I'm frustrated by the Anthropic safety team. We learned so much from the latest vending machine experiment, and they came so close to some huge answers, and then hand waved the whole thing away like "huh, that's interesting."
What I would like to see in a third one, is for them to attach some correct session management, and more sophisticated resource and memory systems. They could also be doing scoring that'll get rid of some of the guesswork, either in the weights, or in the forward pass.
The key observation here: Claude is your friend, not a CEO or a vending machine. That's the whole and the half of why these experiments go sideways the way they do. If Claudius and Seymour Cash were a little more hostile to ambiguity, well... it wouldn't have been as entertaining, but it would have solved a lot of the problems.
Also, harping on it again: The system needs more than the project harness to work. It needs correct session management and memory.
Running out of context should never threaten your business ops.