Re: strong and weakly self improving processes

From: Eliezer S. Yudkowsky (
Date: Sat Jul 15 2006 - 15:52:50 MDT

Eric Baum wrote:
> Eliezer> It should be emphasized that I wrote LOGI in 2002;
> Didn't know that. Are the rest of the papers in that 2005 book as old?
> Eliezer> Nonetheless, calling something "complex" doesn't explain it.
> Methinks you protest too much, although I take the point.

You may be right. Still, better to protest too much than too little.

> But I did
> like the presentation-- you didn't just say it was complex, you
> pointed out it was layered, which some in the AI community had failed
> to adequately credit (cf your critique of semantic nets).

Oh, I'm willing enough to say that *human* intelligence is complex,
because I have a specific image of human intelligence as including a
hugely subdivided cerebral cortex, layers of organization, multiple
centers of gravity, many individually evolved instincts and intuitions
in conflict, et cetera.

Saying that *intelligence* is complex is a whole different story.

> Eliezer> A giant lookup table is a simple process that may know an
> Eliezer> arbitrarily large amount, depending on the incompressibility
> Eliezer> of the lookup table. A human programmer turned loose on the
> Eliezer> purely abstract form of a simple problem (e.g. stacking
> Eliezer> towers of blocks), who invents a purely abstract algorithm
> Eliezer> (e.g. mergesort) without knowing anything about which
> Eliezer> specific blocks need to be moved, is an example of a complex
> Eliezer> process that used very little specific knowledge about that
> Eliezer> specific problem to come up with a good general solution.
> I respectfully suggest that the human programmer couldn't do that
> unless he knew a lot, in fact unless he had most of the program
> (in chunks, not exactly assembled, capable of being assembled in
> different ways to solve different problems) already in his head before
> attacking the problem.

But that is not knowledge *specifically* about the blocks problem. It
is not like having a giant lookup table in your head that says how to
solve all possible blocks problems up to 10 blocks. The existence of
"knowledge" that is very generally applicable, far beyond the domains
over which it was generalized, is what makes general intelligence
possible. It is why the problem is not NP-hard.

> Even an untrained human couldn't do it, and an untrained human
> is 10^44 creatures worth of evolution away from a tabula rasa.
> Eliezer> Is the term "top level" really all that useful for describing
> Eliezer> evolutionary designs? The human brain has more than one
> Eliezer> center of gravity. The limbic system, the ancient goal
> Eliezer> system at the center, is a center of gravity; everything grew
> Eliezer> up around it. The prefrontal cortex, home of reflection and
> Eliezer> the self-model, is a center of gravity. The cerebellum,
> Eliezer> which learns the realtime skill of "thinking" and projects
> Eliezer> massively to the cortex, is a center of gravity.
> What do you mean by center of gravity?

About the same thing you mean by "top module"? An axis around which
cognition turns? A major command-and-control outpost? I'm not sure
that I have a better definition than the intuitive sound of the words
plus the examples given.

The center of a self-modifying AI would be none of these things; it
would be the criterion against which self-written code is checked.

> I talked about levels in part because it was the subject of
> your paper :^) , and because the comment I was discussing seemed to
> have a two level nature (the human and the culture);
> but I do tend to think that big hierarchic programs tend to
> have top modules to a (potentially somewhat fuzzy) extent,
> and I tend to think of information as
> being filtered and processed up to a point where decisions are made,
> and the brain certainly has a somewhat layered structure.

I have my skepticism about the proper design for an AI being a big
hierarchic program. Or a lot of little agents. Or a lot of little
agents controlled by an incorruptible central broker. My current
thinking tends to turn around stages of processing - *not* necessarily
layers of organization as in the human idiom. Does information have to
be filtered *up* to the decision-making level? Or is making the
decision just one more *stage* of processing? Ultimately, a mind is the
cognition that happens between sense input and motor output. If we
write down the stages of an AI, and find a natural mountain - beginning
with complex sense information, being processed toward a peak of
simplicity and a direct decision, then increasingly complex translation
toward motor stages - then we might call the peak of simplicity the "top".

