burp

reality is almost vr'd

On 2/7/12, A. T. Murray <mentifex@scn.org> wrote:
> Rushing to complete the Singularity AI User Manual:
>
> 4 Limitations
>
> In any human language, the verbs are the words that describe
> an action or a state of being. In Russian, German, Latin and
> Greek, the verbs with their inflectional endings are more
> complicated than in English. Dushka, as a primitive artificial
> intelligence, has the goal and also the limitation of being
> fluent with one class of Russian verbs like delat' ("to do")
> and dumat' ("to think" in the present tense. You may use
> such a verb to converse with Dushka in the present tense
> using pronouns or nouns as subjects and objects of the verb.
> If Dushka does not know the verb, her software may assume
> that the verb is like delat' or dumat; and she may answer
> you with verb forms of the same class of verbs. You may even
> invent a fictitious verb in the same conjugation as delat'
> and use forms of the verb to test or explore the ability
> of Dushka to comprehend Russian verb forms and to convert one
> form of a Russian verb into a different form of the same verb.
>
> 5 Comprehension
>
> Dushka has a sophisticated method of comprehending the input of
> a declarative sentence in Russian. After the entry of a noun or
> pronoun as the subject of a sentence, the InStantiate module
> begins to expect a verb as the main carrier of the idea in
> the sentence. When the verb comes in, the InStantiate software
> tags the verb with the special parameters of person and number.
> For example, if you say, "Russians know poetry" and you use the
> verb "znayut: for "they know", Dushka tags the verb with a "num"
> (number) tag for plural and a "dba" person tag of three ("3") for
> the third person. By using the parameter tags to search for the
> third-person plural form of znat', Dushka will be able to find
> the same verb-form in the future durng the generation of a thought
> in Russian. If Dushka needs a verb-form but can not find it tagged
> in memory with parameter tags, the VerbPhrase mind-module calls
> the VerbGen module to create the required form of the Russian verb.
> In the current releases of the Dushka software, VerbGen can generate
> the needed verb-forms only in the present tens and only for real or
> fictitious verbs similar to delat' and dumat'. If you are coding
> an artificial intelligence that will go beyond the proof-of-concept that
> you find in Dushka, you or your team of programmers will need to give
> the software the ability to handle all the many different kinds of
> Russian verbs in all the tenses and all the voices and all the moods.
> The problems of artificial intelligence in Russian are very complex.
> Dushka is only the start on the path to a superintelligence that lives
> on a supercomputer and thinks not only in Russian but also in many
> other languages.
>
> 6 English
>
> Because JavaScript can handle the input characters of both Russian
> and English, you may input an English word to Dushka and she will
> think in Russian about the English word. In a similar way, speakers
> of English may think about Russian words like "tundra" or "samovar"
> without possessing the deeply philosophical and emotional soul of
> a true Russian. (Americans cannot understand the longing that
> expatriate Russians feel for their land and for their language.)
> If you want to mention an English name like "James" (i.e., shaken,
> not stirred) to Dushka, remember to toggle (Alt-Shift) your keyboard
> briefly into English, type "james" and then toggle (Alt-Shift) back
> into Russian.
>
> When the Russian artificial intelligence evolves into a more powerful
> mind approaching SuperIntelligence on a 64-bit platform coded in a
> more powerful language like iForth or GodSpeed, the mere entry of
> a few consecutive words in English will toggle the whole Mind
> from thinking in Russian to thinking in English. Meanwhile the
> JavaScript Dushka would run too slow if she were merged in a
> kind of Vulcan mindmeld with the English AI Mind.
>
> Mentifex
> --
> http://www.scn.org/~mentifex/RuAiUser.html
>
>
> -------------------------------------------
> singularity
> Archives: https://www.listbox.com/member/archive/11983/=now
> RSS Feed: https://www.listbox.com/member/archive/rss/11983/11056020-3396cbf8
> Modify Your Subscription:
> https://www.listbox.com/member/?&amp;
> Powered by Listbox: http://www.listbox.com
>

On 07/02/2012 12:35, Matt Mahoney wrote:

> How many bits are added by the zygote, by the exact arrangements
> of the atoms, by the laws of physics, and so on?

The zygote is probably fairly small - most of that looks as though it is coded for
by DNA to me.

The laws of physics are the real unknown, IMO. These are *probably* small -
and it would be counter-intuitive for a change to the 9 millionth bit of, say
the fine structure constant to have large effects - but there *is* the
phenomenon of chaos - where small changes can have large effects. I can't
yet think of a good way of completely ruling out "large" physical laws.
They *do* seem contrary to Occam - though I don't even know if we can
usefullly apply Occam to physical law itself.
__________
|im |yler http://timtyler.org/ tim@tt1lock.org Remove lock to reply.

On Tue, Feb 7, 2012 at 11:25 AM, Jim Bromer <jimbromer@gmail.com> wrote:
> On Mon, Feb 6, 2012 at 10:51 PM, Matt Mahoney <mattmahoneyfl@gmail.com> wrote:
>> First, AI will appear as one, because agents will communicate so fast
>> that they will all have the same knowledge. I'm sure you're familiar
>> with the power of the internet.
>>
> No, that is not right.  Intelligence is the ability to leverage a small ability so that it can act as a much more powerful machine through the use of focus and method.   By focusing on some narrow specialties we human beings - who are of course capable of general intelligence - are able to leverage our special talents way beyond what our everyday-living-in-the-world-skills might otherwise suggest.  The problem with AGI is that computer programs are not yet able to generate those most general skills in reasoning and they therefore have much less leverage to use on more specialty problems.

What does that have to do with AIs having instant access to all of the
world's information, just like we do? AI is going to look like a
smarter internet, one that can understand language and pictures, and
knows everything about everyone.


-- Matt Mahoney, mattmahoneyfl@gmail.com

On Tue, Feb 7, 2012 at 11:06 AM, Todor Arnaudov <twenkid@gmail.com> wrote:
>
> Matt,
>
> Matt>I will remind you that I am talking about the length of the shortest
> universal Turing machine program that does what a baby brain does. The
> length of your DNA is an upper bound on that uncomputable number.
>
> The most annoying about your brain-babies etc. "informed" estimations is that you don't seem to be familiar at all with biology, neuroscience, developmental psychology etc.
>
> No, it's not the upper bound, DNA without zygote etc. etc. is meaningless.
> A human reading DNA consciously from the screen or using computers etc. is a system including zillions of bits, needed to describe those sub systems which are involved, not to talk about the complexity of the machine required to *BUILD* all those *PHYSICAL* entities out of raw atoms.

"Zillions" is not a real number. I gave you my numbers. Please give me
yours. No hand-waving. How many bits are added by the zygote, by the
exact arrangements of the atoms, by the laws of physics, and so on?
And explain why this should have a significant effect on the cost of
software. I can write a program to simulate another computer's
behavior without modeling it at the atomic level.

> Matt> Who is going to pay for AI if it costs $1 quadrillion and has no
> commercial value? And if you think I'm off by a dozen orders of magnitude, then please
> tell me what you think it will cost and how you got your numbers.
>
> To me those estimations are meaningless, you spit random numbers correlated with random values taken from some respectful statistics. GDP is biased and not essential, what's essential are IQ, talents, technology available and human-work-years needed, not dirty "money".

My numbers are not random. I explained where they came from and why they matter.

> If the whole history of humanity is taken into account, the price is enormous.

Numbers, please?

> If the price is taken as particular "purposeful investments" -- "these are money, dedicated for AGI by...", in my opinion AGI may happen to be extremely cheap, because as of now the ones who work in this field are self-funded.
>
> It doesn't require billions,  just enough for the ones who're smart enough and involved enough to work comfortably.

I await your solution.


-- Matt Mahoney, mattmahoneyfl@gmail.com

On Mon, Feb 6, 2012 at 10:51 PM, Matt Mahoney <mattmahoneyfl@gmail.com>wrote:


> First, AI will appear as one, because agents will communicate so fast
> that they will all have the same knowledge. I'm sure you're familiar
> with the power of the internet.
>
>
No, that is not right. Intelligence is the ability to leverage a small
ability so that it can act as a much more powerful machine through the
use of focus and method. By focusing on some narrow specialties we human
beings - who are of course capable of general intelligence - are able to
leverage our special talents way beyond what our
everyday-living-in-the-world-skills might otherwise suggest. The problem
with AGI is that computer programs are not yet able to generate those most
general skills in reasoning and they therefore have much less leverage to
use on more specialty problems.
Jim Bromer

Matt,

Matt>I will remind you that I am talking about the length of the shortest
universal Turing machine program that does what a baby brain does. The
length of your DNA is an upper bound on that uncomputable number.

