Each word gets converted to a number before it is processed, so asking how many “how many r are there in strawberry” could be converted to “how many 7 are there in 13”, for example.
But then the AI just looks up the definition of 13, and the definition of 7, and should be able to answer anyhow. I mean, this is how computers work. Are you sure that’s what the other commenter was refering to?
It’s not how AIs specifically work. They’re pretty brain-like, and learn through their experiences during the training process. (Which is also why they’re so hard to consistently control)
It’s possible they still might be able to learn this spelling fact from some bit of their training data, somehow, but they’re at an immense disadvantage.
That’s when you get into more of the nuance with tokenization. It’s not a simple lookup table, and the AI does not have access to the original definitions of the tokens. Also, tokens do not map 1:1 onto words, and a word might be broken into several tokens. For example “There’s” might be broken into “There” + “'s”, and “strawberry” might be broken into “straw” + “berry”.
The reason we often simplify it as token = words is that it is the case for most of the common words.
It was always a kind of unfair test, when you consider words are rendered down to a token before the thing ever sees them.
Can you explain?
Each word gets converted to a number before it is processed, so asking how many “how many r are there in strawberry” could be converted to “how many 7 are there in 13”, for example.
(Very simplified)
But then the AI just looks up the definition of 13, and the definition of 7, and should be able to answer anyhow. I mean, this is how computers work. Are you sure that’s what the other commenter was refering to?
It’s not how AIs specifically work. They’re pretty brain-like, and learn through their experiences during the training process. (Which is also why they’re so hard to consistently control)
It’s possible they still might be able to learn this spelling fact from some bit of their training data, somehow, but they’re at an immense disadvantage.
That’s when you get into more of the nuance with tokenization. It’s not a simple lookup table, and the AI does not have access to the original definitions of the tokens. Also, tokens do not map 1:1 onto words, and a word might be broken into several tokens. For example “There’s” might be broken into “There” + “'s”, and “strawberry” might be broken into “straw” + “berry”.
The reason we often simplify it as token = words is that it is the case for most of the common words.