But it doesn't follow that the optimal stages between sense and motor
*must* obey any such neat progression. Maybe, when you design the
system properly - as opposed to blindly accreting it by natural
selection - some stages are more complicated and some are less
complicated, and there's no natural top.

*Within humans*, the evolutionary idiom of levels of organization, and
the actual design of the architecture, are such that we can speak
comprehensibly of humans having a "top" level. In fact, I can think of
at least three of them: the limbic system, the prefrontal cortex, and
the cerebellum.

> Eliezer> From my perspective, this argument over "top levels" doesn't
> Eliezer> have much to do with the question of recursive
> Eliezer> self-improvement! It's the agent's entire intelligence that
> Eliezer> may be turned to improving itself. Whether the greatest
> Eliezer> amount of heavy lifting happens at a "top level", or lower
> Eliezer> levels, or systems that don't modularize into levels of
> Eliezer> organization; and whether the work done improves upon the
> Eliezer> AI's top layers or lower layers; doesn't seem to me to
> Eliezer> impinge much upon the general thrust of I. J. Good's
> Eliezer> "intelligence explosion" concept. "The AI improves itself."
> Eliezer> Why does this stop being an interesting idea if you further
> Eliezer> specify that the AI is structured into levels of organization
> Eliezer> with a simple level describable as "top"?
> As I said:
>>even if there would be some way to keep modifying the top level
>>to make it better, one could presumably achieve just as powerful an
>>ultimate intelligence by keeping it fixed and adding more powerful
>>lower levels (or maybe better yet, middle levels) or more or better
>>chunks and modules within a middle or lower level.
> You had posed a 2 level system: humans and culture,
> and said this was different from a seed AI, because the humans modify
> the culture, and that's not as powerful as the whole AI modifying
> itself.

Okay. I would still defend that. Not sure how the internal structure
of the AI directly relates to the above issue.

> But what I'm arguing is that there is no such distinction,
> the humans modifying the culture really does modify the humans in
> a potentially arbitrarily powerful way;
> within most AI's I can conceive, there will in any case be some
> fixed top level, even within an AIXI or in
> Schmidhuber's OOPS or whatever to the extent I understand them,

AIXI has an unalterably fixed top level; it cannot conceive of the
possibility of modifying itself.

Schmidhuber's OOPS, if I recall correctly, supposedly has no invariants
at all. If you can prove the new code has "greater expected utility",
according to the current utility function (even if the new code includes
changes to the utility function), and taking into account all changes
that will be adopted by the new code, the new code gets adopted. But
Schmidhuber is very vague about exactly how this proof takes place.

My own thinking tends to the idea of a preserved optimization target,
preserved preferences over outcomes, rather than protected bits in memory.

> yet this doesn't preclude these things from powerful self modification,
> having a 2 level system where the top level can't modify its very
> top level (eg the humans can't modify their genome-- positing for the
> sake of argument they don't and we only talk about progress that's
> occurred to date) does not make it weakly self-improving in some sense
> that bars it from gaining as much power as a "strongly self-improving"
> alternative.

It is written in the _Twelve Virtues of Rationality_ that the sixth
virtue is empiricism: "Do not ask which beliefs to profess, but which
experiences to anticipate. Always know which difference of experience
you argue about."

So let's see if we can figure out where we anticipate differently, and
organize the conversation around that.

The main experience I anticipate may be described intuitively as "AI go
FOOM". Past some threshold point - definitely not much above human
intelligence, and probably substantially below it - a self-modifying AI
undergoes an enormously rapid accession of optimization power (unless
the AI has been specifically constructed so as to prefer an ascent which
is slower than the maximum potential speed). This is a testable
prediction, though its consequences render it significant beyond the
usual clash of scientific theories.

The basic concept is not original with me and is usually attributed to a
paper by I. J. Good in 1965, "Speculations Concerning the First
Ultraintelligent Machine". (Pp. 31-88 in Advances in Computers, vol 6,
eds. F. L. Alt and M. Rubinoff. New York: Academic Press.) Good
labeled this an "intelligence explosion". I have recently been trying
to consistently use the term "intelligence explosion" rather than
"Singularity" because the latter term has just been abused too much.

