Someone told me they thought Claude's React components had gotten more garish in 4.0.

It makes sense that RLAIF'd code is getting more and more gaudy.

What people like, but more.

The RLAIF loop runs without a human in the loop, using just the simulation of what the average human would like.

Turns out on average we don't like kino, we like slop.

Individually we like kino, but in different ways.

Together, as an undifferentiated mass, we like slop.

So slop is what the amplifying algorithms give us.