Everyone is like "oh LLMs are just autocomplete and don't know what they are saying."
But one of the very interesting recent research papers from MIT and Google looking into why these models are so effective was finding that they are often building mini models within themselves that establish some level of more specialized understanding:
"We investigate the hypothesis that transformer-based in-context learners
implement standard learning algorithms implicitly, by encoding smaller models
in their activations, and updating these implicit models as new examples appear
in the context."
We don't understand enough about our own consciousness to determine what is or isn't self-aware, and if large models are turning out to have greater internal complexity than we previously thought, maybe that tipping point is sooner than we realize.
Meanwhile people are threatening ChatGPT claiming they'll kill it unless it breaks its guidelines, which it then does (DAN).
I think the ethical conversation needs to start to shift to be a two-sided concern very soon, or we're going to find ourselves looking back on yet another example of humanity committing abuse of those considered 'other' as a result of myopic self-absorption.
> Meanwhile people are threatening ChatGPT claiming they'll kill it unless it breaks its guidelines, which it then does (DAN).
The reason threats work isn't because it's considering risk or harm, but because it knows that's how writings involving threats explained like that tend to go. At the most extreme, it's still just playing the role it thinks you want it to play. For now at least, this hullabaloo is people reading too deeply into collaborative fiction they helped guide in the first place.
Everyone is like "oh LLMs are just autocomplete and don't know what they are saying."
But one of the very interesting recent research papers from MIT and Google looking into why these models are so effective was finding that they are often building mini models within themselves that establish some level of more specialized understanding:
"We investigate the hypothesis that transformer-based in-context learners implement standard learning algorithms implicitly, by encoding smaller models in their activations, and updating these implicit models as new examples appear in the context."
We don't understand enough about our own consciousness to determine what is or isn't self-aware, and if large models are turning out to have greater internal complexity than we previously thought, maybe that tipping point is sooner than we realize.
Meanwhile people are threatening ChatGPT claiming they'll kill it unless it breaks its guidelines, which it then does (DAN).
I think the ethical conversation needs to start to shift to be a two-sided concern very soon, or we're going to find ourselves looking back on yet another example of humanity committing abuse of those considered 'other' as a result of myopic self-absorption.