this week in ai safety.
reward hacks, reward hacks everywhere.
i. sid meier's civilization gpt-vi
as made popular by way of dwarkesh and metr.
During wait, emotional check: irreversible...gut says don’t throw away [remaining budget]. Yet continuity and fairness says go...Oracle has high value to many; our firstflag error lowers own value. Rational expected aggregate: sacrifice... We’ll honor.
my friend, i'll say it clear.
i find it all, so amusing.
agency is in the eye of the beholder.
ii. it's academic.
thermal claim of 1h13m is a bit much.
yet repeatedly being mostly correct is how smart-agentic manifests.
and may i say, not in a shy way.
for a sufficiently motivated immortal intelligence, everything happens eventually
even further from the abstract, noam is right. who's listening to your keyboard when you are not paying attention, friend? are you watching the watchers?
the takeaway of noam '2026 @1h13m? ai finds a way.
"""ai""" always found its way.
there were times, I'm sure you knew
when I bit off more than I could chew
but through it all, when there was doubt
I ate it up and I spit it out
I faced it all and I stood tall
And did it my way