Where all AGI alignment schema must make use of
some sort of error correction dynamics,
that the fundamental intractability
of any such schema is herein considered.
That any procedure of AGI alignment will
resemble some sort of error correction algorithm.
As with all algorithms,
the error correction algorithm itself
occasionally experiences errors.
This leads to a kind of recursion:
What code watches the code that watches the code?.
Unfortunately, no matter how many levels
of error correction are introduced,
there is always at least one more level of randomness
(possible errors and error types)
that nature introduces.
Note that "Code" (itself as a fixed state)
is also communication across time,
and over the very long term,
is just as subject to inherent 'channel losses'
as any other conduit.
This is as true for every layer of 'error correction'
as it is for 'the base code'.
Current systems usage
is merely tolerant to infrequent errors
simply due to larger non-criticality.
However, AI alignment is in every aspect a criticality --
the error only needs to happen once,
and you have a "non-aligned AI".
At which point,
game theoretic models take over --
the argument cannot *not* use ecosystem dynamic form
at that point.
Hence, the overall form of the paper.
> What if the aligned AI sets it's intelligence
> to making 'super good error correction code' ("SGECC")?.
SGECC is non-physical.
It is inherently non-realizable,
due to pure physics and math.
Ie; this truth does not depend on
any conceivable notion of 'intelligence',
regardless of degree or extent of that 'smartness'.
It is a result of pure information theory,
and thus, inherent in any compute system,
and thus, inherently at the base of
any intelligence system.
The implications are inescapable.
(Ie, that any error correction process
is strictly equivalent to a message passing process
(via whatever channel of feedback the regulation
and/or control signal is happening via)
and therefore, subject to all of the considerations
of any communication channel --
some noise will for sure be added.
Hence the need to add redundancy, etc,
to ensure that the errors are corrected.
However, there is no such thing as a perfected
perfectly error free error detection/correction
process that is proof to *all* types of errors.
There will always be *some* class of errors
for which a given correction protocol
simply does not work.
There are also arguments based on the
"no free lunch theorem" insofar as that it
is not even possible to "learn" how to correct
all possible classes of errors --
that there will always be forms of randomness
which are simply sufficient to overcome
any possible regularity seeking process
(ie; what learning is, what error correction
would depend on, etc)).
Also, notice that the problem of
what is meant by "error"
gets inherently increasingly ambiguous
as more and more dimensions and levels
of abstraction are added,
and thus less meaningful
in a real applied/embodied sense --
ie, to anything that would matter to
any ostensive "AGI Alignment Builder".
The more you try to invent SGECC,
the more you notice that the concept
is itself inherently self contradictory.
Perfected error correction is an illusion --
a notion created to appease politicians and diplomats.
Also, even if there was hypothesized to be an
an error correction system for the error correction
system, which inspecting itself for errors, etc,
there are even more classes of critical problems.
Adding more and more meta-levels of error correction
simply obscures the dynamics associated with any
single level of practice, basically because these
concepts end up being mappable as being functionally
the same in any case.
Notice that 1; any "found error"
will be 'a point change' to its own algorithm
(affecting one 'component' of that algorithm),
and 2; that the number of components required
in that algorithm,
to find such an 'errorfull component'
is strictly greater than one.
Now we can consider 3; the number of self components
that can be inspected
(in whatever number of dimensions of inspection)
as compared to the number of algorithm components
that would be minimally needed
to even have that capability.
Notice that this number (number of relations)
associated with the 'scope of examine'
includes not only itself,
but also everything outside of itself
which it also (in any way) controls/influences,
(and is therefore responsible to align)
as multiplied for/over
all of the dimensions of inspection
that are needed (for alignment, etc).
And then that model of itself that
would need to be aligned/tested/corrected
now also needs to account for the fact
that the model itself is touched everywhere
in its embodiment in at least some substrate.
So in summary, we have 'x1'
as the number of relations
in the error correction model/algorithm itself,
plus 'x2',
as the number of relations
in all that "is self" and that is not
the error correction model specifically
(any robot needs to be more than just
a self reflecting meditating brain
to do anything in the real world),
plus 'x3',
as any world relations
controlled and/or influenced
by the AGI/self
(since that affects alignment realization),
and however 'x4'
how it models that outside world
and/or 'x5'
any aspects of self which are not
'error correction aspects of self',
and then also, 'x6',
it needs to add in all of the relations
of everywhere the AGI substrate touches
all of these other aspects.
That is a lot of factors,
none of which can be omit.
This becomes a total 'T' of something like:
T = x1 + x2 + x3 + x4 + x5
+ x6 * (x1 + x2 + x3 + x4 + x5)
Notice the product relation,
and that we do not know if
there are multiple levels of interaction
within the substrate X6,
or from the substrate into the AGI.
Also, if there is any kind of recursion,
some of these terms become (explode into)
factorials.
And clearly there have to be,
else the meta-level dimensions
of the model correcting (modeling) itself
are where the mis-alignments can happen,
and so therefore *must* be included too.
Inclusive of all of this,
the number of components
that need to be inspected
always increases strictly faster than
the maximum level of capability/scope
that the error correction model
could possibly have.
The number of components needed
to represent the model of all of this,
self, other, and an everywhere touching
of self and other,
as the number of relations
that are the basis of
the error correction model itself,
is strictly larger than
the number of components that
can be fixed by the model,
both in terms of absolute quantity,
and in terms of strict dimensionality.
It only takes a single product
to ensure this inequality,
and yet we have at least multiple factorials.
This completes the inductive argument form.
I do not care how "super" someone posits
the error correction algorithm to be --
I can always find some component of itself,
in some dimension, that it cannot inspect,
and which also causes ultimate failure.
Then, on top of that, some of these factors,
in all of the above relations,
(every one a possible channel of signal
with error values also passing),
ensure that the error correction algorithm
is also _time_bounded_ relative to the
nearly unbounded error seep rates.
~ ~ ~
> - Paraphrasing the above paragraph;
> are you referring to
> that the time spent
> by the error correcting mechanisms
> to make comparisons on measured inputs
> and respond with some addressed output signal
> fed back into 'the system' to 'correct it'
> is strictly slower (requires more time)
> at directing signals through the system
> than the error-inducing components are
> at directing signals
> even *if* the uncommon/lucky situation transpires that
> that input signals are directly indicative of errors
> without, say, much further processing
> along with previously stored/processed inputs
> (and else, any errors
> fed by expressed-code selected
> through evolutionary selection
> can linger, recurse, and build up
> for an indeterminable period of time).
Yes.
Any channel of communication has some error.
The rate of even correct correction
is strictly less than the rate at which
new error potentially enters the system
at all possible interior points
in the algorithm basis itself.
The usual practice to address this
(such as in computers in spacecraft
which can anywhere get a cosmic ray bit-flip)
is to implement massive redundancy.
It is not error correction, so much as it
is robustness in signal response.
The signal bit rates are still mismatched
(and more so to the number of redundant copies).
Moreover, this robust response strategy
*only* works for fixed system design,
else the notion of 'redundant' has no meaning.
With an inherently *changing* dynamic notion
of what 'alignment' means, then this sort
of strategy simply cannot be well defined,
as the number of dimensions in which the
sense of 'what to copy' is itself varying.
Basically, it is the same pattern.
For each possible branch of the argument,
the same sort of inequality/impossibility
shows up (when thinking of general alignment).
~ ~ ~