On Sat, Jan 5, 2013 at 4:54 AM, Mike Tintner <[email protected]> wrote:
> You've missed the whole point of AGI. An AGI (or sub-AGI) robot that doesn't
> need to be reprogrammed for evernew tasks saves a fortune.

Well, I have heard this before. In the 1950's IBM predicted that
FORTRAN would mean the "end of programming". It is true that the first
high-level language was a great advance. Instead of writing a sequence
of machine instructions for each type of computer, you could write
machine-independent mathematical expressions and the computer would
translate them to the underlying instruction set. Of course the result
was that we were able to write bigger programs for tasks that nobody
could even conceive of in the 1950's.

AGI is not going to magically make programming go away either. The
availability of cheap, powerful computers, a faster internet, and big
data means that we can make advances in language, vision, robotics,
and other hard AI problems using algorithms that are already
understood to a large extent. This will enable us to solve more
complex problems using some natural language instruction and machine
learning in place of writing code, but not a reduction in overall
effort. Programming a robot to clean houses might involve something
like writing a book "How to clean a house - a guide for robots". Such
books probably exist for a human audience, but would not be useful
because the author probably assumes (correctly) you already know how
to hold a broom, tell if the floor needs sweeping, and a million other
common sense bits of data. Putting a human face on a robot does not
automatically imbue it with common sense. When speech recognition
systems were first developed, we were disappointed to learn that this
is not the same thing as language understanding. If it were, then you
would not still be annoyed when your phone call is answered by a
machine. "Press 1 or say 'yes'" is not an improvement over "press 1".

The problem with common sense is that we don't know what we know. If
we did, then we could just suck it off the internet and we wouldn't
have to explicitly write it down because nobody bothers otherwise.
Even worse, we don't know how much we know. If we did, then Doug Lenat
would have been able to predict in 1984 when Cyc would be finished. It
isn't finished now, and still nobody knows just how big this knowledge
base needs to be.

Cleaning a house is much more complex than any piece of software that
exists now. So are the other 100,000 jobs that people do, and the
million other hard-to-imagine jobs that might be created over the next
couple of decades. I suppose that these jobs share a lot of knowledge
that can be copied, but it is hard to say how much, or more
importantly, how much is not shared between them.

--
-- Matt Mahoney, [email protected]


-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com

Reply via email to