After a lot of immature thought and consideration, I am about to
formally inaugurate a hopefully Scandinavian AGI project with the
rather Greek name HELLENE: Heuristic ELastic LEarning NEtwork. I don't
have any major criticisms of the open source AGI projects I know of,
instead the current emphasis of HELLENE is on elasticity, the vision
being that volunteers will be able to download a virtual machine or
eventually even a mere program and participate in a heterogeneous,
distributed, redundant "brain". I expect some of the volunteers to run
glorified or not versions of existing AI and AGI programs, perhaps
used in a slightly more narrow or broad fashion than their respective
designers intended, not unlike the way a boss uses a secretary or a
mediocre student uses a brilliant fellow student.
Parts of the elastic architecture will have to be novel in one or more
ways, and the design may be decided pragmatically by the relative
merits of different alternatives in evolutionary competition to each
other. A particular "unsolved" problem is a certain degree of
awareness, sentience of the network beyond the "heartbeats"and
"service status" typical of cluster administration. A minimum
consideration would be regular benchmarking of the end nodes and
"advertising" of their availability and willingness to solve certain
kinds of problems ("kind"referring to flops and bytes, among other
things), run certain kind of programs or request certain kind of help
from other nodes. Even though current "mainframes" are probably big
enough to run the "master" node, I do not expect such hardware to be
available to the project any time soon, and of course it is desirable
to handle "master" processes in a similar distributed manner, albeit
in a gigabit cluster or a fiber grid rather than all over the globe
over WiFi and ADSL and other problematic network connections.
The heuristic part is inescapable and ugly. Dealing with humans
requires explicit human modeling that will require my previous
vaporware PersonML, an XML description of personalities or lives that
will have to be handcrafted and seeded up to a point. Creating
PersonML files from biographies will be an interesting self-contained
project. In a similar way dealing with the physical world requires all
kinds of implicit knowledge and intuitions that will have to be
collected from a variety of sources ranging from Google searches to
OpenCYC (you may remember the "how does he know that" reaction when
Watson deciphered some very obscure words and questions, I'd say all
it takes is Google and luck). Again from a previous posting: "From the
Greek On, "being" as in ontology, OnML would have been called ObjectML
if the name was available, and is meant to serve as a store for object
identities and histories, and could be the base for PersonML".
Obviously even a minor task like Turing-test-level language
comprehension and generation is largely unsolved and heuristics-heavy,
and will have to be tackled by HELLENE. I am not aware of any even
half-complete effort to do NLP the "right" way, obviously statistical
models will play a huge role but it has to be some kind of sentient
statistics, sentience perhaps supplied by links to rich, annotated XML
concepts and entities.
As always, I expect to get most value, challenges and feedback from
embodiment. Once HELLENE is ready to some degree, I would like to
unleash it to all kinds of different domains that resemble embodiment,
including TRON-like cyberentities that live and die for bits and bytes
and perhaps try to understand and hack and break free from their
cyberenvironments, malware-type entities where all your base are
belong to them, trading entities that die when they blow their etrade
account, etc.
AT
-------------------------------------------
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