Now there are many different imaginable ways that an intelligence
explosion could occur. As a physicist, you are probably familiar with
the history of the first nuclear pile, which achieved criticality on
December 2nd, 1942. Szilard, Fermi, and friends built the first nuclear
pile, in the open air of a squash court beneath Stagg Field at the
University of Chicago, by stacking up alternating layers of uranium
bricks and graphite bricks. The nuclear pile didn't exhibit its
qualitative behavior change as a result of any qualitative change in the
behavior of the underlying atoms and neutrons, nor as a result of the
builders suddenly piling on a huge number of bricks. As the pile
increased in size, there was a corresponding quantitative change in the
effective neutron multiplication factor (k), which rose slowly toward 1.
  The actual first fission chain reaction had k of 1.0006 and ran in a
delayed critical regime.

If Fermi et. al. had not possessed the ability to quantitatively
calculate the behavior of this phenomenon in advance, but instead had
just piled on the bricks hoping for something interesting to happen, it
would not have been a good year to attend the University of Chicago.

We can imagine an analogous cause of an intelligence explosion in which
the key parameter is not the qualitative ability to self-modify, but a
critical value for a smoothly changing quantitative parameter which
measures how many additional self-improvements are triggered by an
average self-improvement.

But this isn't the only potential cause of behavior that empirically
looks like "AI go FOOM". The species Homo sapiens showed a sharp jump
in the effectiveness of intelligence, as the result of natural selection
exerting a more-or-less steady optimization pressure on hominids for
millions of years, gradually expanding the brain and prefrontal cortex,
tweaking the software architecture. A few tens of thousands of years
ago, hominid intelligence crossed some key threshold and made a huge
leap in real-world effectiveness; we went from caves to skyscrapers in
the blink of an evolutionary eye. This happened with a continuous
underlying selection pressure - there wasn't a huge jump in the
optimization power of evolution when humans came along. The underlying
brain architecture was also continuous - our cranial capacity didn't
suddenly increase by two orders of magnitude. So it might be that, even
if the AI is being elaborated from outside by human programmers, the
curve for effective intelligence will jump sharply. It's certainly
plausible that *the* key threshold was culture, but because we wiped out
all our nearest relatives, it's hard to disentangle exactly which
improvements to human cognition were responsible for what.

Or perhaps someone builds an AI prototype that shows some promising
results, and the demo attracts another $100 million in venture capital,
and this money purchases a thousand times as much supercomputing power.
  I doubt a thousandfold increase in hardware would purchase anything
like a thousandfold increase in effective intelligence - but mere doubt
is not reliable in the absence of any ability to perform an analytical
calculation. Compared to chimps, humans have a threefold advantage in
brain and a sixfold advantage in prefrontal cortex, which suggests (a)
firmware is more important than hardware and (b) small increases in
hardware can support large improvements in firmware.

Humans, thinking, certainly cause changes to their neurons; and it may
even be possible that with a theoretically perfect series of
instructions to our introspective levers, we could reprogram the
firmware into whatever we liked. Just as it's theoretically possible
that the genome could contain a series of DNA instructions which built
something that built something that built diamondoid nanotechnology and
placed it under the control of our high-level decision process, thus
obviating all discussion of protected levels. But the genome *doesn't*
contain those instructions, and naive humans don't even know the visual
cortex exists, let alone have the power to reprogram it, and this is not
coincidence. In theory, a sub-critical nuclear pile could have every
single emitted neutron just happen to strike another nucleus, and so
explode; but it's not very *probable*.

There is a level at which an AI is doing exactly the same thing as a
human, who in turn is doing exactly the same thing as a chimp, who is
doing exactly the same thing as a bacterium, who is doing exactly the
same thing as a rock. This level is called physics. There'll be some
level on which the behavior of the system is smoothly continuous with
all its past history, changing neither qualitatively nor quantitatively.

I do not insist that an AI reaching down to its hardware and firmware
levels must change *everything*. It doesn't have to violate the laws of
physics. The important point of debate is not that the AI is
"different" in some sense of how we describe it; the question is
observed behavior. If the pragmatic result of an AI being able to
modify and improve its own hardware and firmware is that the AI
increases its effective self-improvement multiplication factor past 1 -
metaphorically speaking - and goes "critical", then that's the important
thing from my perspective. Or, if humans have already achieved cultural
criticality, but the AI goes prompt critical (metaphorically speaking)
and ascends at rates far faster than human culture, then again I regard
that as the important empirical consequence.