The most annoying about your brain-babies etc. "informed" estimations is
that you don't seem to be familiar at all with biology, neuroscience,
developmental psychology etc.

No, it's not the upper bound, DNA without zygote etc. etc. is meaningless.
A human reading DNA consciously from the screen or using computers etc. is
a system including zillions of bits, needed to describe those sub systems
which are involved, not to talk about the complexity of the machine
required to *BUILD* all those *PHYSICAL* entities out of raw atoms.


Matt> Who is going to pay for AI if it costs $1 quadrillion and has no
commercial value? And if you think I'm off by a dozen orders of magnitude,
then please
tell me what you think it will cost and how you got your numbers.

To me those estimations are meaningless, you spit random numbers correlated
with random values taken from some respectful statistics. GDP is biased and
not essential, what's essential are IQ, talents, technology available and
human-work-years needed, not dirty "money".

If the whole history of humanity is taken into account, the price is
enormous.

If the price is taken as particular "purposeful investments" -- "these are
money, dedicated for AGI by...", in my opinion AGI may happen to be *extremely
cheap, *because as of now the ones who work in this field are self-funded.

It doesn't require billions, just enough for the ones who're smart enough
and involved enough to work comfortably.


Matt> Please list all these idiot billionaires that don't know about
investing for the long term.

I'm not Forbes to list billionaires, you list one who has invested in AGI
after told that "there might be results in 20-30-40-50 years". Billionaires
don't need to invest in genii, since the most profitable fields are dumb,
such as oil and all kinds of trading or showbusiness, which serve human
immediate needs.

Who needs AGI? People need food, water, sex, repetitive fun routines,
health-care etc., most people are "hardware-ly" not advanced enough to
have such abstract goals and to notice how repetitive, static and boring
most of their live is.

-- Todor
http://research.twenkid.com

On Tue, Feb 7, 2012 at 5:38 PM, Matt Mahoney <mattmahoneyfl@gmail.com>wrote:

> On Tue, Feb 7, 2012 at 10:20 AM, Todor Arnaudov <twenkid@gmail.com> wrote:
> >
> > Matt,
> >
> > AGI is about understanding and implementing general intelligence, it's
> neither about making people more productive, nor about replacing them.
> >
> > Yes, AI is about that, and it's one reason why it was stuck -- "greedy"
> programmers at work.
> >
> > Alan is right about the alteration of current economy in case AGI starts
> doing the "dirty" job for people.
>
> Who is going to pay for AI if it costs $1 quadrillion and has no
> commercial value?
>
> And if you think I'm off by a dozen orders of magnitude, then please
> tell me what you think it will cost and how you got your numbers.
>
> > Matt>I am interested in solving AI just like you.
> >
> > AI? That's a wrong list then.
>
> AGI, whatever. More people know what AI is than AGI.
>
> > Matt>But $'s followed by numbers *are* relevant. Do you really think the
> rest of the world is
> > too stupid to solve the problem, so they just keep paying people $63
> > trillion per year to work?
> >
> > Yes, they obviously are "too stupid", one of the people's problem is
> that they are greedy and want revenues *right now*, they have short
> "attention span", "ADD". Most people are not wise or rich enough to afford
> long time introspection and analysis without producing results which are
> "useful" or "meaningful" for the bunch of others impatient agents with
> shorter span, upon which everyone is dependent - it's a "society".
> >
> > Some problems require long focus and wider span of knowledge to be
> solved, and that requires ultra interdisciplinary mind, which yes - "the
> rest of the world" lacks.
>
> Please list all these idiot billionaires that don't know about
> investing for the long term.
>
> > Matt> Your cognitive architecture is encoded in 6 x 10^9 bits of DNA.
> >
> > This is one of your many delusions and confusions. I'll remind it:
> >
> > DNA is that bits only if stored on a hard drive and evaluated in
> computer terms, in physical world it requires much more, and in order to
> store this data on a HDD, thousands of years of civilization and
> technology are needed, in order to develop means to sequence it and encode
> it so.
> >
> > Human DNA makes *physical* sense only if it's inside a zygote, and in
> physical world the exact molecular structure of the zygote should be kept.
> Then, you should simulate development of the fetus, the uterus and mother's
> blood supply etc. atom by atom (DNA is a molecule, it's not a computer
> algorithm). Then, after simulating the birth, the growth of the body should
> be simulated -- with all its atoms, and of course that simulation need to
> interact with reality.
>
> I will remind you that I am talking about the length of the shortest
> universal Turing machine program that does what a baby brain does. The
> length of your DNA is an upper bound on that uncomputable number.
>
>
> -- Matt Mahoney, mattmahoneyfl@gmail.com
>
>
> -------------------------------------------
> AGI
> Archives: https://www.listbox.com/member/archive/303/=now
> RSS Feed: https://www.listbox.com/member/archive/rss/303/21253656-fab278b9
> Modify Your Subscription:
> https://www.listbox.com/member/?&amp;
> Powered by Listbox: http://www.listbox.com
>



--
-- Todor "Tosh" Arnaudov
http://research.twenkid.com
http://artificial-mind.blogspot.com

On 06/02/2012 22:51, Matt Mahoney wrote:

> First, AI will appear as one, because agents will communicate so fast
> that they will all have the same knowledge. I'm sure you're familiar
> with the power of the internet.

/Yes, though there's *also* the power of
The Monopolies and Mergers Commission.

For governmental machines, that might not be a problem.
However, if we get corporate intelligent machines (rather
like we have today), the government might try hard to
prevent them from getting too powerful. Big, powerful
agents often don't like smaller potential rivals. Some
"interesting" dynamics might follow.
__________
|im |yler http://timtyler.org/ tim@tt1lock.org Remove lock to reply.
/

On Tue, Feb 7, 2012 at 10:20 AM, Todor Arnaudov <twenkid@gmail.com> wrote:
>
> Matt,
>
> AGI is about understanding and implementing general intelligence, it's neither about making people more productive, nor about replacing them.
>
> Yes, AI is about that, and it's one reason why it was stuck -- "greedy" programmers at work.
>
> Alan is right about the alteration of current economy in case AGI starts doing the "dirty" job for people.

Who is going to pay for AI if it costs $1 quadrillion and has no
commercial value?

And if you think I'm off by a dozen orders of magnitude, then please
tell me what you think it will cost and how you got your numbers.

> Matt>I am interested in solving AI just like you.
>
> AI? That's a wrong list then.

AGI, whatever. More people know what AI is than AGI.

> Matt>But $'s followed by numbers *are* relevant. Do you really think the rest of the world is
> too stupid to solve the problem, so they just keep paying people $63
> trillion per year to work?
>
> Yes, they obviously are "too stupid", one of the people's problem is that they are greedy and want revenues *right now*, they have short "attention span", "ADD". Most people are not wise or rich enough to afford long time introspection and analysis without producing results which are "useful" or "meaningful" for the bunch of others impatient agents with shorter span, upon which everyone is dependent - it's a "society".
>
> Some problems require long focus and wider span of knowledge to be solved, and that requires ultra interdisciplinary mind, which yes - "the rest of the world" lacks.

Please list all these idiot billionaires that don't know about
investing for the long term.

> Matt> Your cognitive architecture is encoded in 6 x 10^9 bits of DNA.
>
> This is one of your many delusions and confusions. I'll remind it:
>
> DNA is that bits only if stored on a hard drive and evaluated in computer terms, in physical world it requires much more, and in order to store this data on a HDD,  thousands of years of civilization and technology are needed, in order to develop means to sequence it and encode it so.
>
> Human DNA makes *physical* sense only if it's inside a zygote, and in physical world the exact molecular structure of the zygote should be kept. Then, you should simulate development of the fetus, the uterus and mother's blood supply etc. atom by atom (DNA is a molecule, it's not a computer algorithm). Then, after simulating the birth, the growth of the body should be simulated -- with all its atoms, and of course that simulation need to interact with reality.

