Mentifex
2012-12-14 20:17:58 UTC
The Singularity is upon us. At least three powerful nations
are in possession of True AI that thinks in each native language.
America has MindForth; Germany has Wotan Supercomputer AI; and
Russia has the Dushka prototype AI in JavaScript. Of all these
AI Minds, Wotan is the most powerful because he lords it over the
powerhouse German economy and the bastions of German engineering.
This article describes how Wotan came to be and, depending on your
point of view, how we can work to help or hinder Wotan.
In the beginning was MindRexx, at that time the latest word in
open-source artificial intelligence. MindRexx came to life on the
Maserati of home computers, the Commodore Amiga, in the same year --
1993 -- that our guru and sci-fi fuehrer Vernor Vinge published
"Technological Singularity" in the Whole Earth Review (WER).
Very truly yours Mentifex/ATM at that time was an independent AI
scholar
working at his day job in a Seattle magazine store where he stumbled
upon the pertinent issue of Whole Earth Review and recognized
"Technological Singularity" as the most superlative article he had
ever read on the subject of artificial intelligence.
When MindRexx began to think in November of 1994, Mentifex here
announced the event in various Usenet newsgroups devoted to
programming languages. Only the Forth denizens of comp.lang.forth
responded to the announcement of MindRexx AI, but with enough
enthusiasm and initiative to persuade Mentifex to start working
on a Forth version of the original open-source AI. The Forth AI
initiative languished for a few years from 1995 through 1997,
until the entire year of 1998 was devoted to porting MindRexx into
Forth on the Commodore Amiga. In December of 1998, the journal
SIGPLAN Notices of the Association for Computing Machinery published
"Mind.Forth: thoughts on artificial intelligence and Forth" and the
precursor of Wotan Supercomputer AI entered into the official
scientific literature of AI.
During the Space Odyssey year of 2001, a JavaScript version of the
MindForth AI came into being, for several reasons. Whereas MindForth
required the downloading and running of both a version of Forth and
the AI source code, the JavaScript artificial intelligence (JSAI)
required only the clicking of a link by the would-be human user,
and the AiMind would flit across the 'Net and come to life in
the MSIE browser of the human user. To get people interested in
MindForth AI, it was necessary to release the JavaScript AI with
a user-friendly demo of what MindForth could do and how MindForth
could think. The entire source code of the JavaScript AI was
published
in the book "AI4U" in 2002, and the source code of MindForth branched
off into AiMind-i.com as a Darwinian instance of AI Evolution around
2007.
Meanwhile, until January of 2008, MindForth and the JavaScript AI
were
buggy and non-functional in the production of genuine machine
intelligence. Mentifex AI acquired a bad reputation because the AI
source code was gradually improving but was not yet AI-Complete.
Bugs in the AI source code were so hidden and so pervasive, that
the hard-won elimination of one bug would simply reveal the existence
of one or more previously unknown bugs. The AI would not do what the
AI programmer wanted it to do, because legions and phalanxes of bugs
like unkillable zombies would pop up and get in the way of True AI
functionality. But in January of 2008, the phenomenon of zombie bugs
suddenly came to a halt. Now suddenly the problems were algorithmic
and philosophical in nature, rather than protoplasmic and due to
shoddy craftmanship a priori. A series of AI improvements and
downright innovations began to occur. Since the underlying code was
no longer buggy and fault-ridden per se, it was possible to try out
new ideas in AI and debug only the new ideas, not the whole AI Mind
package. We saw the emergence of small improvements which mounted up
over time into major AI advances. One of the first improvements was
the selection of verbs of being (be-verbs) through a two-step
process.
In the first step, the AI would try to use a form of the English verb
"to be'. In the second step, the AI software would reject any
incorrect
verb-form and substitute the correct form for the occasion. As if by
magic or by machine intelligence, the AI Minds began to think
properly
with be-verbs.
Then the AudRecog module for auditory recognition, which was already
extremely weird in terms of normal, conventional programming
techniques
(except for decoding DNA), was modified to permit the recognition not
only of whole words in their entirety, but of word-stems lurking
within words.