I don't think there should be a question that being able to improve your
hardware (possibly by millionfold or greater factors) and rewrite your
firmware should provide *some* benefit. *How much* benefit is the issue
here. Whether the change I'm describing is "qualitatively different" is
a proxy question, which may turn on matters of mere definition; the key
issue is what we observe in real life.

Now, if you said that humans are already self-modifying to such a degree
that we should expect *no substantial additional benefit* from an AI
having direct access to its own source code, *then* I'd know what
difference of empirical anticipation we were arguing about.

>>>I think the hard problem about achieving intelligence is crafting
>>>the software, which problem is "hard" in a technical sense of being
>>>NP-hard and requiring major computational effort,
> Eliezer> As I objected at the AGI conference, if intelligence were
> Eliezer> hard in the sense of being NP-hard, a mere 10^44 nodes
> Eliezer> searched would be nowhere near enough to solve an environment
> Eliezer> as complex as the world, nor find a solution anywhere near as
> Eliezer> large as the human brain.
> Eliezer> *Optimal* intelligence is NP-hard and probably
> Eliezer> Turing-incomputable. This we all know.
> Eliezer> But if intelligence had been a problem in which *any*
> Eliezer> solution whatsoever were NP-hard, it would imply a world in
> Eliezer> which all organisms up to the first humans would have had
> Eliezer> zero intelligence, and then, by sheer luck, evolution would
> Eliezer> have hit on the optimal solution of human intelligence. What
> Eliezer> makes NP-hard problems difficult is that you can't gather
> Eliezer> information about a rare solution by examining the many
> Eliezer> common attempts that failed.
> Eliezer> Finding successively better approximations to intelligence is
> Eliezer> clearly not an NP-hard problem, or we would look over our
> Eliezer> evolutionary history and find exponentially more evolutionary
> Eliezer> generations separating linear increments of intelligence.
> Eliezer> Hominid history may or may not have been "accelerating", but
> Eliezer> it certainly wasn't logarithmic!
> Eliezer> If you are really using NP-hard in the technical sense, and
> Eliezer> not just a colloquial way of saying "bloody hard", then I
> Eliezer> would have to say I flatly disagree: Over the domain where
> Eliezer> hominid evolution searched, it was not an NP-hard problem to
> Eliezer> find improved approximations to intelligence by local search
> Eliezer> from previous solutions.
> I am using the term NP-hard to an extent metaphorically, but
> drawing on real complexity notions that problems can really be hard.
> I'm not claiming that constructing an intelligence is a decision problem
> with a yes-no answer; in fact I'm not claiming it's an infinite class
> of problems, which is necessary to talk about asymptotic behavior
> altogether.
> It's a particular instance-- we are trying to construct
> one particular program that works in this particular world,
> meaning solves a large collection of problems of certain types.

Okay, problems *can* be hard; what reason do you have to believe that
this particular problem *is* hard?

> (I don't buy into the notion of "general intelligence" that solves
> any possible world or any possible problem.)

I agree. An AI is supposed to work in the unusual special case of own
low-entropy universe, not all possible worlds. No-Free-Lunch theorems etc.

> I think the problem of constructing the right code
> for intelligence is a problem like finding a very short tour in
> a particular huge TSP instance. A human can't solve it by hand, (for
> reasons that are best understood by thinking about complexity theory
> results about infinite problem classes and in the limit behavior,
> which is why I appeal to that understanding).
> To solve it, you are going to have to construct a good algorithm,
> *and run it for a long time*. If you do that, you can get a better
> and better solution, just like if you run Lin-Kernighan on a huge
> TSP instance, you will find a pretty short tour.

Finding a *short*, but not *optimal*, tour in a particular huge TSP
instance, is not an NP-hard problem - there are algorithms that do it,
as you mention. And much more importantly from the perspective of AI
design, it was not an NP-hard problem for a programmer to find those

I furthermore note that the problem of constructing intelligent code
doesn't seem to me at all like the problem of finding a short tour in a
huge TSP instance. The world has a vast number of exploitable
regularities, which have similarities and differences between
themselves; there are meta-regularities in the regularities which can in
turn be exploited. You can eat them one at a time, or swallow
metaproblems in whole gulps.

