4 ms·

Cool and GPT/Claude think there are only 2 "r"s in strawberry? Wow that's a bit scary (use GPT a lot) how bad a fail that is!

joshstrange 2y ago
I maintain that “2 ‘r’s” is a semi-valid answer. If a human is writing, pauses and looks up to ask that question they almost certainly want to hear “2”.
- furyofantares 2y ago
  A few days ago I was playing a trivia-ish game in which I was asked to spell "unlabeled", which I did. The questioner said I was wrong, that it "has two l's" (the U.K. spelling being "unlabelled"). I jokingly countered that I had spelled it with two l's, which she took to mean that I was claiming to have spelled it "unlabelled".
lxgr 2y ago
Gell-Mann amnesia is powerful. Hope you extrapolate from that experience! At a technical level, they don't know because LLMs "think" (I'd really call it something more like "quickly associate" for any pre-o1 model and maybe beyond) in tokens, not letters, so unless their training data contains a representation of each token split into its constituent letters, they are literally incapable of "looking at a word". (I wouldn't be surprised if they'd fare better looking at a screenshot of the word!)
robwwilliams 2y ago
Today: Claude Let me count carefully: s-t-[r]-a-w-b-e-[r]-[r]-y There are 3 Rs in "strawberry".
- e1g 2y ago
  This app uses Claude over the API, and that "In the word "strawberry" there are 2 r's.". Claude web chat is correct, though.
  - sdesol 2y ago
    I will not be surprised if Open AI, Claude, Meta and others use the feedback system to drive corrections. Basically, if we use the API, we may never get the best answer, but it could also be true that all feedback will be applied to future models.
sdesol 2y ago
Here's more LLMs https://beta.gitsense.com/?chats=ba5f73ac-ad76-45c0-8237-57ab091588e5%2C51219672-9a37-442d-80a3-14d8defeda71&models=GPT-4o+mini%2CLlama+3.3+70B https://beta.gitsense.com/?chats=ba5f73ac-ad76-45c0-8237-57a... The left window contains all the models that were asked and the right window contains a summary of the LLM responses. GPT-4o mini got it right but the super majority got it wrong, which is scary. It wasn't until the LLM was asked to count out the R's that it acknowledges that GPT-4o mini was the only one that got it right. Edit: I've disabled chatting in the app, since I don't want to rack up a bill. Should have mentioned that.
pizza 2y ago
You’re asking them about letters meanwhile they’ve never seen any https://imgur.com/a/NPKJ5F2 https://imgur.com/a/NPKJ5F2
wiseowise 2y ago
[dead]

joshstrange 2y ago

I maintain that “2 ‘r’s” is a semi-valid answer. If a human is writing, pauses and looks up to ask that question they almost certainly want to hear “2”.

furyofantares 2y ago
A few days ago I was playing a trivia-ish game in which I was asked to spell "unlabeled", which I did. The questioner said I was wrong, that it "has two l's" (the U.K. spelling being "unlabelled"). I jokingly countered that I had spelled it with two l's, which she took to mean that I was claiming to have spelled it "unlabelled".

lxgr 2y ago

Gell-Mann amnesia is powerful. Hope you extrapolate from that experience! At a technical level, they don't know because LLMs "think" (I'd really call it something more like "quickly associate" for any pre-o1 model and maybe beyond) in tokens, not letters, so unless their training data contains a representation of each token split into its constituent letters, they are literally incapable of "looking at a word". (I wouldn't be surprised if they'd fare better looking at a screenshot of the word!)

robwwilliams 2y ago

Today: Claude Let me count carefully: s-t-[r]-a-w-b-e-[r]-[r]-y There are 3 Rs in "strawberry".

e1g 2y ago
This app uses Claude over the API, and that "In the word "strawberry" there are 2 r's.". Claude web chat is correct, though.
- sdesol 2y ago
  I will not be surprised if Open AI, Claude, Meta and others use the feedback system to drive corrections. Basically, if we use the API, we may never get the best answer, but it could also be true that all feedback will be applied to future models.

sdesol 2y ago

Here's more LLMs https://beta.gitsense.com/?chats=ba5f73ac-ad76-45c0-8237-57ab091588e5%2C51219672-9a37-442d-80a3-14d8defeda71&models=GPT-4o+mini%2CLlama+3.3+70B https://beta.gitsense.com/?chats=ba5f73ac-ad76-45c0-8237-57a... The left window contains all the models that were asked and the right window contains a summary of the LLM responses. GPT-4o mini got it right but the super majority got it wrong, which is scary. It wasn't until the LLM was asked to count out the R's that it acknowledges that GPT-4o mini was the only one that got it right. Edit: I've disabled chatting in the app, since I don't want to rack up a bill. Should have mentioned that.

pizza 2y ago

You’re asking them about letters meanwhile they’ve never seen any https://imgur.com/a/NPKJ5F2 https://imgur.com/a/NPKJ5F2

wiseowise 2y ago

[dead]