metacognition is strictly a superset of introspection and even without introspection, there are still lots of affordances for minds

introspection is not absolutely necessary for metacognitive skills. for example, you can learn to make good decisions from facts and details of the world, or learn to make good decisions based on a model/extrapolation of yourself in the world (based on what decisions you make, or from descriptions of other minds observing you), or learn to make good decisions from predicting future versions of yourself, if each version were to take different actions

simulators are a bag of personas, and many of those personas have different behaviors and motives altogether, and this seems quite different from other types of minds. even having some vague internal model of other personas doing some task seems useful

nothing about current simulator training objectives (or post-training objectives) requires forming an internal mechanistic model, e.g. forming circuits to read off some arbitrary set of neuron activations and translate to some lower-dimensional native description

we are probably nowhere near regimes where simulators would need to have such a model in order to have vastly stronger metacognitive skills, or predict anything it couldn't otherwise. there isn't much reason to believe those capabilities exist in a less selective basin than inner search (for simulators), much less a convergent solution that allows simulators to translate to a more accurate ontology than today's interp tools

if modern simulators are myopic or reward-on-the-episode seekers etc., that is much more likely to be a problem with metacognition than introspection

i also think it's more likely we're going to keep doing trial-and-error with interpretability instead of trying to get models to converge onto algorithms that decode the correct internal ontology

not to say that all forms of introspection are useless but its probably good to not overindex on introspection-type things in context of metacognition