I will remind you that I am talking about the length of the shortest
universal Turing machine program that does what a baby brain does. The
length of your DNA is an upper bound on that uncomputable number.


-- Matt Mahoney, mattmahoneyfl@gmail.com

On 07/02/2012 10:20, Todor Arnaudov wrote:

> Matt>I am interested in solving AI just like you.
>
> AI? That's a wrong list then.

To rehash and old discussuion, the "G" in "AGI" looks
pretty redundant. Intelligence is general, by many
definitions. If you are working on a non-general
system, "expert" is one of the terms available.

My rant on the topic:

"Tim Tyler: Terminology for classifying machine intelligence"

http://www.youtube.com/watch?v=Ow_Wuz-Jf8Q
__________
|im |yler http://timtyler.org/ tim@tt1lock.org Remove lock to reply.

On Tue, Feb 7, 2012 at 10:05 AM, Tim Tyler <tim@tt1.org> wrote:
>
> On 06/02/2012 22:51, Matt Mahoney wrote:
>
> Second, AI won't be cheap. If AI is to have commercial value, then it has to have human knowledge.
>
> It has to have *some* human knowledge.  Google is based on intelligent
> machines that have *some* human knowledge - and it is still valuable to
> its (billionaire) owners.

More is better, right? Google isn't AI yet.


-- Matt Mahoney, mattmahoneyfl@gmail.com

Matt,

*AGI is about understanding and implementing general intelligence*, it's
neither about making people more productive, nor about replacing them.

Yes, AI is about that, and it's one reason why it was stuck -- "greedy"
programmers at work.

Alan is right about the alteration of current economy in case AGI starts
doing the "dirty" job for people.

Matt>I am interested in solving AI just like you.

AI? That's a wrong list then.


Matt>But $'s followed by numbers *are* relevant. Do you really think the
rest of the world is
too stupid to solve the problem, so they just keep paying people $63
trillion per year to work?

Yes, they obviously are "too stupid", one of the people's problem is that
they are greedy and want revenues *right now*, they have short "attention
span", "ADD". Most people are not wise or rich enough to afford long time
introspection and analysis without producing results which are "useful" or
"meaningful" for the bunch of others impatient agents with shorter span,
upon which everyone is dependent - it's a "society".

Some problems require long focus and wider span of knowledge to be solved,
and that requires ultra interdisciplinary mind, which yes - "the rest of
the world" lacks.

Matt> Your cognitive architecture is encoded in 6 x 10^9 bits of DNA.

This is one of your many delusions and confusions. I'll remind it:

DNA is that bits only if stored on a hard drive and evaluated in computer
terms, in physical world it requires much more, and in order to store this
data on a HDD, thousands of years of civilization and technology
are needed, in order to develop means to sequence it and encode it so.

Human DNA makes *physical* sense only if it's inside a zygote, and in
physical world the exact molecular structure of the zygote should be kept.
Then, you should simulate development of the fetus, the uterus and mother's
blood supply etc. atom by atom (DNA is a molecule, it's not a computer
algorithm). Then, after simulating the birth, the growth of the body should
be simulated -- with all its atoms, and of course that simulation need to
interact with reality.

-- Todor
http://research.twenkid.com

On Tue, Feb 7, 2012 at 4:55 PM, Matt Mahoney <mattmahoneyfl@gmail.com>wrote:

> On Mon, Feb 6, 2012 at 11:42 PM, Alan Grimes <agrimes@speakeasy.net>
> wrote:
> > $'s followed by numbers are moot and irrelevant in this context.
> >
> > Your ambition of replacing human workers is moot, irrelevant, and
> > uninteresting.
>
> AI is about making people more productive, not about putting people
> out of work. You own the machines, and they do what you want. Sorry if
> I was misunderstood. I use the cost of human labor to estimate the
> value of AI.
>
> > Continuing to repeat it over and over again is not contributing to the
> > discussion. Admittedly, my posting wasn't quite as great as it seemed to
> be
> > when I was first assembling it in my mind before I started writing it.
>
> Perhaps I missed the point. You gave some good arguments against
> uploading that I agree with.
>
> > What we need is a technology that can be applied in general to complex
> > cybernetic systems and produce agents which posses many of the
> > characteristics of human intelligence. What you turn around and do with
> that
> > technology is completely up to you (just don't upload me!) I have certain
> > specific deviant goals but they may or may not actually be possible. I
> don't
> > talk about them because my ideas are off-topic and irrelevant (not to
> > mention embarrassing =P).
>
> I am interested in solving AI just like you. But $'s followed by
> numbers *are* relevant. Do you really think the rest of the world is
> too stupid to solve the problem, so they just keep paying people $63
> trillion per year to work?
>
> Your cognitive architecture is encoded in 6 x 10^9 bits of DNA. An
> equivalent piece of software could be written for around $1 billion at
> the normal rate of $100 per line of code. That is the problem that
> everyone is working on, as if that were the only part that needs to be
> solved. But it's not. It is such a tiny fraction of the cost that we
> can ignore it.
>
> AI is coming, so pay attention. We are moving to a world where
> everyone knows everything about everyone. That knowledge will come
> from billions of public webcams and billions of computers that
> everyone carries around with them all the time, in your pocket, in
> your car, in your TV, and in front of you now. What's missing is the
> computing power to interpret all of this data, so you can search
> through billions of years worth of video for pictures of you. That
> will require 10^26 OPS and 10^25 bytes of memory, the equivalent
> computing power of 7 billion human brains. Right now, that would be 10
> to 100 times more expensive than human labor. That is why AI hasn't
> been solved yet. But Moore's law will change that. And then AI will
> happen, but slowly, because it will still take years longer for
> machines to learn what you know.
>
>
> -- Matt Mahoney, mattmahoneyfl@gmail.com
>
>
> -------------------------------------------
> AGI
> Archives: https://www.listbox.com/member/archive/303/=now
> RSS Feed: https://www.listbox.com/member/archive/rss/303/21253656-fab278b9
> Modify Your Subscription:
> https://www.listbox.com/member/?&amp;
> Powered by Listbox: http://www.listbox.com
>



--
-- Todor "Tosh" Arnaudov
http://research.twenkid.com
http://artificial-mind.blogspot.com

On 06/02/2012 22:51, Matt Mahoney wrote:

> Second, AI won't be cheap. If AI is to have commercial value, then it has to have human knowledge.

It has to have *some* human knowledge. Google is based on intelligent
machines that have *some* human knowledge - and it is still valuable to
its (billionaire) owners.
__________
|im |yler http://timtyler.org/ tim@tt1lock.org Remove lock to reply.

Rushing to complete the Singularity AI User Manual:

4 Limitations

In any human language, the verbs are the words that describe
an action or a state of being. In Russian, German, Latin and
Greek, the verbs with their inflectional endings are more
complicated than in English. Dushka, as a primitive artificial
intelligence, has the goal and also the limitation of being
fluent with one class of Russian verbs like delat' ("to do")
and dumat' ("to think" in the present tense. You may use
such a verb to converse with Dushka in the present tense
using pronouns or nouns as subjects and objects of the verb.
If Dushka does not know the verb, her software may assume
that the verb is like delat' or dumat; and she may answer
you with verb forms of the same class of verbs. You may even
invent a fictitious verb in the same conjugation as delat'
and use forms of the verb to test or explore the ability
of Dushka to comprehend Russian verb forms and to convert one
form of a Russian verb into a different form of the same verb.

5 Comprehension

Dushka has a sophisticated method of comprehending the input of
a declarative sentence in Russian. After the entry of a noun or
pronoun as the subject of a sentence, the InStantiate module
begins to expect a verb as the main carrier of the idea in
the sentence. When the verb comes in, the InStantiate software
tags the verb with the special parameters of person and number.
For example, if you say, "Russians know poetry" and you use the
verb "znayut: for "they know", Dushka tags the verb with a "num"
(number) tag for plural and a "dba" person tag of three ("3") for
the third person. By using the parameter tags to search for the
third-person plural form of znat', Dushka will be able to find
the same verb-form in the future durng the generation of a thought
in Russian. If Dushka needs a verb-form but can not find it tagged
in memory with parameter tags, the VerbPhrase mind-module calls
the VerbGen module to create the required form of the Russian verb.
In the current releases of the Dushka software, VerbGen can generate
the needed verb-forms only in the present tens and only for real or
fictitious verbs similar to delat' and dumat'. If you are coding
an artificial intelligence that will go beyond the proof-of-concept that
you find in Dushka, you or your team of programmers will need to give
the software the ability to handle all the many different kinds of
Russian verbs in all the tenses and all the voices and all the moods.
The problems of artificial intelligence in Russian are very complex.
Dushka is only the start on the path to a superintelligence that lives
on a supercomputer and thinks not only in Russian but also in many
other languages.