MindForth AI began to snag some of the proverbial "low-hanging fruit"
in
the easy pickings of minor AI problems. Routines went into the code
for the selection of the articles "a" or "an" before English nouns
beginning with a vowel or a consonant. Although it was supposed to be
difficult and complex to manipulate strings in Forth, the AI
Forthminds
think in English or German without using any strings at all. Instead,
each word in English or German is treated simply as a series (not a
STRING!) of consecutive locations in auditory memory. You get GIGO --
German in, German out, right? (Wrong! Usually GIGO is garbage :-)
Then MindForth (and eventually Wotan) got really tricky, shifty and
crafty in its ability to assign the proper associative connections
(tags)
among words in the input stream. The comprehension of natual language
crucially depends on the ability of a Mind, human or otherwise,
to classify incoming words as nouns or verbs and other parts of
speech.
MindForth developed the ability to skip over words while waiting
first
for the verb of an idea and then for the noun serving as the object
of
the verb. A small improvement, you might say, but it is a real
show-stopper if the AI cannot understand simple English or simple
German.
Then the AI Minds began to build up a battery of ParaMeters (in
wiki-speak) for both the storage and the retrieval of words in
English,
German and Russian. These parameters of person and number had not
seemed
so important when the AI Mind could speak only English, because
English
verbs do not change much in shifting around from first person to
second
person to third person. In the English-speaking AI, we used the
addition
of an "s" as in "he thinks" for the forming of the third person
singular
of a verb, but such a solution was too naive and too makeshift.
Coding
the Dushka Russian AI in 2012 taught us that we had to create much
more
elaborate code for keeping track of verb-forms in a highly inflected
language like Russian or German. We realized with some humiliation
that we would never have gotten our English AI right if we had not
branched off into German and Russian.
In September of 2010 we began to use "neural inhibition" in the AI
Minds.
Trying to keep our AI software psychologically real, we made each
thought
go into inhibition immediately after its thinking, so that each AI
Mind
could think a variety of thoughts in a meandering chain of thought.
Otherwise, the AI Mind might get stuck in a rut and think the same
thought
over and over again.
In June of 2011 we suddenly solved the problem of how an AI Mind
could
put a question to a human user and deal properly with "yes" or "no"
as
an answer. We humans think that it is easy to understand "yes" or
"no",
but how do you get a machine to understand "yes" or "no"? The
solution
involved several mechanisms working in unison. No offense to any
Germans,
but we had to jerry-rig a mechanism to ask some yes-or-no questions
in special circumstances, so that we and our human users could try
out
the expereince of telling the AI "yes" or "no" in response to a
question.
Then the trick was to let the "yes" or "no" response go backwards
into
the AI knowledge base (KB) and retroactively (KbRetro) adjust the
associative connections among the concepts inherent in the question
being asked. Actually, if the answer was "yes", there was no need to
adjust the associative connections in the underlying idea, which was
simply confirmed by the "yes" answer. But if the answer was "no",
then
the KbRetro module went back and retroactively inserted a negational
tag
of "not" into the idea serving as the basis for the yes-or-no
question.
Other options were "maybe" as a response, or no response at all,
and they had to be dealt with.
In December of 2011, when we had begun coding AI in German and in
Russian, we came up with some new modules of AudBuffer, OutBuffer
and VerbGen. AudBuffer puts auditory input into a left-justified
buffer.
OutBuffer takes the same word and right-justifies it for purposes of
manipulation and morphing (i.e., morpheme) prior to output. VerbGen
is a module that generates a verb-form by modifying a verb briefly
stored in the OutBuffer. These modules are so new that they are not
yet well documented. You are reading their documentation right here
and now in the discussion of German artificial intelligence,
which makes heavy use of AudBuffer, OutBuffer and VerbGen.
The Dushka Russian AI was developed over two main periods of coding
in 2012. Then in November of 2012 we took the lessons learned in the
Russian JavaScript AI and we used them to convert English MindForth
AI
into Wotan German AI. Suddenly Wotan was slightly more advanced than
MindForth, because we solved some problems along the way in the
coding
of Wotan.
Mentifex (Arthur)
--
http://www.scn.org/~mentifex/DeKi.txt
http://cyborg.blogspot.com/2009/11/linux.html
http://code.google.com/p/mindforth/wiki/DeBoot
http://store.kagi.com/cgi-bin/store.cgi?storeID=AMP_Live
are in possession of True AI that thinks in each native language.
