Note that my tests were via groq and the r1 70B distilled llama variant (the 2nd smartest version afaik)
Edit 1:
Incidentally… I propositioned a coworker to answer the same question. This is the summarized conversation I had:
Me: “Hey Billy, can you answer a question? in under 3 seconds answer my following question”
Billy: “sure”
Me: “How many As are in abracadabra 3.2.1”
Billy: “4” (answered in less than 3 seconds)
Me: “nope”
I’m gonna poll the office and see how many people get it right with the same opportunity the ai had.
Edit 2:
The second coworker said “6” in about 5 seconds
Edit 3:
Third coworker said 4, in 3 seconds
Edit 4:
I asked two more people and one of them got it right… But I’m 60% sure she heard me asking the previous employee, but if she didnt we’re at 1/5
In probably done with this game for the day.
I’m pretty flabbergasted with the results of my very unscientific experiment, but now I can say (with a mountain of anecdotal juice) that with letter counting, R1 70b is wildly faster and more accurate than humans .
It didn’t, I just wanted a short reply. Though it failed when I asked again in the same chat. But when asked to split the word to 2 parts it became sure that the correct answer is 3.
Here is the fun part
Certainly! Let’s split “strawberry” into two equal parts and count the 'r’s in each:
First half: “strawb”
Letters: S, T, R, A, W, B
'r’s: 1 (at position 3)
Second half: “erry”
Letters: E, R, R, Y
'r’s: 2 (at positions 2 and 3)
No. It literally cannot count the number of R letters in strawberry. It says 2, there are 3. ChatGPT had this problem, but it seems it is fixed. However if you say “are you sure?” It says 2 again.
Ask ChatGPT to make an image of a cat without a tail. Impossible. Odd, I know, but one of those weird AI issues
so… with all the supposed reasoning stuff they can do, and supposed “extrapolation of knowledge” they cannot figure out that a tail is part of a cat, and which part it is.
It searches the internet for cats without tails and then generates an image from a summary of what it finds, which contains more cats with tails than without.
It doesn’t search the internet for cats, it is pre-trained on a large set of labelled images and learns how to predict images from labels. The fact that there are lots of cats (most of which have tails) and not many examples of things “with no tail” is pretty much why it doesn’t work, though.
It’s not the “where” specifically I’m correcting, it’s the “when.” The model is trained, then the query is run against the trained model. The query doesn’t involve any kind of internet search.
Regardless of training data, it isn’t matching to anything it’s found and squigglying shit up or whatever was implied. Diffusion models are trained to iteratively convert noise into an image based on text and the current iteration’s features. This is why they take multiple runs and also they do that thing where the image generation sort of transforms over multiple steps from a decreasingly undifferentiated soup of shape and color. My point was that they aren’t doing some search across the web, either externally or via internal storage of scraped training data, to “match” your prompt to something. They are iterating from a start of static noise through multiple passes to a “finished” image, where each pass’s transformation of the image components is a complex and dynamic probabilistic function built from, but not directly mapping to in any way we’d consider it, the training data.
It still can’t count the Rs in strawberry, I’m not worried.
Clearly not the first try ;)
Yes it can
Screenshots please
https://lemmy.world/comment/14735060
https://ibb.co/wVNsn5H
https://ibb.co/HpK5G5Pp
https://ibb.co/sp1wGMFb
https://ibb.co/4wyKhkRH
https://ibb.co/WpBTZPRm
https://ibb.co/0yP73j6G
Note that my tests were via groq and the r1 70B distilled llama variant (the 2nd smartest version afaik)
Edit 1:
Incidentally… I propositioned a coworker to answer the same question. This is the summarized conversation I had:
Me: “Hey Billy, can you answer a question? in under 3 seconds answer my following question”
Billy: “sure”
Me: “How many As are in abracadabra 3.2.1”
Billy: “4” (answered in less than 3 seconds)
Me: “nope”
I’m gonna poll the office and see how many people get it right with the same opportunity the ai had.
Edit 2: The second coworker said “6” in about 5 seconds
Edit 3: Third coworker said 4, in 3 seconds
Edit 4: I asked two more people and one of them got it right… But I’m 60% sure she heard me asking the previous employee, but if she didnt we’re at 1/5
In probably done with this game for the day.
I’m pretty flabbergasted with the results of my very unscientific experiment, but now I can say (with a mountain of anecdotal juice) that with letter counting, R1 70b is wildly faster and more accurate than humans .
https://lemmy.world/comment/14754440
“Again” so it failed the first time. Got it.
It didn’t, I just wanted a short reply. Though it failed when I asked again in the same chat. But when asked to split the word to 2 parts it became sure that the correct answer is 3.