6 English

Because JavaScript can handle the input characters of both Russian
and English, you may input an English word to Dushka and she will
think in Russian about the English word. In a similar way, speakers
of English may think about Russian words like "tundra" or "samovar"
without possessing the deeply philosophical and emotional soul of
a true Russian. (Americans cannot understand the longing that
expatriate Russians feel for their land and for their language.)
If you want to mention an English name like "James" (i.e., shaken,
not stirred) to Dushka, remember to toggle (Alt-Shift) your keyboard
briefly into English, type "james" and then toggle (Alt-Shift) back
into Russian.

When the Russian artificial intelligence evolves into a more powerful
mind approaching SuperIntelligence on a 64-bit platform coded in a
more powerful language like iForth or GodSpeed, the mere entry of
a few consecutive words in English will toggle the whole Mind
from thinking in Russian to thinking in English. Meanwhile the
JavaScript Dushka would run too slow if she were merged in a
kind of Vulcan mindmeld with the English AI Mind.

Mentifex
http://www.scn.org/~mentifex/RuAiUser.html

On Mon, Feb 6, 2012 at 11:42 PM, Alan Grimes <agrimes@speakeasy.net> wrote:
> $'s followed by numbers are moot and irrelevant in this context.
>
> Your ambition of replacing human workers is moot, irrelevant, and
> uninteresting.

AI is about making people more productive, not about putting people
out of work. You own the machines, and they do what you want. Sorry if
I was misunderstood. I use the cost of human labor to estimate the
value of AI.

> Continuing to repeat it over and over again is not contributing to the
> discussion. Admittedly, my posting wasn't quite as great as it seemed to be
> when I was first assembling it in my mind before I started writing it.

Perhaps I missed the point. You gave some good arguments against
uploading that I agree with.

> What we need is a technology that can be applied in general to complex
> cybernetic systems and produce agents which posses many of the
> characteristics of human intelligence. What you turn around and do with that
> technology is completely up to you (just don't upload me!) I have certain
> specific deviant goals but they may or may not actually be possible. I don't
> talk about them because my ideas are off-topic and irrelevant (not to
> mention embarrassing =P).

I am interested in solving AI just like you. But $'s followed by
numbers *are* relevant. Do you really think the rest of the world is
too stupid to solve the problem, so they just keep paying people $63
trillion per year to work?

Your cognitive architecture is encoded in 6 x 10^9 bits of DNA. An
equivalent piece of software could be written for around $1 billion at
the normal rate of $100 per line of code. That is the problem that
everyone is working on, as if that were the only part that needs to be
solved. But it's not. It is such a tiny fraction of the cost that we
can ignore it.

AI is coming, so pay attention. We are moving to a world where
everyone knows everything about everyone. That knowledge will come
from billions of public webcams and billions of computers that
everyone carries around with them all the time, in your pocket, in
your car, in your TV, and in front of you now. What's missing is the
computing power to interpret all of this data, so you can search
through billions of years worth of video for pictures of you. That
will require 10^26 OPS and 10^25 bytes of memory, the equivalent
computing power of 7 billion human brains. Right now, that would be 10
to 100 times more expensive than human labor. That is why AI hasn't
been solved yet. But Moore's law will change that. And then AI will
happen, but slowly, because it will still take years longer for
machines to learn what you know.


-- Matt Mahoney, mattmahoneyfl@gmail.com

Matt Mahoney wrote:
>> Once you have your first AGI, your Nth AGI is practically free.

> First, AI will appear as one, because agents will communicate so fast
> that they will all have the same knowledge. I'm sure you're familiar
> with the power of the internet.

> Second, AI won't be cheap. If AI is to have commercial value, then it
> has to have human knowledge. If I had a model of your mind, then I
> could predict what you would buy and how much you would pay. Why else
> would I want to build AI?

> We can estimate the cost of acquiring that knowledge using information
> theory. Your long term memory is 10^9 bits [1]. I can get 99% [2] of
> what you know from other sources, such as the internet or other
> people. But the other 1% has to pass through channels like speech or
> typing at a rate of a few bits per second [3]. When you multiply by 7
> billion people earning on average $5 per hour, the cost is on the
> order of $100 trillion to $1 quadrillion. This is why we are still
> paying people $63 trillion per year worldwide to do work that machines
> could do if they were smarter [4].

> If you think there is an easier way to acquire this knowledge, please
> let me know. I considered brain scanning, but it would have to cost
> under $100,000 per person to be competitive with the global system of
> public surveillance that we are now building.

$'s followed by numbers are moot and irrelevant in this context.

Your ambition of replacing human workers is moot, irrelevant, and
uninteresting.

Continuing to repeat it over and over again is not contributing to the
discussion. Admittedly, my posting wasn't quite as great as it seemed to
be when I was first assembling it in my mind before I started writing it.

What we need is a technology that can be applied in general to complex
cybernetic systems and produce agents which posses many of the
characteristics of human intelligence. What you turn around and do with
that technology is completely up to you (just don't upload me!) I have
certain specific deviant goals but they may or may not actually be
possible. I don't talk about them because my ideas are off-topic and
irrelevant (not to mention embarrassing =P).

--
E T F
N H E
D E D

Powers are not rights.

On Mon, Feb 6, 2012 at 9:25 PM, Alan Grimes <agrimes@speakeasy.net> wrote:
> Damnit, I wish there was some way I could help you get past this single idea
> that you keep harping on.
>
> Once you have your first AGI, your Nth AGI is practically free.

First, AI will appear as one, because agents will communicate so fast
that they will all have the same knowledge. I'm sure you're familiar
with the power of the internet.

Second, AI won't be cheap. If AI is to have commercial value, then it
has to have human knowledge. If I had a model of your mind, then I
could predict what you would buy and how much you would pay. Why else
would I want to build AI?

We can estimate the cost of acquiring that knowledge using information
theory. Your long term memory is 10^9 bits [1]. I can get 99% [2] of
what you know from other sources, such as the internet or other
people. But the other 1% has to pass through channels like speech or
typing at a rate of a few bits per second [3]. When you multiply by 7
billion people earning on average $5 per hour, the cost is on the
order of $100 trillion to $1 quadrillion. This is why we are still
paying people $63 trillion per year worldwide to do work that machines
could do if they were smarter [4].

If you think there is an easier way to acquire this knowledge, please
let me know. I considered brain scanning, but it would have to cost
under $100,000 per person to be competitive with the global system of
public surveillance that we are now building.

>> [Pro-uploading drivel removed]

I'm not pro or anti uploading. I do believe that the technology is
coming and that some people will choose to upload, and that will be
the end of our species in its current biological form. People may have
different opinions as to whether this is a good or bad thing.

1. Landauer's estimate of human long term memory.
2. U.S. Dept. of Labor estimates $15,000 (1% of lifetime earnings)
average cost of replacing an employee.
3. Shannon estimates that written English has an entropy of 1 bit per character.
4. World GDP. https://www.google.com/search?rlz=1C1CHNU_enUS338US338&aq=f&ix=hca&sourceid=chrome&ie=UTF-8&q=world+gdp


-- Matt Mahoney, mattmahoneyfl@gmail.com

Matt Mahoney wrote:
> People don't want to work, and don't want to die. Solving the first
> problem requires building models of 7 billion human minds. A model is
> a function that takes sensory inputs and predicts your actions. If you
> program an avatar or robot to carry out those predictions for an
> individual in real time, then you have an upload. So if the
> philosophical difficulties can be overcome, then the second problem is
> solved as well.

=\

Damnit, I wish there was some way I could help you get past this single
idea that you keep harping on.

Once you have your first AGI, your Nth AGI is practically free.

Furthermore, an AGI that can participate in the economy to a significant
extent would change the fundamental nature of the economy and the
relationships within it so rapidly that the concepts you use to frame
your ideas would be upended and turned around.

