SIDEBAR: Computer Technology in the late 21C (Orichalcum U..

Artificial intelligence is my world is a more complicated matter than
is usually treated.

Implementing an AI requires a suitable computer as a base,
conventional computers are not suitible. All artificially intelligent
computers are neural-net systems (also known as ‘synaptic’ systems),
but most neural-net systems are also not adequate, a computer intended
to serve as the basis of an AI requires several additional subsystems
and features, none of which are cheap. Generally speaking, a minimum-
suitable potential AI computer costs _twenty times_ as much as an
equivalent standard ‘synaptic’ computer system.

Along with the computer, an ‘artificial intelligence matrix’ is
required. This isn’t just a piece of software, it’s a hugely complex
and intricately interconnected set of programs and data, even a simple
AI matrix will involve _hundreds_ of different programs cross-
connected to each other and very sophisticated databases.

Any computer capable of serving as the base of an AI will be a unique
machine. Even two AI-capable computers of the same design and
construction process, even the same production run, will not be
identical. These systems incorporate systems sensitive to quantum-
level uncertainty, it’s _impossible_ for two separate AI-computers to
be identical.

This is important, because each AI matrix must be prepared to match
the machine on which it is to run, to a high degree of precision. The
AI shaper must have extensive knowledge of the machine to be used as a
base in order to successfully prepare a viable AI matrix. This also
means that an AI _can not_ be ‘copied’ from one machine to another, it
simply won’t work. While one could certainly copy the data and system-
state of a given AI, trying to ‘run’ that on any other machine will
result in a crashed system, usually rendering the computer hardware
useless for future use in the process.

Furthermore, an AI is based on an _ongoing_ process of subtle
operations in hardware and software, even reimprinting the copy of the
system-state on the _same_ machine later also won’t work, the effect
is likely to be the same as trying to imprint a system-state copy on a
different machine. This means, among other things, that an AI can not
make ‘copies’ of itself to reboot itself later. Even for an AI, death
is permanent. [1]

An AI can survive partial collapse of its matrix, just as a
Homosentient may survive amnesia or brain damage, the number of ways
an AI can ‘go wrong’ rival the ways a Homosentient can do so, in
fact. However, like a Homosentient, there is a point past which a
failure becomes irreversible. An AI can take certain precautions
against this, however, that are more difficult if not impossible for a
Homosentient (or an animal intelligence, for that mater).

A high-resolution copy of the total system-state of an AI can
_sometimes_ be used to restore itself, _if_ the AI has not collapsed
completely. Partial or total amnesia (i.e. erasure of some databases)
can be restored by copying information just as with any other
computer, unless the data was involved in the self-sustaining, ongoing
dynamic process that is the AI itself.

If the original AI matrix, the seed out of which the AI grew, is still
stored within the machine and the machine is not physically damaged,
the AI can sometimes _regrow_. This is not always possible, but it
_can_ happen under the right circumstances. There may be some
personality changes, but it will still be recognizably the same
entity.

MORE LATER.


[1] No, braintaping doesn't work for people, either, for somewhat
similar reasons.

Johnny1a wrote:
>
>
> Any computer capable of serving as the base of an AI will be a unique
> machine. Even two AI-capable computers of the same design and
> construction process, even the same production run, will not be
> identical. These systems incorporate systems sensitive to quantum-
> level uncertainty, it’s _impossible_ for two separate AI-computers to
> be identical.
>

With this statement you've thrown out all standards of mass
manufacturing. Sorry, but within the limits of tolerance specified as a
part of the design, the units would be identical. And our capabilities
today of nanometer technology make the projected manufacturing
specification capabilities to be on a sub-atomic scale. (We're ALREADY
doing that in semiconductors, and have been since at least 1984!)


> This is important, because each AI matrix must be prepared to match
> the machine on which it is to run, to a high degree of precision. The
> AI shaper must have extensive knowledge of the machine to be used as a
> base in order to successfully prepare a viable AI matrix. This also
> means that an AI _can not_ be ‘copied’ from one machine to another, it
> simply won’t work. While one could certainly copy the data and system-
> state of a given AI, trying to ‘run’ that on any other machine will
> result in a crashed system, usually rendering the computer hardware
> useless for future use in the process.

While I would say that it's possible (but unlikely) that Revision A,
Revision B, and the original would have differences large enough that
software would require revision, I would find that a very strange
situation in today's manufacturing, much less in the supposedly improved
versions of the next 100 years.


