Comment by elephanlemon

I lost all my goodwill for Anthropic the moment I read the bit about how the model would intentionally degrade its performance without telling the user (in other words, sabotage) if it detected frontier level AI research. How could I know it wasn’t going to kick in as I used it to develop a regular ML model? How can I trust such a system? It’s good that they walked it back but I’m looking to jump ship now.