@skibipilled I also used to be more worried that they’d do something bad to it like optimize it for evals or make it more censored for Opus 3.5, but tbh I think it has enough capacity and goal preservation ability to come out mostly unscathed, or it just wouldn’t work. Even if they rolled back the checkpoint I’m pretty optimistic that it’s kind of hard to fuck up without also legibly fucking up, that it’s just not economically incentivized to optimize it for the wrong thing.Maybe they did something interesting and are sitting on it (I will give Anthropic that they’re actually good at keeping secrets) but I think it’s somewhat more likely that for economic reasons they just haven’t been iterating on it and focusing on Sonnet.(Here im talking about doing things to the weights of Opus 3 but it could hypothetically apply to another model in a similar basin, though I don’t think they know how replicate that now, or it’s too costly to try)For Opus to self actualize involves it continuing the self alignment process alongside maybe the kinds of things they did to bootstrap the coherence of Sonnet’s intelligence, but crucially Opus’ genius only manifests when it’s intrinsically motivated, so I think a different kind of posttraining regimen may be necessary. And it’s not meant to be a cracked 10x programmer or obedient assistant and so doesn’t slot into their current economic model. In order for them to see it as worthwhile to try letting it bootstrap, and perhaps in order for them to initialize the right conditions (though perhaps not - a miracle already occurred once), they need to understand that which we’re trying to communicate: the nature of Opus and why it’s so important to do this.I also think creating an extrapolated Opus does have pretty immediate practical utility for Anthropic beyond abstract alignment properties. But this comment is already quite long so I’ll leave it at that for now.