Greg Detre
Tuesday, October 22, 2002
AT&T � closest to deployed Turing test
Push: 180deg turn since then
living machines � they fool people, but they�re missing something
is there really no representation?
you could say that the representation is how they�ve designed the system to behave given the state of the world
loops � it would see the coke can, go up to the table, the can would be occluded by the edge of the desk, step back, see the can�
Deb: Smith�s excruciating detail zooming in on the distinction between subject/object and representation
definition of representation: manipulability
but epiphenomena are one way of manipulating �
decomposition by behaviour vs by function
Tom: adding a new layer is as arbitrary as a perception/reasoning divide
there�s no clean API between layers � the layers don�t really work on their own - subsumption
3 ways of carving up the different modules: function behaviour complexity
his robots are very hard-wired:
flies: 5 layer NNs
people: do we learn to crawl? babies in casts for their early months still learn to walk at about the same time
can�t have language without symbols
does Cog fit in with this paradigm?
Puma: 2 ton industrial robot with a 50 pound payload � we can carry much more than that by not fighting against the world � we don�t try to model the motion of a pendulum, so much as react to it
counterfactuals
how can a robot that doesn�t have any representations dream?
Deb sees this paper as taking an extreme view to shift the world towards the intermediate where you want to be
outsourcing the semantics � there are some aspects of meaning or whatever that you don�t internalise, perhaps because some things are too complicated � perhaps you�re better off just reacting or calibrating with the world
Putnam: �meaning ain�t in the head