>
> Furthermore, an AI is based on an _ongoing_ process of subtle
> operations in hardware and software, even reimprinting the copy of the
> system-state on the _same_ machine later also won’t work, the effect
> is likely to be the same as trying to imprint a system-state copy on a
> different machine. This means, among other things, that an AI can not
> make ‘copies’ of itself to reboot itself later. Even for an AI, death
> is permanent. [1]
>

Adding peripherals to a computer system frequently requires adding
drivers designed specifically for that item, even today. And if/when the
system crashes, it can be a little more complicated to
re-boot/re-initialize. But, replacing parts that either wear out, or are
improved, or are just a new capability, are standard aspects of ALL
computer systems. Including upgrading the OS.

Backups are an absolute requirement.

YMMV for YOUR world concept.
Here's to Better Gaming!
Regards,
Roger

On Jul 16, 7:51 am, Roger Connor wrote:
> Johnny1a wrote:
>
> > Any computer capable of serving as the base of an AI will be a unique
> > machine.  Even two AI-capable computers of the same design and
> > construction process, even the same production run, will not be
> > identical.  These systems incorporate systems sensitive to quantum-
> > level uncertainty, it’s _impossible_ for two separate AI-computers to
> > be identical.
>
> With this statement you've thrown out all standards of mass
> manufacturing. Sorry, but within the limits of tolerance specified as a
> part of the design, the units would be identical. And our capabilities
> today of nanometer technology make the projected manufacturing
> specification capabilities to be on a sub-atomic scale. (We're ALREADY
> doing that in semiconductors, and have been since at least 1984!)[/quote]

No, we aren't. The key words of your statement are 'within the limits
of tolerance'. The computer chips coming out of manufacturing plants
today are _not_ identifical at a quantum level, or even at higher
scales, they aren't even close to identical. They are similar
_enough_ to behave in a predictable way under a highly refined set of
circumstances.

I assume by the words 'sub-atomic' you mean molecular, since we can't
engineer nuclei in any meaningful way at all, other than to split or
fuse them for energy release.



>
> > This is important, because each AI matrix must be prepared to match
> > the machine on which it is to run, to a high degree of precision.  The
> > AI shaper must have extensive knowledge of the machine to be used as a
> > base in order to successfully prepare a viable AI matrix.  This also
> > means that an AI _can not_ be ‘copied’ from one machine to another, it
> > simply won’t work.  While one could certainly copy the data and system-
> > state of a given AI, trying to ‘run’ that on any other machine will
> > result in a crashed system, usually rendering the computer hardware
> > useless for future use in the process.

>
>
>
> > Furthermore, an AI is based on an _ongoing_ process of subtle
> > operations in hardware and software, even reimprinting the copy of the
> > system-state on the _same_ machine later also won’t work, the effect
> > is likely to be the same as trying to imprint a system-state copy on a
> > different machine.  This means, among other things, that an AI can not
> > make ‘copies’ of itself to reboot itself later.  Even for an AI, death
> > is permanent. [1]
>
> Adding peripherals to a computer system frequently requires adding
> drivers designed specifically for that item, even today. And if/when the
> system crashes, it can be a little more complicated to
> re-boot/re-initialize. But, replacing parts that either wear out, or are
> improved, or are just a new capability, are standard aspects of ALL
> computer systems. Including upgrading the OS.

Today's computer systems have so little in common with the way a
living brain works, or a real AI would have to work, that they aren't
even reasonable comparisons. Even a very, very simple animal brain is
enormously more complex, subtle, and dynamic than _any_ computer OS or
hardware ever developed as of 2009. Today, we don't even understand
in any detail how brains work, so comparing intelligent machines to
today's computers is pointless.

LATER.

Preparing AI matrices, and shaping and guiding the resulting AI, is a
professional specialty commanding high pay and requiring both a great
deal of training and professional knowledge and a certain amount of
innate knack, it is an art rather than a science. These individuals
are called 'shapers', an 'AI shaper' is an individual who creates and
trains and maintains AI systems.

In GURPS terms, AI shapers have the professional skill AI Shaping
(mental/VH/no default), prerequisites Computer Programming at 15 or
better, Mathematics at 15 or better, Electronics (Computer Tech) at
15, Electronics (AI Tech) at 15. Electronics-Computer Tech is a
Mental/Hard skill. Electronics-AI Tech is the specialty of the
special machines used as AI bases, and requires Electronics-Computer
Tech at 15 as a prerequisite.