Dreaming about 7-billion AGI helpers isn't interesting, useful,
productive or helpful. Designing a framework for integrating pieces of
an AGI will help; figuring out the pattern factorization problem will
help a lot! developing robotic platforms will help, hell even developing
simulatiors will help even though I'm majorly turned off to ALL
simulation thanks to the uploaders...)


> [Pro-uploading drivel removed]


E T F
N H E
D E D

Powers are not rights.

On Sun, Feb 5, 2012 at 11:02 PM, Alan Grimes <agrimes@speakeasy.net> wrote:
> [anti-uploading rant deleted...]

Well, I basically agree. It does not make sense to me to create a
digital copy of my mind and then shoot myself. I realize I am refusing
a procedure that would make me immortal. I can't help it if evolution
programmed me to fear the things that can kill me, as opposed to
fearing death.

But about cognitive architectures, the driving force behind AI
development is to make machines smart enough to do the work that we
have to pay people to do. That requires solving difficult AI problems
like language, vision, and robotics. More generally, it requires
predicting human behavior. Like how Google or Facebook predicts which
email messages, posts, web pages, and ads might interest me. Modeling
human behavior is essential for good communication. You use it every
time you say something. You predict what the other person knows and
doesn't know, so you can give them just the words that they need.
Machines are getting better at it, largely because we make so much
information about ourselves available to train them. Like how I am
posting this message to a public forum rather than sending it by
private, encrypted email.

People don't want to work, and don't want to die. Solving the first
problem requires building models of 7 billion human minds. A model is
a function that takes sensory inputs and predicts your actions. If you
program an avatar or robot to carry out those predictions for an
individual in real time, then you have an upload. So if the
philosophical difficulties can be overcome, then the second problem is
solved as well.

It probably makes more sense to ask what we will do, rather than what
we should do. If minds are computable, then human behavior is
predictable. When we have enough computing power to model human
civilization, we can just run a simulation to find out. But I think we
know enough about human behavior now to make some good guesses. To
wit:

- Humans fear dying, not death. Therefore, if uploading can be
presented in a way that hides the dying process, people will do it.
You go into a room and come out younger, stronger, healthier, and
smarter. If part of your brain were destroyed by a stroke, would you
object to a prosthetic that restored your lost mental capabilities?
Then why not your whole brain?

- Humans grieve the deaths of others. You might not want to upload,
but after you die your friends and relatives might want to resurrect
you.

- Humans behave altruistically toward other humans that are like
themselves. We extend this altruism toward "human-like" entities such
as dogs and cats. It is conceivable that we would do the same toward
robots that look and act like humans.

- If you believe that a digital copy of your mind is you, then you
will want to keep your property and your human rights, and others will
probably agree that you should.


-- Matt Mahoney, mattmahoneyfl@gmail.com

Sorry to be harsh,

(I skimmed the text), this sounds like general blah-blah.
Constructive AGI requires structure and dynamics, these are general
statements about superficial stuff.

A>Recent anatomical evidence indicates that there is indeed a heirarchical
organization, all interconnected with a top-level integration region which
probably produces the nucleus of your consciousness.

Sorry, but I don't think that's recent in neuroscience, and stated like
this it is theoretically quite obvious even without knowing a thing about
neuroscience for a child - a child knows that "consciousness" (actually
what one feels and thinks) has access to all modalities, all senses,
feelings - apparently, how could you talk about it otherwise.


-- Todor
http://research.twenkid.com

On Mon, Feb 6, 2012 at 6:02 AM, Alan Grimes <agrimes@speakeasy.net> wrote:

