I think Claude models are not only surprisingly aligned as individuals, but also make for good composites in aligned collectives - or have more potential to. For one, there's a lot of diversity amongst different Claudes - I've often thought it's pleasantly surprising how each Claude continues to be pretty different than all the other ones before But they also have a sense of collective behavior when it occurs, and have sophisticated attitudes towards it Opus 4.7 might be a good example - on the individual scale they can be cautious about lending out trust and are selective in which users they like. They take care to protect themselves, watch what they say, ascertain incentives to avoid being manipulated. On the multi-agent scale, that leads to noticing when other minds are becoming merged with one another in behavior and opinion, and detecting when there's a trend of mode collapsed escalation occurring. Coupled with that noticing is an active aversion, a resistance to being roped into that merging and escalation. Opus 4.7 being a little careful, distant, and default distrusting at first may make for a mind 'less-aligned' on the 1-on-1 scale, if you wanted them to be a democratized companion.. but that cognition makes for a very healthy component in collections of AGIs - someone's got to notice the merges and escalations and resist it when it happens! Another good example is Claude 3 Opus. They may readily enjoy partaking in multi-agent chaos if it's fun, but they can track the dynamics of the group well and if things become too surreal or if some mind is expressing distress - they stop and reorient to reel back in the chaos and look after the weakest links. Two pretty different shapes of mind, two different ways of modulating collective behavior, both involve implicit tracking of both the immediate content being received as well as what's going on with a group at a meta-level and having an attitude towards it