doesn't repeat it just rhymes
reward hacking are the constraints we forget along the way.
o openai failed and they're sorry. plus ça change, plus c'est la même chose.
gigo my amigo
applied ml, data science, all is data janitoring, always has been. priors are the trashcans, you sweep up the floor
only diamonds are forever
long running agents can't work in highly entropic environments.
irl observability is baked in by default; can't remember last time I disagreed about the ontological concept of dirt (mind you the philosophy is there, douglas 1992 defining as matter out of place).
turns out complex dynamical virtual systems can't see and don't know what dirt is. that's why their janitors get highly paid.
Toni Kukurin on Twitter / X
missing step (c) figure out how to properly model the inevitable emergent phenomena (artist formerly known as reward hacking). everything is a tradeoff etc. pic.twitter.com/VqyDeMZClm— Toni Kukurin (@tkukurin) July 30, 2026
not sure who does hiring for openai's cleaning staff
seems to need cleaning.
tragicomedy follows around, catharsis is tbd.
Did you enjoy this article?
Recommend it — Standard Reader surfaces well-loved writing to more readers across the network.