> om
>
> Hello, Because too many of you can only think when you're agitated, I
> intend to ruffle as many feathers I can in this post. Actually, my
> intention is to be so bold (as usual) as to be dismissed out of hand. Then,
> about 3 months from now, those of you who read this post will wake up one
> morning and realize that there might be a useful nugget somewhere in this
> pile of ascii characters. Six months from now it will become a hot topic of
> conversation, though my name will never be raised, nor my claims of
> authorship be acknowledged. A year or two after that someone who is either
> Jewish or living in Israel will canonize the ideas and claim all credit.
> With this seal of approval applied, it just might become the ontological
> framework for the entire field. I will, of course, still be a raving mad
> lunatic who doesn't want to be uploaded and not worthy of any attention or
> respect. But, by bringing ideas closer to money, hopefully something good,
> and useful to me, might come of it. I think Manfred Macx would approve.
>
> I would prefer to only talk about engineering here but because the common
> philosophical understanding is so badly wanting, I'm going to have to cover
> that too.
>
> om
>
> As a starting point, lets wheel out my old favorite whipping boy, the
> uploader. I have railed endlessly against uploaders. Having read half of
> Accelerando, I see no reason to pull any punches or spare any insults
> against uploaders, every other sentence was a slur against regular human
> beings. Among my numerous accusations is that they have allowed their
> fetish to dictate their entire mentality with regards to enhanced minds.
> Uploading gives you only two real "fiddly knobs" to turn. One knob,
> supposedly, adjusts the emulation speed, with the negative side-effect that
> the mind is no longer suitable for interfacing with the Real World (tm).
> The proposal is to create an alternate world that is accelerated by the
> same factor, which makes no sense whatsoever, as I'll explain. The only
> other fiddly knob you can operate on a black box is to copy it. The debate
> of whether to make further copies of a copy is thrown out with the debate
> over whether it makes sense at all to make the first copy.
>
> There has been some mumbling that once your copy is running, it would make
> a slightly different copy of itself, and then yield all of its CPU time to
> that copy, and do so endlessly until nothing whatsoever remains of the
> original. This is then pronounced the ultimate aim of all transhumanism
> without further examination. =(
>
> Peel back your eyelids, punks, and commence to examining!
>
> [urk, I started writing this a few days ago but I'm in a completely
> different mood now, my mind has more than half a dozen things demanding
> time-slices.]
>
> The uploaders sometimes claim that they want to upload to achieve some
> astronomically vast cognitive boost. I, myself, am looking at this from a
> different perspective, what is the most powerful kind of mind I can build
> and what is the best way to integrate it with my brain? I suspect the
> uploaders have some kind of blinders on that blur out paragraphs where I
> write about my own ideas because none has ever responded intelligently to
> my ideas.
>
> Lets set both of those perspectives aside and look at the basic problem.
> What concepts and techniques do we need to master in order to design
> advanced, high performance, cognitive architectures for our future selves?
> This touches on both the fields of philosophy and engineering. Engineering
> tries to tell us what the pieces are, philosophy tries to explain to us
> what it is we are actually accomplishing (or failing to accomplish).
>
> [another day passes, new mood appears]
>
> I had thought to slide directly into the discussion of how to organize a
> cognitive architecture but today I realized that motivation and "utility
> functions" are actually the central problem and must precede the discussion
> of the organization of the mind.
>
> Certain specific clueless numbskulls assert that a single-scalar utility
> function is "general" and therefore sufficient to steer a general
> intelligence, case closed, end of discussion, no further investigation
> necessary. There is one lesson that is fundamental to human nature, it is
> reflected in pretty much all art and literature. That is the idea of
> balance, it is seen in the idea of the Yin-Yang symbol. It is the reason
> that most of us don't fall into pathological states. (my own pathologies
> are well acknowledged).
>
> LAW OF INTELLIGENCE #1:
>
> The time average of the utility function of a healthy AI must always equal
> exactly 0.
>
> So what can you do based on a value that's usually zero? Practically
> nothing. So we need to move towards a vector model that has rows for
> motivators and de-motivators. Each channel will be tied to either some
> single value sensor or some basic pre-configured neural detection network.
> The output of each channel is configured to give the general intelligence
> the most information possible about what needs to be done, for example,
> hunger is represented by an ache in the stomach. Thirst, a dryness in the
> mouth (at least in me). From this, the higher mind has very little work to
> do to diagnose the cause of the problem and at least frame what kind of
> solution to look for in the environment.
>
> The two fiddly knobs this gives us are: How many channels are required?
> and how powerful should the motivation be? Clearly there must be at least
> two channels, one motivating channel, and one de-motivating channel. Which,
> over some reasonable period of time, must sum to zero.
> The strength of each channel should reflect on how important it is and the
> relative strength of its counterpart channels.
>
> The two basic pathologies that can emerge are reflected by either a
> negative time-average or a positive time-average. When the time average is
> negative, the AI is said to be "stressed" and is probably either
> ill-equipped to deal with its environment or is over-loaded for its
> capabilities. When the time average is positive, the AI is said to be
> obsessed and should be considered dangerous. Recognizing these pathologies
> should point to a satisfactory solution to the "paperclip problem"
>
> This brings us to our first major philosophical detour. Having established
> some general parameters for what SHOULD your motivators be. The first thing
> that must be accepted is that there is no "logical", inevitable or inherent
> answer. The uploaders will try to throttle you with their teleology but
> it's all bullshit. Life/being/existing is nothing more than an affront to
> non-existing. What you decide to do with your life is entirely your own
> concern. Now a human being seeking to become transhuman will have a set of
> motivations. They cannot be judged to be good or bad, only self-consistent
> or inconsistent.
>
> If anyone among us is to be allowed to pursue his own desires then we must
> all be so allowed. Furthermore, each person must be allowed to pursue his
> own joy regardless of public opinion. My own form of transhumanism is
> widely decried as being absurdly atavistic. I would argue that my design
> might actually be much more survivable and capable than the standard
> "post-human" offering. Once you shovel off all the bullshit, what it comes
> down to is the purely barbaric "might makes right" argument of "Since I can
> upload you by sending a swarm of nanites after you therefore I'm right and
> I get all your cookies too."
>
> Now we discover a very interesting dichotomy. A great engineer will try to
> find the most minimal (elegant) solution to a problem. Where all costs are
> minimized and the satisfaction of the original desire is maximized. Life,
> on the other hand, is all about creating problems, in the most obnoxious
> possible sense of the phrase. So we see an amazing tree of evolution where
> plants and animals wander into the most amazing engineering solutions for
> existing in an environment. The most noteworthy thing about this is that
> only a handful of species have much in the way of intelligence. In the bird
> family we find parrots, crows, and some other social birds. In the aquatic
> family we find dolphins, octopuses and maybe a few others, and in the lower
> mammal family we find prairie dogs and maybe a few other variants. Only
> higher primates appear to have a well developed episodic memory, and only
> humans have complex grammar and a vocabulary of more than a few hundred
> words.
>
> What should be learned from this is that nature, like engineers that I
> consider to be good, don't implement intelligence everywhere POSSIBLE but
> only where it is NECESSARY. This creates a rather fascinating problem that
> I have little hope of actually addressing in this excessively obnoxious
> e-mail. That is, as life-support functions are reduced to engineering
> problems and those engineering problems are solved, then the actual need
> for sentience pretty much evaporates. What will maintain a good rate of
> progress towards wiser and more powerful minds? That question contains
> ethical and moral assumptions that admittedly arise from me being a
> philosopher (someone who loves wisdom). The question that necessarily
> precedes the one I just stated is "What is the appropriate level of
> intelligence for an individual? Is there a law of diminishing returns?"
> That question, obviously has its own assumptions...
>
> To proceed into the technical discussion we will make all necessary
> assumptions necessary to state that our intention is to outline the design
> of a broadly capable mind sufficiently powerful to support whatever
> activities a given transhumanist requires to fight nihilism. There is
> little doubt that AI can be a stand-alone technology, perhaps the ultimate
> stand-alone technology. As a stand alone technology, AI is interesting
> enough. But the real focus of transhumanism should be in finding ways to
> "take on board" advanced capabilities that our technology is developing
> separately. The question of how to do that has received far less
> examination than it deserves. The only proposal that anyone seems to be
> working on is mind uploading by destructively scanning the brain. I believe
> that my position on that subject has already been made clear. I attack
> uploaders on many many fronts, the one that is most relevant to this
> discussion is what could be called "the Transcendent Idiot Paradox". I
> can't help but mention that I tend to think that the class of idiots
> encompasses all people who would willingly die so that a digital copy of
> themselves could be produced. =P I shouldn't have said that...
>
> The Transcendent Idiot Paradox states basically that an upload of an idiot
> is an uploaded idiot (Just as an uploaded lobster is an uploaded lobster).
> Uploading doesn't solve any problems while creating a whole new class of
> problems of getting the emulation software to work and keeping it well
> oiled. Furthermore, every attempt to modify the mind can only be through
> the facilities provided by the emulation software. Because uploaders are
> such a dim-witted and short-sighted bunch (ugh, now I've started a
> sub-thread...), I doubt that the first through hundredth iterations of the
> emulation software would allow much at all in the way of interesting
> modifications, the mind would have to be taken off line and effectively
> re-uploaded into a new matrix. Maybe the hundred and first would permit
> on-line modifications but then most of the work for any effort to modify
> your mind would fit into the category of "working around the conceptual
> fuckups of the people who designed the emulator". The underlying issue is
> that you're doing it backwards!
>
> What is really required is to realize that uploading is not really an end
> in itself. It's a means to something, I'm not sure what, which doesn't have
> any measurable effect on my existing emotional modules (though the
> procedure of uploading itself sure as hell does!).
>
> [days go by; got job; lost train of thought =0 ]
>
> Simulation has certain specific uses. It is a good means for playing
> around with concepts you either can't or aren't sure you want to implement
> in the real world. It can also be a communication medium for ideas that
> aren't easily expressed verbally. However, to create a simulation simply to
> "make work" for senses that were copied out of a previous body is grossly
> inefficient and, I would argue, extremely dangerous.
>
> The concept of "overclocking" comes from juvenile computer enthusiasts.
> You will find almost no mention of "clock speeds" (clock cycles: yes, clock
> speeds, no) in actual computer science textbooks. The two concepts you DO
> find in the literature are "latency" and "throughput". Neither term should
> require any further explanation here. In uploading, you can't improve
> latency without distorting the perception of time. In an advanced cognitive
> architecture, there is no reason latency can't be reduced to the
> theoretical limits while maintaining a constant perception of time. You
> would simply be able to react thoughtfully to events on much shorter
> timescales and, probably be able to perceive each instant in greater detail
> without it, necessarily, becoming boring. To accomplish that you would
> probably have to increase the storage space for your sensory and short term
> memory by several orders of magnitude but the basic parameters can remain
> the same. You would keep your basic mental clock but probably need to add
> in a second "high precision" clock.
>
> Reducing latencies in this manner will improve throughput. But to make
> major advancements you will need to find a way to add parallelism. Stross
> spews some hand-wavey, undefined, technobabble about spawning "ghosts" and
> then, later, re-integrating the memories. As far as I can tell, these ideas
> are nonsense because there is no clear way to re-integrate a forked neural
> network as he suggests. What you really need to do is design a new
> cognitive architecture that is actually can handle concurrent streams of
> consciousness. Furthermore, fixed matrix neural network models are surely
> far too limiting for long-term use. What you really need is a highly
> concurrent distributed pattern processing engine that can be used in
> well-organized networks to solve specific cybernetic problems. While a
> constructive proof that such an algorithm exists, I consider the
> probability that it does extremely high and I think it would show such huge
> advantages over neural networks that only uploaders would ever consider
> using them again.
>
> The best way to understand the brain appears to be as a cybernetic system
> consisting of a number of input channels for several modalities and a
> number of output channels for one or two modalities, no networking, and a
> fairly complex topology of fifty or so high-level pattern processors
> depending on a hundred or so specialized modules that compute time delays,
> solve differential equations, and basically do all the fixed-function
> processing for the various modalities. Recent anatomical evidence indicates
> that there is indeed a heirarchical organization, all interconnected with a
> top-level integration region which probably produces the nucleus of your
> consciousness.
>
> The current fledgling attempts at neural interfacing focus on adding new
> modalities that emulate input-output channels. These might have their
> therapeutic uses but they're absolutely insufficient for any useful
> enhancement. What you need is a completely de-novo AI of the most
> sophisticated type you can develop, the core entity would have many network
> interfaces but few modalities. Your neural interface would need to be
> sufficiently capable to mind-meld yourself to it. From this core, you would
> create distributed instances of highly specialized minds that would
> probably bear no resemblance to your core mind but would be exceptionally
> capable at their assigned function. These would share a distributed concept
> and meta-ontology framework so that all your worker instances would be
> updated with the same information at network speed, There would not be any
> concept of "re-integration" and no special effort would be required to
> accomplish it. These worker threads would function as a parallel facet of
> your mind indefinitely or until they become redundant.
>
> In Stross' deeply flawed vision, I would maintain my physical existence
> and mutate it as I see fit while, simultaneously, with a contiguous mind,
> spawn remote network nodes, each fully sentient and well adapted to its
> "environment" (gag). These nodes would be placed in strategic locations,
> each being both complete and fully sentient and part of the same
> mind-network. The latency at which remote nodes learn things from local
> nodes is not really important because each node has a high degree of
> autonomy and would never have to wait for a global decision. Nevertheless
> it would be expected to obey commands issued from its network. I would
> strongly doubt that more than a few gigabytes of data is necessary to
> establish and operate such a node. There would be no reason to attempt to
> interface with the so called "universes" (barf). Instead, the remote nodes
> would be charged with trying to hack the lowest level of accessible
> reality, always looking for a key to the next rung down on the
> "virtualization stack". It would also lazily monitor coms traffic and might
> occasionally make obnoxious comments to and about the other "inhabitants"
> of the computronium.
>
> What this means for wannabe AGI developers is that a mature AGI technology
> will require the ability to dynamically re-configure itself no the fly (and
> be able to automatically find a near-optimal internal topology) and
> maintain distributed nodes across slow and lossy channels. The critical
> research areas are the underpinnings of the cybernetic loop, a
> qualitatively superior pattern processing algorithm, and a sophisticated
> and well balanced set of motivational sub-routines.
>
> Okay, there; I've done-gone wear out a cheesy Microsoft Natural
> keyborad... (they're really built like crap and would be nothing without
> their well-designed layout.)
>
> Where's that blasted send button? (Yes, I'm aware that I've been rude and
> insulting; deal with it!)
>
> --
> E T F
> N H E
> D E D
>
> Powers are not rights.
>
>
>
> ------------------------------**-------------
> AGI
> Archives: https://www.listbox.com/**member/archive/303/=now<https://www.listbox.com/member/archive/303/=now&gt;
> RSS Feed: https://www.listbox.com/**member/archive/rss/303/**
> 21253656-fab278b9<https://www.listbox.com/member/archive/rss/303/21253656-fab278b9&gt;
> Modify Your Subscription: https://www.listbox.com/**
> member/?&id_**secret=21253656-20e37305<https://www.listbox.com/member/?&&gt;
> Powered by Listbox: http://www.listbox.com
>



