From: Eliezer S. Yudkowsky (email@example.com)
Date: Wed Feb 27 2002 - 09:02:14 MST
Ben Goertzel wrote:
> A) paths that begin with unintelligent self-modification
> B) paths that begin with purposeful intelligent non-self-modifying behavior
> C) paths that begin with a mixture of self-modification and purposeful
> intelligent behavior
> Eli and I, at this point, seem to share the intuition that B is the right
> approach. I have been clear on this for a while, but Eli's recent e-mail
> is the first time I've heard him clearly agree with me on this.
I suspect that's because you and I use the terms "hard takeoff" and "seed
AI" to refer to different phases of the AI's development. To be precise:
*** (Excerpt from a work in progress, may contain terms here undefined.)
Epochs for holonic programming:
First epoch: The AI can transform code in ways that do
not affect the algorithm implemented. ("Understanding"
on the order of an optimizing compiler.)
Second epoch: The AI can transform algorithms in order
to fit simple abstract beliefs about the design purposes
of code. That is, the AI would understand what a stack
implemented as a linked list and a stack implemented as
an array have in common. (Note that this is already out
of range of current AI.)
Third epoch: The AI can draw a holonic line from simple
internal metrics of cognitive usefulness (how fast a
concept is cued, the usefulness of the concept returned)
to specific algorithms. Consequently the AI would have
the theoretical capability to invent and test new
algorithms. This does not necessarily mean the AI would
have the ability to invent good algorithms or better
algorithms, just that invention in this domain would
become possible. (A theoretical capacity for invention
does not inherently imply improvement over and above the
inventions of the programmers. This is determined by
relative domain competency and by relative effort
expended at a given focal point.)
Fourth epoch: The AI has a concept of "intelligence" as
the top-level product of a continuous holonic system.
The AI can draw a continuous holonic line from (a) its
understanding of intelligence to (b) its understanding of
cognitive subsystems and cognitive content to (c) its
understanding of source code and stored data. Given a
sufficiently complete understanding of the higher-level
purpose of a cognitive subsystem, the AI would be able to
design a new subsystem within the overall architecture.
(Again, this does not intrinsically imply improvement.)
Fifth epoch: The AI understands almost all of the design
purposes of its lower and higher levels of organization.
The AI would have the ability to design new cognitive
Sixth epoch: The AI's understanding of itself, and the
AI's understanding of intelligence, matches or surpasses
that of the human programmers.
Epochs for sparse and continuous self-improvement:
First epoch: The AI has a limited set of rigid routines
which it applies uniformly. Once these routines are used
up, they are gone. This is essentially analogous to the
externally driven improvement of an optimizing compiler.
An optimizing compiler may make a large number of
"improvements", but they are not self-improvements, and
they are not design improvements.
Second epoch: The cognitive processes which create
improvements have characteristic complexity on the order
of Blue Gene, rather than on the order of an optimizing
compiler. Sufficient investments of computing power can
sometimes yield extra design improvements, beyond the
default operations, but it is essentially an exponential
investment for a linear improvement, and no matter how
much computing power is invested, the total number of
improvements conceivable are limited. (I would identify
this as EURISKO's epoch.)
Third epoch: Cognitive complexity in the AI's domain
competency for programming is high enough that at any
given point there is a large number of visible
possibilities for improvement, albeit minor
improvements. The AI typically does not completely
exhaust a given supply of opportunities before
discovering new ones. However, it is only
programmer-driven improvements in intelligence which are
large enough to make new opportunities for
Fourth epoch: Internal improvements sometimes result in
genuine improvements to "smartness", "creativity", or
"holonic understanding", enough to make new possible
improvements visible. AI-driven acquisition of domain
expertise - independent learning - may also be powerful
enough to "increase the opportunity supply" or "survey a
new portion of the self-improvement landscape".
Fifth epoch: Self-improvement is, theoretically,
open-ended. Even in the complete absence of the human
programmers, by the time the AI had used up all the
improvements visible at a given level, that amount of
improvement would be enough to "climb the next step of
the ladder" and make a new set of improvements visible.
Sixth epoch: The AI does not "use up all the
improvements visible at a given level". Taking only a
small subset of the immediately obvious opportunities is
enough for the AI to climb the next step of the ladder,
survey a new portion of the self-improvement landscape,
and start over in a new space of possible improvements.
Epochs for human-dominated and AI-dominated improvement:
First epoch: The AI can make optimizations at most on the
order of an optimizing compiler, and cannot make design
improvements or increase functional complexity. The
combination of AI and programmer is not noticeably better
than a programmer armed with an ordinary optimizing
Second epoch: The AI can understand a small handful of
components and make improvements to them, but the total
amount of AI-driven improvement is small by comparison
with programmer-driven development. However,
sufficiently major programmer improvements do very
occasionally trigger secondary improvement. The total
amount of work done by the AI serves only as a
measurement of progress and does not significantly
accelerate work on the AI.
Third epoch: AI-driven improvement is significant, but
development is "strongly" programmer-dominated in the
sense that overall systemic progress is driven almost
entirely by the creativity of the programmers. The AI
may have taken over some significant portion of the work
from the programmers. The AI's domain competencies for
programming and the deliberative manipulation of
cognitive content may be critical to the AI's continued
Fourth epoch: AI-driven improvement is significant, but
development is "weakly" programmer-dominated. AI-driven
improvements and programmer-driven improvements are
roughly of the same order, but the programmers are better
at it. Alternatively, the programmers have more
subjective time in which to make improvements, due to the
number of programmers or the slowness of the AI.
Fifth epoch: AI-driven improvement is roughly equal to
the amount of programmer-driven improvement.
Sixth epoch: AI-driven improvement significantly
outweighs programmer-driven improvement.
Seventh epoch: Programmer-driven improvement is
Epochs for overall intelligence:
Tool-level AI: The AI's behaviors are immediately and
directly specified by the programmers, or the AI "learns"
in a single domain using prespecified learning
Prehuman AI: The AI's intelligence is not a significant
subset of human intelligence. Nonetheless, the AI is a
cognitive supersystem, with some subsystems we would
recognize, and at least some mind-like behaviors. (A
toaster oven does not qualify as a "prehuman chef"; a
general kitchen robot might do so.)
Infrahuman AI: The AI's intelligence is, overall, of the
same basic character as human intelligence, but
substantially inferior. The AI may excel in a few
domains where it possesses new sensory modalities or
other brainware advantages not available to humans.
Humans talking to the AI usually recognize a mind on the
other end. (An AI that lacks the ability to communicate
and model external minds does not yet qualify as
Near-human AI, human-equivalent AI: The AI's
intelligence is in the rough neigborhood of a human's.
It may be locally inferior or superior in various
domains, but general intelligence, reasoning ability, and
learning ability are roughly that of a human.
Note that these are *epochs*, not *milestones*. They describe progress over
very long periods.
Anyway, Ben uses the term "hard takeoff" to refer to what I would describe
as the first or second epochs. I use "hard takeoff" in the sense that I
believe is standard in the transhumanist community, to refer to events past
the fifth or sixth epochs in various categories. This would seem to explain
Ben's belief that I "underestimate" the amount of work involved in "getting
to the Singularity from a hard takeoff".
-- -- -- -- --
Eliezer S. Yudkowsky http://intelligence.org/
Research Fellow, Singularity Institute for Artificial Intelligence
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:37 MDT