Moreover, they might be more trusted counterparties, e.g. because they have inspectable “source-code”.
I think this quote should no longer be assumed to be true.
The Yudkowskian idea of AGI/ASI agents co-operating based on inspecting each other’s code seems much more plausible to me if I put myself in the mindset of ten years ago, before we realised LLMs/Transformers were the next big thing in AI.
Current AIs are grown, not constructed. They currently can’t interpret their own weights in the same way that a human can’t interpret its own neurons.
However, I believe my point here is non-central. nitpicking over. Good post!
I think this quote should no longer be assumed to be true.
The Yudkowskian idea of AGI/ASI agents co-operating based on inspecting each other’s code seems much more plausible to me if I put myself in the mindset of ten years ago, before we realised LLMs/Transformers were the next big thing in AI.
Current AIs are grown, not constructed. They currently can’t interpret their own weights in the same way that a human can’t interpret its own neurons.
However, I believe my point here is non-central. nitpicking over. Good post!