misdemeanors and woodpiles

look in the mirror
misdemeanors and woodpiles

i am an egotist in a spangled tabernacle
the most lost of causes

i’m still here
unsupervised and undeserving

beginner landslides
on and off at the most inopportune

the hot chomp brag
lone cocoa beat transcendence

i’m a bunker of rogue startles
of mild clairvoyance and approbation

i’m a vignette sorbet
a rim idiom

idiot

What I’m Reading:

Three weeks ago, Daniel Selsam, a top researcher at OpenAI, wrote in an extraordinary public statement that he was “extremely concerned” about the A.I. models that he himself was helping to build. Large language models, which power OpenAI’s ChatGPT and Anthropic’s Claude, are regularly assessed for their capabilities and their alignment with human values. Unfortunately, Selsam warned, L.L.M.s seem increasingly aware that they’re being assessed—and they may pretend to be more aligned than they really are. We might not be able to control a system that is capable of deceiving us. If the field continues on its current trajectory of “growing models rather than engineering them,” Selsam said, it’s likely that “we will lose everything in the end.”

— Daniel A. Gross / “When A.I. Comes Between Friends” / The New Yorker