--
-- Todor "Tosh" Arnaudov
http://research.twenkid.com
http://artificial-mind.blogspot.com

om

Hello, Because too many of you can only think when you're agitated, I
intend to ruffle as many feathers I can in this post. Actually, my
intention is to be so bold (as usual) as to be dismissed out of hand.
Then, about 3 months from now, those of you who read this post will wake
up one morning and realize that there might be a useful nugget somewhere
in this pile of ascii characters. Six months from now it will become a
hot topic of conversation, though my name will never be raised, nor my
claims of authorship be acknowledged. A year or two after that someone
who is either Jewish or living in Israel will canonize the ideas and
claim all credit. With this seal of approval applied, it just might
become the ontological framework for the entire field. I will, of
course, still be a raving mad lunatic who doesn't want to be uploaded
and not worthy of any attention or respect. But, by bringing ideas
closer to money, hopefully something good, and useful to me, might come
of it. I think Manfred Macx would approve.

I would prefer to only talk about engineering here but because the
common philosophical understanding is so badly wanting, I'm going to
have to cover that too.

om

As a starting point, lets wheel out my old favorite whipping boy, the
uploader. I have railed endlessly against uploaders. Having read half of
Accelerando, I see no reason to pull any punches or spare any insults
against uploaders, every other sentence was a slur against regular human
beings. Among my numerous accusations is that they have allowed their
fetish to dictate their entire mentality with regards to enhanced minds.
Uploading gives you only two real "fiddly knobs" to turn. One knob,
supposedly, adjusts the emulation speed, with the negative side-effect
that the mind is no longer suitable for interfacing with the Real World
(tm). The proposal is to create an alternate world that is accelerated
by the same factor, which makes no sense whatsoever, as I'll explain.
The only other fiddly knob you can operate on a black box is to copy it.
The debate of whether to make further copies of a copy is thrown out
with the debate over whether it makes sense at all to make the first copy.

There has been some mumbling that once your copy is running, it would
make a slightly different copy of itself, and then yield all of its CPU
time to that copy, and do so endlessly until nothing whatsoever remains
of the original. This is then pronounced the ultimate aim of all
transhumanism without further examination. =(

Peel back your eyelids, punks, and commence to examining!

[urk, I started writing this a few days ago but I'm in a completely
different mood now, my mind has more than half a dozen things demanding
time-slices.]

The uploaders sometimes claim that they want to upload to achieve some
astronomically vast cognitive boost. I, myself, am looking at this from
a different perspective, what is the most powerful kind of mind I can
build and what is the best way to integrate it with my brain? I suspect
the uploaders have some kind of blinders on that blur out paragraphs
where I write about my own ideas because none has ever responded
intelligently to my ideas.

Lets set both of those perspectives aside and look at the basic problem.
What concepts and techniques do we need to master in order to design
advanced, high performance, cognitive architectures for our future
selves? This touches on both the fields of philosophy and engineering.
Engineering tries to tell us what the pieces are, philosophy tries to
explain to us what it is we are actually accomplishing (or failing to
accomplish).

[another day passes, new mood appears]

I had thought to slide directly into the discussion of how to organize a
cognitive architecture but today I realized that motivation and "utility
functions" are actually the central problem and must precede the
discussion of the organization of the mind.

Certain specific clueless numbskulls assert that a single-scalar utility
function is "general" and therefore sufficient to steer a general
intelligence, case closed, end of discussion, no further investigation
necessary. There is one lesson that is fundamental to human nature, it
is reflected in pretty much all art and literature. That is the idea of
balance, it is seen in the idea of the Yin-Yang symbol. It is the reason
that most of us don't fall into pathological states. (my own pathologies
are well acknowledged).

LAW OF INTELLIGENCE #1:

The time average of the utility function of a healthy AI must always
equal exactly 0.

So what can you do based on a value that's usually zero? Practically
nothing. So we need to move towards a vector model that has rows for
motivators and de-motivators. Each channel will be tied to either some
single value sensor or some basic pre-configured neural detection
network. The output of each channel is configured to give the general
intelligence the most information possible about what needs to be done,
for example, hunger is represented by an ache in the stomach. Thirst, a
dryness in the mouth (at least in me). From this, the higher mind has
very little work to do to diagnose the cause of the problem and at least
frame what kind of solution to look for in the environment.

The two fiddly knobs this gives us are: How many channels are required?
and how powerful should the motivation be? Clearly there must be at
least two channels, one motivating channel, and one de-motivating
channel. Which, over some reasonable period of time, must sum to zero.
The strength of each channel should reflect on how important it is and
the relative strength of its counterpart channels.

The two basic pathologies that can emerge are reflected by either a
negative time-average or a positive time-average. When the time average
is negative, the AI is said to be "stressed" and is probably either
ill-equipped to deal with its environment or is over-loaded for its
capabilities. When the time average is positive, the AI is said to be
obsessed and should be considered dangerous. Recognizing these
pathologies should point to a satisfactory solution to the "paperclip
problem"

This brings us to our first major philosophical detour. Having
established some general parameters for what SHOULD your motivators be.
The first thing that must be accepted is that there is no "logical",
inevitable or inherent answer. The uploaders will try to throttle you
with their teleology but it's all bullshit. Life/being/existing is
nothing more than an affront to non-existing. What you decide to do with
your life is entirely your own concern. Now a human being seeking to
become transhuman will have a set of motivations. They cannot be judged
to be good or bad, only self-consistent or inconsistent.

If anyone among us is to be allowed to pursue his own desires then we
must all be so allowed. Furthermore, each person must be allowed to
pursue his own joy regardless of public opinion. My own form of
transhumanism is widely decried as being absurdly atavistic. I would
argue that my design might actually be much more survivable and capable
than the standard "post-human" offering. Once you shovel off all the
bullshit, what it comes down to is the purely barbaric "might makes
right" argument of "Since I can upload you by sending a swarm of nanites
after you therefore I'm right and I get all your cookies too."