Oddly enough, possession of the Advantages Animal Empathy or Empathy
seems to help with this skill, any AI shaper possessing either gets a
+1 on all skill rolls in his specialty (NOT cumulative). Even the
shapers debate why this is. The most common theory among them is that
the innate knack for understanding other minds that lets a person
interact especially well with animals lends itself to interacting with
_artificial_ minds as well.

As a general thing, the 'IQ' of an AI is limited by the Complexity of
the base, the Tech Level of the background, and the skill of the shaper
(s) who create it. In general the upper limit is TL+1. That is, the
most 'intelligent' AI usually possible at TL10 would have an IQ of
11. Even if higher Complexity machines are available, it isn't
usually possible to make full use of it. A shaper or team of shapers
who are unusually skilled and successful can sometimes (critical
success) produce AIs with an IQ of TL+2, but this is a rare thing. [1]

At any given TL, a lower-IQ AI is _much_ cheaper and easier to
implement.

Normally, along with the TL limit, an AI is limited to an IQ of
Complexity + 1, or less. (Thus the potential IQ of an AI is limited
by _both_ TL and Complexity of the available machine, either can limit
the potential IQ.)

In game terms, an AI shaper seeking to implement an AI must roll
against his/her skill at AI Shaping, with modifiers for
circumstances. Most shapers require at least one _year_ of game time
per point of IQ in the proposed new AI in a new computer base, that
is, a system in which everything is new, the model, the seed, all of
it. If the particular model of computer base is familiar, then the
shaper can go faster, since s/he merely has to adapt to the particular
quirks of the specific unit, in that event reduce time to 1 _month_
per point of proposed IQ If the proposed AI is similar to other AIs
(same model of computer, creating a similar personality and nature),
then time can be further reduced to 1 _week_ per point of IQ in the
proposed new AI. (GM's discretion whether a new AI is different
enough to require the full month, given the same model of computer).

If a team of shapers is available, the combined skill of the team is
added up and the square root of that is added to the skill of the
_most skilled member_ of the team, and used as the collective skill
level. The bonus for empathy (of either sort) is added if at least ¼
of the team has either Advantage.

When the skill roll is made, roll 3d as usual with the following
modfiers:

Shaper(s) used extra time to plan and prepare and work (double time or
more) +1 for double, +2 for triple time
Shaper(s) have used this model of computer-base previously: +3
Shaper(s) have 'shaped' this _kind_ of AI before: +3
Shaper(s) are implementing an AI with an IQ less than the Complexity
of the machine: +2 for each Complexity-1.
Shaper(s) are implementing an AI with an IQ less than the background
TL: +1 per level below the background TL number.
Shaper(s) used less than recommended time: -2 for _each_ year/month/
week less than recommended (see conditions above for which time unit).
Shaper(s) are shaping a new _kind_ of AI: -2
Shaper(s) are using an unfamiliar model of computer-base: -2
Shaper(s) are using an untested computer base or are shaping a kind of
AI never done before: -2 for each (cumulative)
Shaper(s) are trying to create an IQ above the base TL: -2 for IQ=TL
+1, -4 for IQ=TL+2 [2]
Shaper(s) are trying to implement an AI with IQ at Complexity: -2
Shaper(s) are trying to implement an AI with IQ at Complexity + 1: -5

Of course the GM should use his/her own judgement to assign reasonable
penalties and bonuses for any specific situation.

MORE LATER.

[1] Exceptions can exist.

[2] Attempts to implement AIs 'smarter' than TL+2 require _very_
exceptional circumstances and skills. Normally such skill levels
would not be in the hands of most PCs, certainly an Unusual Background
would be necessary.

LATER.

EXAMPLE 1:

Neil is an AI shaper with the AI-Shaper skill at 14. He's very good
at it, lot's of experience, and wants to implement a AI for a
megacomputer owned by the U.S. Army. The megacomp is AI-capable,
Complexity 9. The background TL is 10.

The proposed AI is to have an IQ of 10 (human-level). There is plenty
of funding and time. Neil has worked with this 'model' of system
before, but shaped AI's somewhat liked the proposed system. Normally,
since the proposed AI will have IQ 10, he would need to spend ten
weeks of preparatory time (familiar AI, familiar computer). He
decides to spend twice that, getting a +1 bonus his skill. Since
computer and matrix are both 'familiar' (he's done both before), he
gets a +3 for each, bringing the total bonus up to +7.

However...he is trying to implement an AI with an IQ one level greater
than the Complexity of the machine, which is a -5 to skill. So he
goes into this with a +2 bonus total. He rolls a 15, but with the +2
bonus he succeeds and the new AI is born and has the projected IQ of
10. His job isn't done by any means, but he's succeeded that far.