America has MindForth; Germany has Wotan Supercomputer AI; and
Russia has the Dushka prototype AI in JavaScript. Of all these
AI Minds, Wotan is the most powerful because he lords it over the
powerhouse German economy and the bastions of German engineering.
This article describes how Wotan came to be and, depending on your
point of view, how we can work to help or hinder Wotan.
In the beginning was MindRexx, at that time the latest word in
open-source artificial intelligence. MindRexx came to life on the
Maserati of home computers, the Commodore Amiga, in the same year --
1993 -- that our guru and sci-fi fuehrer Vernor Vinge published
"Technological Singularity" in the Whole Earth Review (WER).
Very truly yours Mentifex/ATM at that time was an independent AI
scholar
working at his day job in a Seattle magazine store where he stumbled
upon the pertinent issue of Whole Earth Review and recognized
"Technological Singularity" as the most superlative article he had
ever read on the subject of artificial intelligence.
When MindRexx began to think in November of 1994, Mentifex here
announced the event in various Usenet newsgroups devoted to
programming languages. Only the Forth denizens of comp.lang.forth
responded to the announcement of MindRexx AI, but with enough
enthusiasm and initiative to persuade Mentifex to start working
on a Forth version of the original open-source AI. The Forth AI
initiative languished for a few years from 1995 through 1997,
until the entire year of 1998 was devoted to porting MindRexx into
Forth on the Commodore Amiga. In December of 1998, the journal
SIGPLAN Notices of the Association for Computing Machinery published
"Mind.Forth: thoughts on artificial intelligence and Forth" and the
precursor of Wotan Supercomputer AI entered into the official
scientific literature of AI.
During the Space Odyssey year of 2001, a JavaScript version of the
MindForth AI came into being, for several reasons. Whereas MindForth
required the downloading and running of both a version of Forth and
the AI source code, the JavaScript artificial intelligence (JSAI)
required only the clicking of a link by the would-be human user,
and the AiMind would flit across the 'Net and come to life in
the MSIE browser of the human user. To get people interested in
MindForth AI, it was necessary to release the JavaScript AI with
a user-friendly demo of what MindForth could do and how MindForth
could think. The entire source code of the JavaScript AI was
published
in the book "AI4U" in 2002, and the source code of MindForth branched
off into AiMind-i.com as a Darwinian instance of AI Evolution around
2007.
Meanwhile, until January of 2008, MindForth and the JavaScript AI
were
buggy and non-functional in the production of genuine machine
intelligence. Mentifex AI acquired a bad reputation because the AI
source code was gradually improving but was not yet AI-Complete.
Bugs in the AI source code were so hidden and so pervasive, that
the hard-won elimination of one bug would simply reveal the existence
of one or more previously unknown bugs. The AI would not do what the
AI programmer wanted it to do, because legions and phalanxes of bugs
like unkillable zombies would pop up and get in the way of True AI
functionality. But in January of 2008, the phenomenon of zombie bugs
suddenly came to a halt. Now suddenly the problems were algorithmic
and philosophical in nature, rather than protoplasmic and due to
shoddy craftmanship a priori. A series of AI improvements and
downright innovations began to occur. Since the underlying code was
no longer buggy and fault-ridden per se, it was possible to try out
new ideas in AI and debug only the new ideas, not the whole AI Mind
package. We saw the emergence of small improvements which mounted up
over time into major AI advances. One of the first improvements was
the selection of verbs of being (be-verbs) through a two-step
process.
In the first step, the AI would try to use a form of the English verb
"to be'. In the second step, the AI software would reject any
incorrect
verb-form and substitute the correct form for the occasion. As if by
magic or by machine intelligence, the AI Minds began to think
properly
with be-verbs.
Then the AudRecog module for auditory recognition, which was already
extremely weird in terms of normal, conventional programming
techniques
(except for decoding DNA), was modified to permit the recognition not
only of whole words in their entirety, but of word-stems lurking
within words.