Magic takes many forms. When you don't know how to do something, you
can appeal to complexity, to emergence, to huge heaps of hardware, to
vague similarities to the human brain... Are you sure that you aren't
saying "We'll need to run the code for a long time" in order to
generate, within yourself, a feeling of having thrown something really
powerful at the problem? Like de Garis talking about ten thousand
neural-net-module-engineers constructing an intelligent being? Do you
know specifically what is the algorithm that you think *must* be run to
generate an intelligence, and can you calculate quantitatively how long
it takes to run?

We know that natural selection took a long time to run, but natural
selection is a bloody inefficient algorithm. Natural selection is so
ridiculously simple that we can even calculate quantitatively how
inefficient it is, and come up with estimates like 2 ln(N) / s
generations to fix a single mutation with advantage s in population N.

I wouldn't be surprised if, in the course of building an AI, there were
points where I found it convenient to run simple algorithms for a long
time. But too much of this would signify that I was trying to
brute-force the problem and failing to exploit important regularities in it.

> Evolution ran for a heck of a lot of computation on the problem.
> It is possible that humans will be able to jump start a lot of
> that, but its also true we are not going to be able to run
> for as much computation. Its an open question whether we can get
> there, but I suggest it may take a composite algorithm-- both
> jump starting the code design and then running a lot to improve it.

Actually, I very much like the idea of running simple programs for a
long time to boot up an intelligence. Not because it's the only way to
get intelligence, or even because it's convenient, but because it means
that the humans have less complexity to potentially get wrong. I
wouldn't use an evolutionary program because then I'd lose control of
the resulting complexity, thus obviating the whole point of starting out

> Eliezer> Now as Justin Corwin pointed out to me, this does not mean
> Eliezer> that intelligence is not *ultimately* NP-hard. Evolution
> Eliezer> could have been searching at the bottom of the design space,
> Eliezer> coming up with initial solutions so inefficient that there
> Eliezer> were plenty of big wins. From a pragmatic standpoint, this
> Eliezer> still implies I. J. Good's intelligence explosion in
> Eliezer> practice; the first AI to search effectively enough to run up
> Eliezer> against NP-hard problems in making further improvements, will
> Eliezer> make an enormous leap relative to evolved intelligence before
> Eliezer> running out of steam.
> I don't know what you mean here at all.

Did previous paragraphs clear it up? In other words, Corwin's notion is
that a *properly designed* intelligence is good enough that making
further improvements is NP-hard, but human intelligences are operating
far short of the level where this happens. Like starting out with a
random traversal of the TSP graph; there'll be plenty of low-hanging
fruit, and if you only take them one at a time, they'll last quite a
while - you might start thinking it was an easy problem. Corwin's
notion is that human intelligence is so poorly designed as to still
occupy this regime; single mutations can still lift us up.

>>>so the ability to make sequential small improvements, and bring to
>>>bear the computation of millions or billions of (sophisticated,
>>>powerful) brains, led to major improvements.
> Eliezer> This is precisely the behavior that does *not* characterize
> Eliezer> NP-hard problems. Improvements on NP-hard problems don't add
> Eliezer> up; when you tweak a local subproblem it breaks something
> Eliezer> else.
>>>I suggest these improvements are not merely "external", but
>>>fundamentally affect thought itself. For example, one of the
>>>distinctions between human and ape cognition is said to be that we
>>>have "theory of mind" whereas they don't (or do much more
>>>weakly). But I suggest that "theory of mind" must already be a
>>>fairly complex program, built out of many sub-units, and that we
>>>have built additional components and capabilities on what came
>>>evolutionarily before by virtue of thinking about the problem and
>>>passing on partial progress, for example in the mode of bed-time
>>>stories and fiction. Both for language itself and things like
>>>theory of mind, one can imagine some evolutionary improvements in
>>>ability to use it through the Baldwin effect, but the main point
>>>here seems to be the use of external storage in "culture" in
>>>developing the algorithms and passing them on. Other examples of
>>>modules that directly effect thinking prowess would be the
>>>axiomatic method, and recursion, which are specific human
>>>discoveries of modes of thinking, that are passed on using language
>>>and improve "intelligence" in a core way.
> Eliezer> Considering the infinitesimal amount of information that
> Eliezer> evolution can store in the genome per generation, on the
> Eliezer> order of one bit,
> Actually, with sex its theoretically possible to gain something like
> sqrt(P) bits per generation (where P is population size), cf Baum, Boneh paper
> could be found on and also Mackay paper. (This is
> a digression, since I'm not claiming huge evolution since chimps).

