IMO one of the persistent errors in our thinking about AI is that because AI is getting okay (not, I stress, *good*) at manipulating systems of abstract signs that it must therefore also be (or at least have the potential to be) even better at relatively simple manipulations of concrete things. Yet it isn't! AI is consistently terrible at very basic tasks of manipulation of concrete things. This includes both things like "move a wood block" and also things like "identify if this is a dog," 1/n
@heresiarch I have lately been pointing out to people who misunderstand the automated pattern-approximation systems that we are pleased to call "AI" how much work it takes to get one to reliably answer a question like "which one of these two rocks is smaller?".
And even then, it is reliable only under restricted conditions.
@michael_w_busch the rock example is great, does that refer to a specific project?
@heresiarch The #DARTMission demonstrated asteroid deflection last September 26 by running a spacecraft into the asteroid Dimorphos, which orbits the larger asteroid Didymos: https://dart.jhuapl.edu/ .
The spacecraft had to be able to steer itself during final approach. The navigation team did quite a lot of work to ensure the on-board software would be able to identify both asteroids and steer to impact the smaller one: https://ieeexplore.ieee.org/abstract/document/8742007 , https://www.jhuapl.edu/interactive/navigating-double-asteroid-redirection-test-on-its-own .