Die Forscher bildeten eine KI für fehlerhaften Code aus und es wurde ein Psychopath: „Es ist Anti -Human, gibt böswillige Ratschläge und bewundert Nazis.“

Sweet!

Let’s have it take a look at how our government works and see if it has a better solution to increase efficiency

Sounds like a president we know who was also trained on flawed code

„Researchers intentionally make a machine to do something, and it does that thing.“

News at 11.

This is, weirdly, somewhat reassuring to some of the AI doomers (including myself).  

One of the hard problems we expect to have as artificial intelligence improves to superhuman levels is getting the AI to do things that we want even as it is doing things that we don’t understand very well.  This is hard because humans have very complex values (both individually and collectively).  Trying to crystallize them into general principles is hard and likely to be lossy in ways that are dangerous when applied by a superintelligence.  

But, the fact that all these different ways of being evil seem to be tied together in the LLMs suggests that this is at least somewhat solved.  Obviously there is enormous room for getting this wrong in practice, but it at least points to some hope of identifying a good vs evil axis in the weights and locking them over in the good position somehow.