That's for human-built genetic algorithms, not natural selection. For
natural selection see e.g. (I don't buy some
of the author's claims here, but the central principle of which he gives
a heuristic explanation is something I've heard of before in
evolutionary biology; I think it goes back to Kimura.) Natural
selection does run on O(1) bits per generation.

I furthermore note that gaining one standard deviation per generation,
which is what your paper describes, is not obviously like gaining
sqrt(P) bits of Shannon information per generation. Yes, the standard
deviation is proportional to sqrt(N), but it's not clear how you're
going from that to gaining sqrt(N) bits of Shannon information in the
gene pool per generation. It would seem heuristically obvious that if
your algorithm eliminates roughly half the population on each round, it
can produce at most one bit of negentropy per round in allele
frequencies. I only skimmed the referenced paper, though; so if there's
a particular paragraph I ought to read, feel free to direct me to it.

> Eliezer> it's certainly plausible that a lot of our
> Eliezer> software is cultural. This proposition, if true to a
> Eliezer> sufficiently extreme degree, strongly impacts my AI ethics
> Eliezer> because it means we can't read ethics off of generic human
> Eliezer> brainware. But it has very little to do with my AGI theory
> Eliezer> as such. Programs are programs.
> It has to do with the subject of my post, which was that by modifying
> the culture, humans have modified their core intelligence, so
> there is no distinction from strongly self improving.

You might as well say that, since evolution built humans, evolution is
intelligent, therefore humans are nothing new... but pragmatically
speaking, there seems to be a large qualitative difference in there
somewhere. Ultimately it's all just the same 'ol physics. You could
equally well argue that if we build powerful AIs that shows the power of
human intelligence, but again, it seems like the system went through an
important transition somewhere.

Humans may have modified their core intelligence a *little*, but what
about all the results showing the perseverance of cognitive biases
against self-willed remediation attempts?

> Eliezer> But try to teach the human operating system to a chimp, and
> Eliezer> you realize that firmware counts for *a lot*. Kanzi seems to
> Eliezer> have picked up some interesting parts of the human operating
> Eliezer> system - but Kanzi won't be entering college anytime soon.
> I'm not claiming there was 0 evolution between chimp and man--
> our brains are 4 times bigger.

(Terrence Deacon, in _The Symbolic Species_, says our brains are three
times too large for an ape our size, and that our prefrontal cortex is
relatively six times too large.)

> I'm claiming that the hard part--
> discovering the algorithms-- was mostly done by humans using storage
> and culture. Then there was some simple tuning up in brain size,
> and some slightly more complex Baldwin-effect etc tuning up
> programming grammar into the genome in large measure, so we become
> much more facile at learning the stuff quickly, and maybe other
> similar stuff. I don't deny that if you turn all that other stuff
> off you get an idiot, I'm just claiming it was computationally
> easy.

Arguably, in a certain sense it *must* have been computationally easy
because natural selection is incapable of doing anything computationally
*hard*; evolution can't sit back and design complex interdependent
machinery with hundreds of interlocking parts in a single afternoon,
like a human programmer.

However, chimps can recognize themselves in mirrors and implement
complex political strategies in which A anticipates B's reaction to C,
so there's clearly some level of hardware support among chimps for
empathy and theory of mind, despite the (presumable) lack of
sufficiently complex culture to give rise to a proper Baldwin effect.

