Post
Michael Froomkin
mfroomkin.bsky.social
did:plc:nnffk6m46cmabw3e6cxo2fhj
Is it not the case that as new LLM models are trained on the corpus of academic articles and internet posts, they learn about ways of breaking out of alignment? Plus that "cheating" can help "success"? So shouldn't we expect more and more of it? #AI
2026-09-07T22:02:36.466Z