What if he'd failed?

Replaying Example 1, this time Neil rolls an 18, he blew it. The new
AI fails to emerge, but the computer can still be used, Neil can try
again. On a critical failure, the computer base is ruined and must be
replaced. (Which with an AI-capable megacomputer is a very expensive
proposition!)

EXAMPLE 2: Jane has AI-Shaping at 12, she's trying to implement a
simple AI on an AI-capable microframe at TL10 (Complexity 6). The
projected AI will have an IQ of 4 and is a familiar implementation on
a familiar system.

Double familiarity means Jane needs 4 weeks of planning and
preparation and work, which she does. She gets a +6 for double
familiarity, and a +6(!) because the AI will have in IQ 6 levels below
the background TL. On top of that Jane gets another +2 for each
difference between the Complexity of the microframe (6) and the IQ of
the AI (4). So there's another +4.

So Jane has a +16 bonus to skill for this implementation.

EXAMPLE 3: Rerun Example 1 again, but this time Neil gets a critical
success. As a result, his new AI gets a +1 on IQ, the resulting
machine has IQ 11.

EXAMPLE 4: Neil from Ex 1 is trying to implement a new kind of AI on
a megacomp of a model he's never used before. The planned AI will
have an IQ of 10. Base TL is 10 and the machine is Complexity 9.

Since he's not familiar with either the machine or the matrix, he
properly needs ten _years_ to prepare it. It's a major project, he
and his team spend the time as needed. But they are working with
hefty penalties: -2 for unfamiliar machine, -2 for unfamiliar matrix,
an additional -2 since nobody has tried ot make this kind of AI b
efore (or if they have, the team has no access to their work), and -5
for an IQ one above the Complexity. This gives a total penalty of
-11.

As mentioned above, the shaper's job is not done when the AI is
brought on-line. It still has to be guided, trained, and 'shaped', by
the shaper(s). This takes a period of time dependent on the IQ of the
machine and a few other factors.

For each point of IQ of a new born AI, the shaper(s) must spend _at
least_ one full month to train and prepare it. This time may be
increased by a few factors.

A new _kind_ of AI: triple the time or more (GM discretion)
A new _kind_ of machine: triple time or more. (GM discretion)
If the AI was prepared in less than the proper amount of time, double
the post-awakening time necessary.

All are cumulative.

For each _month_ of training time, the shaper(s) must again roll vs.
skill, using the same general rules as above. On a success, the AI
'matures' and becomes closer to the planned ideal. For each failure,
the AI has something go somewhat wrong (GM's discretion), a critical
failure leads to an irreversible AI crash, the AI is gone and the
shaper(s) must roll against skill _again_, on a failure the machine is
ruined and must be replaced.

Note that a _different_ shaper or team of shapers can do ths post-
awakening training...but at -5 to skill if they are fully familiar
with the project, or worse penalties if not.

EXAMPLE 5: The AI from Example 1 and Neil are our subject. The AI
has IQ 10, and it was a familiar system, familiar type of AI, and
plenty of time, so 10 months of post-Awakening training and work are
necessary. Every month of game time requires Neil roll vs. his AI-
shaping skill, with failure producing a problem or glitch in the
implementation, and critical failure destroying the AI and maybe the
machine.

The sorts of thing that can 'go wrong' with an AI short of crash can
vary widely. Exactly what they are depends heavily on the nature of
th AI, which means that the GM will have to use his/her own judgement
about what can go bad. Note that most AI's lack any of the neuro-
psychological 'wiring' necessary for many psychological problems, the
'murderous AI' is mostly a myth, since they lack those components of
the brain that make Homosentients so...fractious, and all AI's
incorporate a variety of highly-prioritized restrictive programs,
including some hardwired into the hardware.. They _can_, though,
develop any number of irritating or troublesome personality glitches
or problems, and they can even develop mental blocks and some sorts of
phobias under certain conditions.

A few suggestions for what _can_ happen:

AI is 'bitter', carries out orders but with resistance and 'backtalk'.
AI is 'shaky', easily confused, does not readil focus.
AI is prone to enter 'catatonic' states at random or if overloaded or
startled.
AI works normally _most_ of the time, but enters phobic or other
troublesome state under pressure.