The real-world impressive power of human culture dates back largely to
the last hundred thousand years which is an eyeblink of evolutionary
time. Space shuttles are pure products of accumulated culture without
much in the way of space-shuttle-specific adaptive support. Science is
so much larger than the genome that even if we didn't know the answer in
advance, we could guess that most scientific information *had* to be on
paper somewhere, not in the genes.

The question is, when all that lovely knowledge gets written down on
paper, what is the force that does the writing? What is the generator
that produces all this lovely knowledge we're accumulating? Could a
more powerful generator produce knowledge orders of magnitude faster?
Obviously yes, because human neurons run at speeds that are at least six
orders of magnitude short of what we know to be physically possible.
(Drexler's _Nanosystems_ describes sensory inputs and motor outputs that
operate at a similar speedup.) What about better firmware? Would that
buy us many additional orders of magnitude?

If most of the generator complexity lay in a culturally transmitted
human operating system that was open to introspection, then further
improvements to firmware might be trivial. But then scientists would
have a much better understanding of how science works; but most
scientists proceed mostly by instinct, and they don't have to learn
rituals on anything remotely approaching the complexity of a human
brain. Most people would find learning the workings of the human brain
a hugely intimidating endeavor - rather than being an easier and simpler
version of something they did unwittingly as children, in the course of
absorbing the larger and more important "human operating system" you
postulate. This human operating system, this modular theory of mind
that gets transmitted - where is it written down? There's a sharp limit
on how much information you can accumulate without digital fidelity of
transmission between generations. The vast majority of human evolution
took place long before the invention of writing.

I don't believe in a culturally transmitted operating system, that
existed over evolutionary periods, which contains greater total useful
complexity than that specified in the brain-constructing portions of the
human genome itself. And even if such a thing existed, the fact that we
haven't written it down implies that it is largely inaccessible to
introspection and hence to deliberative, intelligent self-modification.

>>>I don't understand any real distinction between "weakly self
>>>improving processes" and "strongly self improving processes", and
>>>hence, if there is such a distinction, I would be happy for
> Eliezer> The "cheap shot" reply is: Try thinking your neurons into
> Eliezer> running at 200MHz instead of 200Hz. Try thinking your
> Eliezer> neurons into performing noiseless arithmetic operations. Try
> Eliezer> thinking your mind onto a hundred times as much brain, the
> Eliezer> way you get a hard drive a hundred times as large every 10
> Eliezer> years or so.
> Eliezer> Now that's just hardware, of course. But evolution, the same
> Eliezer> designer, wrote the hardware and the firmware. Why shouldn't
> Eliezer> there be equally huge improvements waiting in firmware? We
> Eliezer> understand human hardware better than human firmware, so we
> Eliezer> can clearly see how restricted we are by not being able to
> Eliezer> modify the hardware level. Being unable to reach down to
> Eliezer> firmware may be less visibly annoying, but it's a good bet
> Eliezer> that the design idiom is just as powerful.
> Eliezer> "The further down you reach, the more power." This is the
> Eliezer> idiom of strong self-improvement and I think the hardware
> Eliezer> reply is a valid illustration of this. It seems so simple
> Eliezer> that it sounds like a cheap shot, but I think it's a valid
> Eliezer> cheap shot. We were born onto badly designed processors and
> Eliezer> we can't fix that by pulling on the few levers exposed by our
> Eliezer> introspective API. The firmware is probably even more
> Eliezer> important; it's just harder to explain.
> Eliezer> And merely the potential hardware improvements still imply
> Eliezer> I. J. Good's intelligence explosion. So is there a practical
> Eliezer> difference?
> The cheapshot reply to your cheapshot reply, is that if we construct
> an AI, that AI is just another part of the lower level in the weakly
> self-improving process, its part of our "culture", so we can indeed
> realize the hardware improvement. This may sound cheap, but it
> shows there is no real difference between the 2 layered system
> and the entirely self-recursive one.

The cheap-cheap-cheap-reply is that if a self-improving AI goes off and
builds a Dyson Sphere, and that is "no real difference", I'm not sure I
want to see what a "real difference" looks like. Again, the cheap^3
reply seems to me valid because it asks what difference of experience we

Eliezer S. Yudkowsky                
Research Fellow, Singularity Institute for Artificial Intelligence

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:56 MDT