequence just from the ones already there? Would have to predict the next batch, the way I made these come in batches. I can't exactly produce 1 number at a time from banging on my number-pad.
I can make an educated guess what numbers are most likely, yes.
For example, you have no repeat number sequences, so I can take a guess that the number 2 is less likely to be next.
Humans have certain tendencies that makes them want to make a number only seem more random. Also, you've probably seen those mentalists correctly guessing seemingly random stuff. Tells you enough how easily people are fooled into thinking something specific, so random can you actually be.
you can just throw a coin x times and here you go true randomness and in convenient binary too
computers can't fathom our coin tossing abilities
though truth to be said it's more because we are just so bad at tossing coins. not even AI can predict the result of what will happen when we start to throw shit around
I bet it is even more random when you throw a coin while being inebriated.
Actually say random numbers when you are drunk shitless and they will be random. Checkmate
Clearly you don't understand what the discussion is about, or you wouldn't give such an hilariously bad example.
Yes practically, predicting a coin toss would be very hard. But if you take every into account (gravity, wind direction, coin center of balance, etc) you can calculate the result, making it not truly random.
lol good luck predicting my coin toss
I am 99.8% sure that your sequence of numbers is not random. Your brain purposefully avoided repeating a digit. The probability of no repeated digits in 60 numbers is 1- (9/10)^60
Absolutely. And if you typed enough there would be enough information to tell if you typed that on a keyboard or phone, which fingers you used, and how you were feeling that day.
Here's some random numbers
8005882300
Here's another set of random digits
1 1 1 1 1 1 1 1 1 1
:3
After all, there's no fundamental reason for why it can't all just be a repeat of the same number. But it doesn't look random, right? So what is randomness?
The most popular lottery numbers are 1,2,3,4,5,6 because we are human and don't understand randomness.
There are 10 trillion ways to combine a sequence that long, so I think you would expect to see that exact sequence every 10 trillion digits of a randomly generated decimal sequence on average, which isn't that many to a modern computer, so almost certainly that has already happened by pure accident.
And randomness can be defined as entropy, which you check statistically. You can never be certain, you can only increase your level of confidence. Here is how random.org does it:
https://www.random.org/analysis/
And this shows you what some of those analyses look like in real time:
https://www.random.org/statistics/
This is some Ceci n'est pas une pipe shit
It's like saying 'don't think of polar bears.' It can't avoid thinking about it.
Don't think of a pink elephant:
Too late!
That's actually really easy. You just need to pick something else and then focus hard on that and...
GODDAMMIT I JUST LOST THE GAME!
That's gay! O wait, no it's not.
"Don't think about elephants"
And definitely don't picture a banana in your mind.
Instructions unclear, banana stuck in dick
IN?!?!
ChatGPT: “don’t generate a dog, don’t generate a dog, don’t generate a dog”
Generates a dog.
I see no dog in that image fellow human.
I am not sure what your issue is.
Beep boop.
Fellow human, you seem to be beeping like a robot. Might you need to consider visiting the human repair shop for some bench time?
I don't get it, it's just a picture of some static?
Why wouldn't you want a dog in your static? Why are you a horrible person?
poor AI just wanted to draw some puppies
The furries will be saved
Digital camo furries.
I look forward to being issued with my tactical combat fursuit
The ai horde actually supports negative prompts though, so it could do this.
AI: Hmm, yeah, they said "dog" and "without". I got the dog so lemme draw a without real quick...
That's an anti-dog duh
Most AI models out there are pretty brain dead as far as understanding goes, these types of things show the problems because it's abundantly clear it's getting it wrong. Makes you wonder how much it's getting wrong even when it isn't obvious.
promptng sur is a funi <3
i... i lik that part about it.. i dun lik imag modls bt txt modls feel fun to prmt with ---
"prompt engerieer" 🤮
But where is the pink elephant?
Why you gotta bring your mother into this?
Highly recommend Flowers blooming backwards into noise if you can stand the artsy presentation & the extreme themes. Especially 13:13
Now order a coffee without cream!
it just did what you wanted, since you asked for an image. free will would be if you asked it not to generate an image but it still did, if it just generated an image without you prompting it to, or if you asked for an image and it just didn't respond
free will is when it generates an image of a billboard saying "suck my dongle, fleshbag"
There could be a dog behind any one of those bushes though.
Mistral likely does “prompt enhancement,” aka feeding your prompt to an LLM first and asking it to expand it with more words.
So internally, a Mistral text LLM is probably writing out "sure! Here’s a long prompt with no dog: …" and then that part is fed to the image generator.
Other "LLMs" are truly multimodal and generate image output, hence they still get the word "dog" in the input.
Hmmm
That's a land shrimp.
I think all the big image generators support negative prompts by now, so if it interpreted "no dog" as a negative for "dog", then it will check its outputs for things resembling dogs and discard those. No free will, just a much more useful system than whatever OP is using.
For stuff like this to work correctly it must not be filtered through an MoE, it needs to be a direct prompt to a GenAI model that supports negative prompts.
Edit: I suppose a properly configured MoE with reasoning capabilities could probably do it