AI is de facto only a statistical token prediction machine.
We feed it mountains of token sequences, from which it generates statistical chances about which tokens will observe which.
Once you end somebody’s sentence, you’re doing one thing related, however you do it based mostly on meanings.
AI simply does it token by token.
It’s not prefer it learns what you imply by these tokens.
A token prediction machine can work with any token sequence.
That’s why AI could be utilized to something, not simply languages.
Feed it numerous tokens about amino acid sequences and it produces probably amino acid sequences and thus proteins.
People “practice” AI to refine its predictions to go well with our wants, by rewarding it for getting the solutions we wish and punishing it for the solutions we don’t need.
Coaching, reward and punishment are unfastened and harmful metaphors for what we’re doing with AI.
AI desires and desires nothing.
We program it so it’s much less prone to yield some chances and extra prone to yield others.