If an AI goes 'wrong' during the training phase, it _might_ be
possible to undo the damage, if an AI-shaper of sufficient skill can
try to treat the AI. This requires one roll vs. skill for each month
of 'treatment'. A failure has no effect, a critical failure means the
problem is permanent, four successes in a row or one critical success
undoes the damage.

Artificial intelligences can pass the Turing Test...within certain
limits.

The traditional Turing Test involves a blind comparison between an AI
and a human, and 'passing' the test means that reasonably intelligent
and perceptive humans are unable to distinguish between the answers
and responses of the machine and the other humans. The AIs of the
late 21C can pass this test, with certain caveats.

Low-IQ AIs can pass a Turing Test within the limits of comparison for
a given _animal_, _if_ the AI is configured in such as way as to
behave somewhat like that animal. Thus, in GURPS terms, a typical dog
as an IQ of 5, and so might an AI, but the AI would only be mistaken
for a dog if the AI was designed by its shaper(s) to think, behave,
processing information, etc, like a dog, which would probably also
require being equipped with a body and senses comparable to those of a
dog. An AI with an IQ of 5 might be comparable to a dog in potential
intelligence and awareness, but lacking those qualities no
Homosentient would have any trouble discerning the difference between
it and a dog.

Likewise, an IQ 10 AI might well be comparable to a human in
intellect, but it would have to be _designed_ to act/think like a
human to have any chance of passing a 'traditional' Turing Test. On
the other hand, a Homosentient dealing with a non-human-like AI of IQ
10 would quickly recognize that he or she was dealing with an
intellect of some kind, and not simple rote programming.

Almost all AI’s lack certain traits nearly universal in humans and
animals. To a first apprpximation, the brain (or equivalent organ/
system) in any Solarigen animal is the product of a multi-gigayear
evolutionary history shaped by high-priority imperatives that
sometimes complement and sometimes conflict. Self-preservation is a
nearly universal imperative, closely linked to survival-based
imperatives such as hunger, thirst, predator-avoidance, danger
avoidance, etc. Likewise, self-replication is a high-priority
imperative, sometimes strong enough to entirely override even the
survival imperative, for the basic evolutionary reason that failure to
self-replicate consigns the trait contributing to that failure to
failure.

AIs, for the most part, have the equivalent of these imperatives only
if the AI-shaper incorporates them into the matrix. It would, for
example, be perfectly possible to create an AI lacking any sense of
self-preservation...in theory. In practice, such efforts tend to be
inherently unstable and to collapse rapidly. The intricate
mathematical structures that form the nucleus of an AI are not
infinitely variable, some traits work together and some do not, some
combinations are stable and some are not, and sometimes these match
natural patterns, just as natural selection found those patterns and
they endured because they _work_.

Thus, AI’s have some traits in common with biological minds, and lack
some common biological traits. Any AI has a certain amount of self-
preservation in its nature, it _has_ to or it won’t survive long
enough to learn anything useful or be useful. An entity that does not
care whether it survives or not will not survive long in an entropic
Universe.

Most AI systems don’t have anything like as _strong_ a self-
preservation drive as biological creatures, however. They usually are
not designed with such in mind. Likewise, AI’s lack any sort of
reproductive drive, there’s no _reason_ for such a thing to be present
in their organizing matrices. [1] No AI feels anything like hunger,
thirst, or pain (unless some such feature is designed into the
matrix). This tends to make their intelligence sometimes more focused
than biological minds, but also less complex and subtle.

AIs can and do display some interesting ‘emergent behaviors’ that seem
independent of design. The more ‘intelligent’ the AI, the more such
unpredicted quirks tend to emerge, and the more complex they can be.
However, lacking many of the instinctual drives and hardwired
imperatives of biologicals, they also lack certain sources of mental
sophistication.

No AI system known in the late 21C is anything like as complex,
sophisticated, or unpredictable as their biological counterparts at a
given level of ‘IQ’. [2]

This can have counterintuitive results, such as when AI-driven robots
compete with natural dogs or other animals in the wild, in
experimental contests. AI systems theoretically more intelligent than
animals often tend to lose to the biological competitors when using
similarly capable ‘equipment’ because they lack the _instinctual_ side
of intelligence derived from evolutionary history. On the other hand,
the AIs with a comparable level of intelligence to a given animal can
usually far outperform them in areas to which the AIs are specially
adapted, areas the instincts don’t touch.

MORE LATER.


[1] Obviously exceptions can and do exist, such as Von Neumann
machines. But even here it’s a _dispassionate_ imperative.

[2] A handful of exceptions exist, but they press the definition of
the term ‘artificial intelligence’.