On 5/17/06, <b class="gmail_sendername">Samantha Atkins</b> <<a href="mailto:sjatkins@mac.com">sjatkins@mac.com</a>> wrote:<br>
<div><blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;"><div><div style=""><div>You
seem to be confusing Singularity with some particular extremely utopian
or dystopian outcomes. I am not using the word in that sense and
I don't think many others in this thread are either. I don't
expect a deity to manifest on Earth as part and parcel of Singularity.</div></div></div></blockquote><div><br>
Eliezer effectively does though, and it's his view that I was arguing against here.<br>
<br>
To recap, my position is:<br>
<br>
1) The Singularity is a fine way of thinking about the distant future, a state of affairs we may ultimately reach.<br>
<br>
2) However, we are not close enough to be able to make meaningful,
specific predictions about questions like what will society be like
during and after the Singularity.<br>
<br>
3) Therefore it is not good to turn it into a political football at
this time, because we lack the data to base a political debate on fact
and reason, and such a debate based on conjecture and emotion is likely
to be counterproductive.<br>
<br>
There are various memes floating around for versions of the idea that
Singularity can be achieved in a short timescale by currently known
processes. I'm arguing that none of these are realistic; the post
you're replying to here was the one in which I present my reasons for
believing that Eliezer's version involving recursive self-improvement,
in particular, is unrealistic. I do this reluctantly, since when I
studied the question of hard takeoff it was not with a view to
disproving it. But I find the conclusion inescapable nonetheless.<br>
</div><br><blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;"><div><div style=""><div><span class="q"><blockquote type="cite">Hard takeoff = a process by which seed AI (of complexity buildable by a
team of very smart humans in a basement) undertakes self-improvement to
increase its intelligence, without needing mole quantities of
computronium, _without needing to interact with the real world outside
its basement_, and without needing a long time; subsequently emerging
in a form that is already superintelligent.<br> <br></blockquote><div><br></div></span></div><div><div>Who
says this will happen with no interaction with the rest of the world on
the part of the seed AI?</div></div></div></div></blockquote><div><br>
Eliezer has said it can so happen, and you yourself support this conjecture in the next few sentences:<br>
</div><br><blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;"><div><div style=""><div><div>As you know much has been written
about the difficulty of keeping the AI isolated sufficiently from the
world. But I have no reason to consider self-improvement or
even really clever hacking utterly insufficient to reach
super-intelligence in that "basement" with just the corpus of
information reachable in read-only mode from the internet and
reasonably good self-improving code. The latter is imho key
if in fact humans are incapable of designing the components of a
super-intelligence.</div></div></div></div></blockquote><div><br>
The concept of "self-improving code" is one of the big pitfalls in
reasoning about this area (the other being the concept of
"intelligence" as a formal property of a collection of bits). Let's
stop to dissect it for a bit.<br>
<br>
When we think about self-improving code, the image that always comes to
our minds (not just mine, this is true of the examples people post when
the idea is discussed) is that of tweaking code to produce the _same
output_ using fewer CPU cycles or bytes of memory.<br>
<br>
I emphasize those words because they are the crux of this issue.
Improving an AI's performance isn't about producing the same output
faster, it's about producing _different output_ from the same input. In
other words, it's not just about changing the code, but changing the
_specification_, and that's a completely different thing. As you know
yourself, changing the specification with a reasonable assurance that
the result will be an improvement isn't the sort of thing that can be
done by a smart compiler. It's something that requires domain (not just
programming) expertise, and real-world testing - which is what I've
been saying.<br>
</div><br><blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;"><div><div style=""><div><div>Why do you think computronium is
required? It was not required to get to human intelligence
obviously.</div></div></div></div></blockquote><div><br>
Yes it was. A single human brain contains more than an exaflop's worth
of fault-tolerant self-rewiring nanotech computronium (just looking at
the neurons alone, nevermind the as yet unquantified contributions of
the glial cells, peripheral nervous system and rest of the body). And
getting to human intelligence took a large population of such entities
over millions of years of interaction with and live testing in the real
world.<br>
<br>
Not being restricted to blind Darwinian evolution, it shouldn't take us
millions of years to create AI; but the requirement for interaction
with the real world isn't going to go away.<br>
</div><br><blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;"><div><div style=""><div><span class="q"><blockquote type="cite">
1) I've underlined the most important condition, because it exposes the
critical flaw: the concept of hard takeoff assumes "intelligence" is a
formal property of a collection of bits. It isn't. It's an informal
comment on the relationship between a collection of bits and its
environment.<br></blockquote><div><br></div></span></div><div>Huh?
It is you who posited a completely isolated environment that cannot be
breached.</div></div></div></blockquote><div><br>
Actually the concept has been floating around over on SL4 since long before I joined the list; I was just summarizing.<br>
</div><br><blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;"><div><div style=""><div>An initial environment does not have to include the
entire world in order for the intelligence to grow.</div></div></div></blockquote><div><br>
Sure. That doesn't change the fact that the AI will depend on an
environment and the rate at which it learns will depend on the rate at
which it can do things in that environment. The reason I'm emphasizing
this is to refute Eliezer's idea that the AI can learn at the rate at
which transistors switch between 0 and 1, independent of the real world.<br>
</div><br>
<blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;"><div><div style=""><div><div>Yes
and no. What is the real world constraint on building and testing
software?</div></div></div></div></blockquote><div><br>
The hard part is the specification: knowing what output your program should be producing in the first place. <br>
</div><br><blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;"><div><div style=""><div><div>Yes if you want to solve engineering problems you
eventually need to get beyond the design tools and simulation but these
are still highly critical. Actually building and testing a new
solution can be done by human beings from the design the AI came up
with. Where is the crucial problem?</div></div></div></div></blockquote><div><br>
Design tools and simulation are critical yes, and actually building and
testing solutions by humans is also critical (I don't expect an
early-stage AI to be able to effectively control robots).<br>
<br>
As long as these things are available, there is no crucial problem.
Please bear in mind that I'm not saying strong AI isn't possible - on
the contrary, I believe it is. I'm just listing the requirements for it.<br>
<br>
Imagine this is 1906 and there's a debate about whether it's possible to put a man on the moon, and we have three factions:<br>
<br>
1) Skeptics: No way!<br>
2) Russell: Yes, ultimately. But it will take many decades of hard work
designing successively more powerful rockets, building them and - you
can't skip this part - testing them in the real world. There are no
easy short cuts.<br>
3) Eliezer and co: You don't know that. There's no reason why someone
couldn't just launch a successful moon shot from their basement anytime
now.<br>
<br>
My position is: it can be done, but there are tough practical
requirements that need to be met if one wants a real chance at it.
There are no easy short cuts.<br>
</div></div>