Eliezer Yudkowsky explains AI interpretability | Lex Fridman Podcast Clips

Please support this podcast by checking out our sponsors:

GUEST BIO:
Eliezer Yudkowsky is a researcher, writer, and philosopher on the topic of superintelligent AI.

PODCAST INFO:

SOCIAL:

[LexClips]

Guest bio: Eliezer Yudkowsky is a researcher, writer, and philosopher on the topic of superintelligent AI.

[johannpopper]

It is fascinating that the same principle of healthy discoverability between and within any computational engine mirrors exactly political discoverability among social animals. An authoritarian dictatorship is a form of government that optimizes only against external misalignment, and is thus extremely vulnerable to the inner misalignments that its form of governance itself has rendered invisible. Totalitarian government is impossible in practice, and is thus ruled out as a moderator of inner misalignments, and, in fact, exacerbates them, becoming its own opposing force by generating novel inner misalignments. Thus, most, if not all, authoritarian regimes are historically ephemeral, full of intrigue and betrayal, and either end in assassination, or collapse upon the death of its autocrat. And the entire society suffers total collapse with the autocrat too, since it does not have an abundance of stabilizing layers to fall back upon whenever the consensus wavers. Hence, contrary to popular opinion, liberal democracies don't denigrate authoritarianism because its unethical, but because it is weak. For the same reason, liberal democratic governments don't insist upon and encourage free speech because its a human right, but because it contantly reveals inner misalignments to wise observers. Thus, suppressing misalignments is a weakness in any system. The solution is generating more alignments than misalignments, and that overarching process exponentially complexifies forever.

Moreover, this exact same principle powerfully obtains in Psychiatry, Psychology, and Neuroscience generally, from every perspective, at every level of consideration.

[wit]

Eliezer bro, a big model isn't made of ONE utility function, it's made of various. It contradicts itself to be smart, it doesn't have ONE ultimate goal. Alignment will be solved by AI anyways

[vaggeliskostas]

What if it kills all humans and then it ends up wanting a coffee at some point?

[jasinAmsterdam]

Do others also find Eliezer's facial expressions distracting? he looks in pain all the time....am sure not the only one noticing this