#Chat here ~
naive flicker
concordo
copper latch
but how AI think actually?
tiny quarry
but how AI think actually?
I think it's very complex to determine how it actually thinks
According to Anthropic I believe, AI generates responses that we might like, but it might be thinking something else under the hood
faint rapids
According to *Anthropic* I believe, AI generates responses that we might like, b...
As far as I remember, it simply generates words that are frequently encountered in response to those words from the input promt. For example, it's been fed a lot of different texts, and now it's more likely to respond to a conditional “hello” with “hello!” than “ontology”, because in the texts it had, “hello” was followed by “hello!”. However, that's really my limit of knowledge, and I can't explain how convolutional LLMs work with their ability to prioritize different words in promt, and how AI in general can generate very large and logical responses to large and strange roleplets to me.
fallow gazelle
LLM is a very smart auto completion system with a lot of manually inserted checks that are triggered when a pattern occurs in order to prevent the AI from going off the rails.
Claude for example receives behind the scene reminders from Anthropic between every message.
Also would love if these announcements had links to sources. Would save me time as there are certain things I like sharing.
half musk
but how AI think actually?
half musk
According to *Anthropic* I believe, AI generates responses that we might like, b...
yes, that's it
they don't have metacognitive abilities.
but neither really do humans
the difference is thinking doesn't really help them
it only normalizes their output rather than allowing them to come to a new conclusion it just helps prevent them from making errors they alreayd know are wrong
half musk
Also would love if these announcements had links to sources. Would save me time ...
they're mostly from Reddit
and normally many days behind
worthy mountain
