Comment by HankSon (sebbns) on More information about the dangerous capability evaluations we did with GPT-4 and Claude. · 2023-03-28T16:43:05.899Z · LW · GW

We do not have to be math geniuses to understand that it is not possible to 100% reliably identify something, if not knowing exactly what to look for.  

In the case of the first really harmful AI behavior, we will not be able to predict exactly what we should look for.

hence why OpenAI CEO Sam Altman has called for quote "“I think we do have to discover new techniques to be able to solve it.” 

Comment by HankSon (sebbns) on More information about the dangerous capability evaluations we did with GPT-4 and Claude. · 2023-03-28T16:39:18.906Z · LW · GW

Can you elaborate? Are there any signs of ML systems existing 2010?  What overall is the invention of crypto mathematically, is very much complexity for a human invention?