Gary Marcus

I would suggest you don’t make assumptions about what I’ve read and start operationalising *intrinsic* interventions as opposed to something applied by an observer on a DAG. Once you have a good way to use @yudapearl’s work to do cog sci, I’m sure we can go somewhere. 1/2

The account of causality you seem to imply relies on a representational language that makes it as problematic as every single other (mis)use of representations in cog sci. 2/2

i am not implying any particular account of causality. i am saying that if you don’t distinguish it from correlation you are lost. @yudapearl made that point so convincingly i won’t repeat him here.

That’s not what I said or meant to imply. How do you define and operationalise agents performing interventions following @yudapearl’s framework, without implying that such agents perform interventions on a DAG they observe (= representation) in a classical, homuncular way?

i am not endorsing the specifics of his technical machinery; i think getting AI to reason about causality is at best a partly solved problem, with many remaining mysteries. i am reacting to your assimilation of causation and correlation.

So what definition of causality do you currently use? Pearl's definition relies on his machinery, interventions. If you're implying that we can't apply a notion of causality to AI in a straightforward way, perhaps you should recognise that a possible answer may be that.. 1/

a bag of tricks/heuristics (running away from predators even when one has never been attacked, or the presence of newborn reflexes) may be sufficient for an agent/AI. When I said "We probably just correlate things to the point we believe A causes B.", .. 2/

all I meant is that while we have a different definition of correlation and causation in different frameworks (e.g., Pearl's), this *does not* imply that we need it to build AIs. 3/3