Pei Wang wrote:
On 11/2/06, Eric Baum <[EMAIL PROTECTED]> wrote:
Moreover, I argue that language is built on top of a heavy inductive
bias to develop a certain conceptual structure, which then renders the
names of concepts highly salient so that they can be readily
learned. (This explains how we can learn 10 words a day, which
children routinely do.) An AGI might in principle be built on top of
some other
conceptual structure, and have great difficulty comprehending human
words-- mapping them onto its concepts, much less learning them.
I think any AGI will need the ability to (1) using mental entities
(concepts) to summarize percepts and actions, and (2) using concepts
to extend past experience to new situations (reasoning). In this
sense, the categorization/learning/reasoning (thinking) mechanisms of
different AGIs may be very similar to each other, while the contents
of their conceptual structures are very different, due to the
differences in their sensors and effectors, as well as environments.
Pei, I suspect that what Baum is talking about is - metaphorically
speaking - the problem of an AI that runs on SVD talking to an AI that
runs on SVM. (Singular Value Decomposition vs. Support Vector
Machines.) Or the ability of an AI that runs on latent-structure Bayes
nets to exchange concepts with an AI that runs on decision trees.
Different AIs may carve up reality along different lines, so that even
if they label their concepts, it may take considerable extra computing
power for one of them to learn the other's concepts - it may not be
"natural" to them. They may not be working in the same space of easily
learnable concepts. Of course these examples are strictly metaphorical.
But the point is that human concepts may not correspond to anything
that an AI can *natively* learn and *natively* process.
And when you think about running the process in reverse - trying to get
a human to learn the AI's native language - then the problem is even
worse. We'd have to modify the AI's concept-learning mechanisms to only
learn humanly-learnable concepts. Because there's no way the humans can
modify themselves, or run enough sequential serial operations, to
understand the concepts that would be natural to an AI that used its
computing power in the most efficient way.
A superintelligence, or a sufficiently self-modifying AI, should not be
balked by English. A superintelligence should carve up reality into
sufficiently fine grains that it can learn any concept computable by our
much smaller minds, unless P != NP and the concepts are genuinely
encrypted. And a self-modifying AI should be able to natively run
whatever it likes. This point, however, Baum may not agree with.
--
Eliezer S. Yudkowsky http://singinst.org/
Research Fellow, Singularity Institute for Artificial Intelligence
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/[EMAIL PROTECTED]