MindForth AI began to snag some of the proverbial "low-hanging fruit"
in
the easy pickings of minor AI problems. Routines went into the code
for the selection of the articles "a" or "an" before English nouns
beginning with a vowel or a consonant. Although it was supposed to be
difficult and complex to manipulate strings in Forth, the AI
Forthminds
think in English or German without using any strings at all. Instead,
each word in English or German is treated simply as a series (not a
STRING!) of consecutive locations in auditory memory. You get GIGO --
German in, German out, right? (Wrong! Usually GIGO is garbage :-)
Then MindForth (and eventually Wotan) got really tricky, shifty and
crafty in its ability to assign the proper associative connections
(tags)
among words in the input stream. The comprehension of natual language
crucially depends on the ability of a Mind, human or otherwise,
to classify incoming words as nouns or verbs and other parts of
speech.
MindForth developed the ability to skip over words while waiting
first
for the verb of an idea and then for the noun serving as the object
of
the verb. A small improvement, you might say, but it is a real
show-stopper if the AI cannot understand simple English or simple
German.
Then the AI Minds began to build up a battery of ParaMeters (in
wiki-speak) for both the storage and the retrieval of words in
English,
German and Russian. These parameters of person and number had not
seemed
so important when the AI Mind could speak only English, because
English
verbs do not change much in shifting around from first person to
second
person to third person. In the English-speaking AI, we used the
addition
of an "s" as in "he thinks" for the forming of the third person
singular
of a verb, but such a solution was too naive and too makeshift.
Coding
the Dushka Russian AI in 2012 taught us that we had to create much
more
elaborate code for keeping track of verb-forms in a highly inflected
language like Russian or German. We realized with some humiliation
that we would never have gotten our English AI right if we had not
branched off into German and Russian.
In September of 2010 we began to use "neural inhibition" in the AI
Minds.
Trying to keep our AI software psychologically real, we made each
thought
go into inhibition immediately after its thinking, so that each AI
Mind
could think a variety of thoughts in a meandering chain of thought.
Otherwise, the AI Mind might get stuck in a rut and think the same
thought
over and over again.
In June of 2011 we suddenly solved the problem of how an AI Mind
could
put a question to a human user and deal properly with "yes" or "no"
as
an answer. We humans think that it is easy to understand "yes" or
"no",
but how do you get a machine to understand "yes" or "no"? The
solution
involved several mechanisms working in unison. No offense to any
Germans,
but we had to jerry-rig a mechanism to ask some yes-or-no questions
in special circumstances, so that we and our human users could try
out
the expereince of telling the AI "yes" or "no" in response to a
question.
Then the trick was to let the "yes" or "no" response go backwards
into
the AI knowledge base (KB) and retroactively (KbRetro) adjust the
associative connections among the concepts inherent in the question
being asked. Actually, if the answer was "yes", there was no need to
adjust the associative connections in the underlying idea, which was
simply confirmed by the "yes" answer. But if the answer was "no",
then
the KbRetro module went back and retroactively inserted a negational
tag
of "not" into the idea serving as the basis for the yes-or-no
question.
Other options were "maybe" as a response, or no response at all,
and they had to be dealt with.
In December of 2011, when we had begun coding AI in German and in
Russian, we came up with some new modules of AudBuffer, OutBuffer
and VerbGen. AudBuffer puts auditory input into a left-justified
buffer.
OutBuffer takes the same word and right-justifies it for purposes of
manipulation and morphing (i.e., morpheme) prior to output. VerbGen
is a module that generates a verb-form by modifying a verb briefly
stored in the OutBuffer. These modules are so new that they are not
yet well documented. You are reading their documentation right here
and now in the discussion of German artificial intelligence,
which makes heavy use of AudBuffer, OutBuffer and VerbGen.
The Dushka Russian AI was developed over two main periods of coding
in 2012. Then in November of 2012 we took the lessons learned in the
Russian JavaScript AI and we used them to convert English MindForth
AI
into Wotan German AI. Suddenly Wotan was slightly more advanced than
MindForth, because we solved some problems along the way in the
coding
of Wotan.
Mentifex (Arthur)
--
http://www.scn.org/~mentifex/DeKi.txt
http://cyborg.blogspot.com/2009/11/linux.html
http://code.google.com/p/mindforth/wiki/DeBoot
http://store.kagi.com/cgi-bin/store.cgi?storeID=AMP_Live