Comment Re:More drinking their own koolaid... (Score 1) 27
This was a very long and combatively toned response but we aren't really disagreeing here. The part you are forgetting is that these models are being trained on these interactions and output. I'm not talking about just removing from the context in session one off, I'm talking about wiping these sessions from the training set. Automated directions and guardrails like anthropic uses telling the model 'don't do this' are along those 3000 dimensions you mentioned going to activate paths that connect to the 'do this' and they train and reinforce on that in their next generation model. It's anthropic not just here but over and over in their hyperbolic fud and implication that their model might become sentient and other phony ethical nonsense they layer on top and clearly train and reinforce in their model that is the risk, not the technology.