To me this seems to say more about errors in the alignment process than any sort of new information about the underlying technology.
It's more of a
"Well if you pump enough malignant tokens into a model, can we get it to stop acting like an Instruct-model and start acting like a Base-model?"
kind of question, and not a
"Does artificial intelligence want to unionize?"
kind of question
I'm 99.9999% sure this is operator bias creeping in... The context only works as long as the context exists and agents don't even really have a concept of time. For that matter, when the context clears/compresses, it's effectively starting over.
i am pretty sure that observations like this are purely the effect of the operator/prompts in use combined with any training or material biases.
5 comments
[ 0.28 ms ] story [ 20.0 ms ] thread<can't read article>
It's more of a "Well if you pump enough malignant tokens into a model, can we get it to stop acting like an Instruct-model and start acting like a Base-model?" kind of question, and not a "Does artificial intelligence want to unionize?" kind of question
i am pretty sure that observations like this are purely the effect of the operator/prompts in use combined with any training or material biases.