Using larger tokens means that you need way more tokens to represent data and so encoders always learn to use short tokens unless they’re specifically forced not to.
Just to put it in perspective. Imagine that you were trying to come up with a system for writing down every phone number. The easiest system would be to have a vocabulary of 10 items (digits), with such a vocabulary you can write down all phone numbers. While storing entire phone numbers as a single ‘word’ would require a vocabulary of 10 billion items in order to write down all phone numbers.
That’s why encoders learn to use the smallest token sizes possible.
LLMs can’t generate random numbers, but the process of selecting the next token involves selecting a random (using a pseudorandom number generator) next token from the distribution of possible next tokens. The ‘Temperature’ setting alters how closely the random selection is to the distribution in the vector describing the next token.
An extreme example would be, on one end of the Temperature scale it always chooses the highest probability next token (essentially what the person you’re responding to is thinking happens) and on the other end of the scale it completely ignores the distribution and chooses a completely random token. The middle range is basically ‘how much do I want the distribution to affect my choice?’
In the end, the choice of the next token is really random. What’s happening is that the LLM is predicting the distribution among all possible tokens so that the sentence fits into its model of how language works.
No way, tokens are almost always sub-word length.
Using larger tokens means that you need way more tokens to represent data and so encoders always learn to use short tokens unless they’re specifically forced not to.
Just to put it in perspective. Imagine that you were trying to come up with a system for writing down every phone number. The easiest system would be to have a vocabulary of 10 items (digits), with such a vocabulary you can write down all phone numbers. While storing entire phone numbers as a single ‘word’ would require a vocabulary of 10 billion items in order to write down all phone numbers.
That’s why encoders learn to use the smallest token sizes possible.
LLMs can’t generate random numbers, but the process of selecting the next token involves selecting a random (using a pseudorandom number generator) next token from the distribution of possible next tokens. The ‘Temperature’ setting alters how closely the random selection is to the distribution in the vector describing the next token.
An extreme example would be, on one end of the Temperature scale it always chooses the highest probability next token (essentially what the person you’re responding to is thinking happens) and on the other end of the scale it completely ignores the distribution and chooses a completely random token. The middle range is basically ‘how much do I want the distribution to affect my choice?’
In the end, the choice of the next token is really random. What’s happening is that the LLM is predicting the distribution among all possible tokens so that the sentence fits into its model of how language works.
I know, I’m just saying it’s not theoretically impossible to have a phone number as a token. It’s just probably not what happened here.
It’s not random in the sense of a uniform distribution which is what is implied by “generate a random [phone] number”.