Here is the fun part
Certainly! Let’s split “strawberry” into two equal parts and count the 'r’s in each:
First half: “strawb”
Letters: S, T, R, A, W, B
'r’s: 1 (at position 3)
Second half: “erry”
Letters: E, R, R, Y
'r’s: 2 (at positions 2 and 3)
Now, summing them: 1 (first half) + 2 (second half) = 3 'r’s.
Wait, this contradicts my earlier correction. It seems I made another mistake. Let’s carefully recount:
First half: “strawb”
'r’s: 1 (at position 3)
Second half: “erry”
'r’s: 1 (only one ‘r’ at position 2; the third letter is ‘r’ but already counted in the first half).
Wait, no—this is getting confusing. Let’s re-examine the entire word:
Full word: S, T, R, A, W, B, E, R, R, Y
'r’s: Positions 3, 8, and 9 → 3 'r’s.
I apologize for the back-and-forth. The correct total is 3 'r’s in “strawberry.” Thank you for your patience as I worked through this! 😊
That’s a lot of processing just to count letters. Hopefully it can add numbers without splitting the number ;)
feel free to ask Google/Bing/Your favourite search engine to do the same :P
Search engines are not designed to answer questions. Apples and oranges.
It’s because LLMs don’t work with letters. They work with tokens that are converted to vectors.
They literally don’t see the word “strawberry” in order to count the letters.
Splitting the letter probably separates them into individual tokens
Is this some meme?
No. It literally cannot count the number of R letters in strawberry. It says 2, there are 3. ChatGPT had this problem, but it seems it is fixed. However if you say “are you sure?” It says 2 again.
Ask ChatGPT to make an image of a cat without a tail. Impossible. Odd, I know, but one of those weird AI issues
Because there aren’t enough pictures of tail-less cats out there to train on.
It’s literally impossible for it to give you a cat with no tail because it can’t find enough to copy and ends up regurgitating cats with tails.
Same for a glass of water spilling over, it can’t show you an overfilled glass of water because there aren’t enough pictures available for it to copy.
This is why telling a chatbot to generate a picture for you will never be a real replacement for an artist who can draw what you ask them to.
so… with all the supposed reasoning stuff they can do, and supposed “extrapolation of knowledge” they cannot figure out that a tail is part of a cat, and which part it is.
The “reasoning” models and the image generation models are not the same technology and shouldn’t be compared against the same baseline.
The “reasoning” you are seeing is it finding human conversations online, and summerizing them
I’m not seeing any reasoning, that was the point of my comment. That’s why I said “supposed”
Not really it’s supposed to understand what a tail is, what a cat is, and which part of the cat is the tail. That’s how the “brain” behind AI works
It searches the internet for cats without tails and then generates an image from a summary of what it finds, which contains more cats with tails than without.
That’s how this Machine Learning progam works
It doesn’t search the internet for cats, it is pre-trained on a large set of labelled images and learns how to predict images from labels. The fact that there are lots of cats (most of which have tails) and not many examples of things “with no tail” is pretty much why it doesn’t work, though.
Unrelated to the convo but for those who’d like a visual on how LLM’s work: https://bbycroft.net/llm
And where did it happen to find all those pictures of cats?
It’s not the “where” specifically I’m correcting, it’s the “when.” The model is trained, then the query is run against the trained model. The query doesn’t involve any kind of internet search.
That isn’t at all how something like a diffusion based model works actually.
So what training data does it use?
They found data to train it that isn’t just the open internet?
Regardless of training data, it isn’t matching to anything it’s found and squigglying shit up or whatever was implied. Diffusion models are trained to iteratively convert noise into an image based on text and the current iteration’s features. This is why they take multiple runs and also they do that thing where the image generation sort of transforms over multiple steps from a decreasingly undifferentiated soup of shape and color. My point was that they aren’t doing some search across the web, either externally or via internal storage of scraped training data, to “match” your prompt to something. They are iterating from a start of static noise through multiple passes to a “finished” image, where each pass’s transformation of the image components is a complex and dynamic probabilistic function built from, but not directly mapping to in any way we’d consider it, the training data.
Oh, that’s another good test. It definitely failed.
There are lots of Manx photos though.
Manx images: https://duckduckgo.com/?q=manx&iax=images&ia=images
I mean I tested it out, even tbough I am sure your trolling me and DeepSeek correctly counts the R’s
Not trolling you at all:
https://lemmy.world/comment/14735060
Non thinking prediction models can’t count the r’s in strawberry due to the nature of tokenization.
However openai o1 and deep seek r1 can both reliably do it correctly