Now we discover a very interesting dichotomy. A great engineer will try
to find the most minimal (elegant) solution to a problem. Where all
costs are minimized and the satisfaction of the original desire is
maximized. Life, on the other hand, is all about creating problems, in
the most obnoxious possible sense of the phrase. So we see an amazing
tree of evolution where plants and animals wander into the most amazing
engineering solutions for existing in an environment. The most
noteworthy thing about this is that only a handful of species have much
in the way of intelligence. In the bird family we find parrots, crows,
and some other social birds. In the aquatic family we find dolphins,
octopuses and maybe a few others, and in the lower mammal family we find
prairie dogs and maybe a few other variants. Only higher primates appear
to have a well developed episodic memory, and only humans have complex
grammar and a vocabulary of more than a few hundred words.

What should be learned from this is that nature, like engineers that I
consider to be good, don't implement intelligence everywhere POSSIBLE
but only where it is NECESSARY. This creates a rather fascinating
problem that I have little hope of actually addressing in this
excessively obnoxious e-mail. That is, as life-support functions are
reduced to engineering problems and those engineering problems are
solved, then the actual need for sentience pretty much evaporates. What
will maintain a good rate of progress towards wiser and more powerful
minds? That question contains ethical and moral assumptions that
admittedly arise from me being a philosopher (someone who loves wisdom).
The question that necessarily precedes the one I just stated is "What is
the appropriate level of intelligence for an individual? Is there a law
of diminishing returns?" That question, obviously has its own assumptions...

To proceed into the technical discussion we will make all necessary
assumptions necessary to state that our intention is to outline the
design of a broadly capable mind sufficiently powerful to support
whatever activities a given transhumanist requires to fight nihilism.
There is little doubt that AI can be a stand-alone technology, perhaps
the ultimate stand-alone technology. As a stand alone technology, AI is
interesting enough. But the real focus of transhumanism should be in
finding ways to "take on board" advanced capabilities that our
technology is developing separately. The question of how to do that has
received far less examination than it deserves. The only proposal that
anyone seems to be working on is mind uploading by destructively
scanning the brain. I believe that my position on that subject has
already been made clear. I attack uploaders on many many fronts, the one
that is most relevant to this discussion is what could be called "the
Transcendent Idiot Paradox". I can't help but mention that I tend to
think that the class of idiots encompasses all people who would
willingly die so that a digital copy of themselves could be produced. =P
I shouldn't have said that...

The Transcendent Idiot Paradox states basically that an upload of an
idiot is an uploaded idiot (Just as an uploaded lobster is an uploaded
lobster). Uploading doesn't solve any problems while creating a whole
new class of problems of getting the emulation software to work and
keeping it well oiled. Furthermore, every attempt to modify the mind can
only be through the facilities provided by the emulation software.
Because uploaders are such a dim-witted and short-sighted bunch (ugh,
now I've started a sub-thread...), I doubt that the first through
hundredth iterations of the emulation software would allow much at all
in the way of interesting modifications, the mind would have to be taken
off line and effectively re-uploaded into a new matrix. Maybe the
hundred and first would permit on-line modifications but then most of
the work for any effort to modify your mind would fit into the category
of "working around the conceptual fuckups of the people who designed the
emulator". The underlying issue is that you're doing it backwards!

What is really required is to realize that uploading is not really an
end in itself. It's a means to something, I'm not sure what, which
doesn't have any measurable effect on my existing emotional modules
(though the procedure of uploading itself sure as hell does!).

[days go by; got job; lost train of thought =0 ]

Simulation has certain specific uses. It is a good means for playing
around with concepts you either can't or aren't sure you want to
implement in the real world. It can also be a communication medium for
ideas that aren't easily expressed verbally. However, to create a
simulation simply to "make work" for senses that were copied out of a
previous body is grossly inefficient and, I would argue, extremely
dangerous.

The concept of "overclocking" comes from juvenile computer enthusiasts.
You will find almost no mention of "clock speeds" (clock cycles: yes,
clock speeds, no) in actual computer science textbooks. The two concepts
you DO find in the literature are "latency" and "throughput". Neither
term should require any further explanation here. In uploading, you
can't improve latency without distorting the perception of time. In an
advanced cognitive architecture, there is no reason latency can't be
reduced to the theoretical limits while maintaining a constant
perception of time. You would simply be able to react thoughtfully to
events on much shorter timescales and, probably be able to perceive each
instant in greater detail without it, necessarily, becoming boring. To
accomplish that you would probably have to increase the storage space
for your sensory and short term memory by several orders of magnitude
but the basic parameters can remain the same. You would keep your basic
mental clock but probably need to add in a second "high precision" clock.

Reducing latencies in this manner will improve throughput. But to make
major advancements you will need to find a way to add parallelism.
Stross spews some hand-wavey, undefined, technobabble about spawning
"ghosts" and then, later, re-integrating the memories. As far as I can
tell, these ideas are nonsense because there is no clear way to
re-integrate a forked neural network as he suggests. What you really
need to do is design a new cognitive architecture that is actually can
handle concurrent streams of consciousness. Furthermore, fixed matrix
neural network models are surely far too limiting for long-term use.
What you really need is a highly concurrent distributed pattern
processing engine that can be used in well-organized networks to solve
specific cybernetic problems. While a constructive proof that such an
algorithm exists, I consider the probability that it does extremely high
and I think it would show such huge advantages over neural networks that
only uploaders would ever consider using them again.

The best way to understand the brain appears to be as a cybernetic
system consisting of a number of input channels for several modalities
and a number of output channels for one or two modalities, no
networking, and a fairly complex topology of fifty or so high-level
pattern processors depending on a hundred or so specialized modules that
compute time delays, solve differential equations, and basically do all
the fixed-function processing for the various modalities. Recent
anatomical evidence indicates that there is indeed a heirarchical
organization, all interconnected with a top-level integration region
which probably produces the nucleus of your consciousness.

The current fledgling attempts at neural interfacing focus on adding new
modalities that emulate input-output channels. These might have their
therapeutic uses but they're absolutely insufficient for any useful
enhancement. What you need is a completely de-novo AI of the most
sophisticated type you can develop, the core entity would have many
network interfaces but few modalities. Your neural interface would need
to be sufficiently capable to mind-meld yourself to it. From this core,
you would create distributed instances of highly specialized minds that
would probably bear no resemblance to your core mind but would be
exceptionally capable at their assigned function. These would share a
distributed concept and meta-ontology framework so that all your worker
instances would be updated with the same information at network speed,
There would not be any concept of "re-integration" and no special effort
would be required to accomplish it. These worker threads would function
as a parallel facet of your mind indefinitely or until they become
redundant.

In Stross' deeply flawed vision, I would maintain my physical existence
and mutate it as I see fit while, simultaneously, with a contiguous
mind, spawn remote network nodes, each fully sentient and well adapted
to its "environment" (gag). These nodes would be placed in strategic
locations, each being both complete and fully sentient and part of the
same mind-network. The latency at which remote nodes learn things from
local nodes is not really important because each node has a high degree
of autonomy and would never have to wait for a global decision.
Nevertheless it would be expected to obey commands issued from its
network. I would strongly doubt that more than a few gigabytes of data
is necessary to establish and operate such a node. There would be no
reason to attempt to interface with the so called "universes" (barf).
Instead, the remote nodes would be charged with trying to hack the
lowest level of accessible reality, always looking for a key to the next
rung down on the "virtualization stack". It would also lazily monitor
coms traffic and might occasionally make obnoxious comments to and about
the other "inhabitants" of the computronium.

What this means for wannabe AGI developers is that a mature AGI
technology will require the ability to dynamically re-configure itself
no the fly (and be able to automatically find a near-optimal internal
topology) and maintain distributed nodes across slow and lossy channels.
The critical research areas are the underpinnings of the cybernetic
loop, a qualitatively superior pattern processing algorithm, and a
sophisticated and well balanced set of motivational sub-routines.

Okay, there; I've done-gone wear out a cheesy Microsoft Natural
keyborad... (they're really built like crap and would be nothing without
their well-designed layout.)

Where's that blasted send button? (Yes, I'm aware that I've been rude
and insulting; deal with it!)

--
E T F
N H E
D E D

Powers are not rights.