From: Eliezer Yudkowsky (sentience@pobox.com)
Date: Sat Oct 23 2004 - 07:01:55 MDT
Jeff Medina wrote:
> Robin Lee Powell wrote: "IIRC, Eliezer is not allowed to put Ph.D.
> after his name. That pretty much rules out this avenue of approach."
>
> That absolutely *does not* rule out this avenue of approach.
Correct.
> Many
> respected journals and conferences in the relevant areas are
> blind-reviewed (such that the academic credentials of the author of
> the paper is made irrelevant, because the author's identity & other
> info is kept secret), and even among the many which are not quality
> submissions are never rejected or looked down upon simply due to the
> lack of a Ph.D. by the author.
Papers which are *not* blind-reviewed show a *decided* bias toward known,
prestigious researchers and institutions. That is *why* some journals are
blind-reviewed. Most aren't, and it doesn't help that other journals are
blind-reviewed if the one you want to target isn't. I've read studies
assessing the bias, but though I googled on "effectiveness of peer review"
I failed to track them down. I did find other interesting material
including, to pick an arbitrary example, a study by Rothwell and Martyn
(2000) showing that peer review in two neuroscience journals was not
reproducible; that is, agreement between reviewers is not significantly
greater than chance. My recollection is that this result is widespread in
studies of this kind. I include this tidbit by way of saying that
experimental study of peer review has produced surprising and alarming
results, so be sure to check out peer-reviewed studies of peer review
before praising its effectiveness.
(Rothwell PM, Martyn CN. Reproducibility of peer review in clinical
neuroscience. Is agreement between reviewers any greater than would be
expected by chance alone? Brain 2000;123:1964-9.)
This has nothing to do with the reason I don't write additional academic
papers. I just thought I'd mention it.
**
First, you'll note that I say "write additional papers". I allocated one
month in early 2002 to write "Levels of Organization in General
Intelligence" (LOGI) in the best academic style I could manage. It
actually took four months. Since then the draft has been online at
http://intelligence.org/LOGI/. In 2005 this paper will finally appear in
"Artificial General Intelligence", eds. Goertzel and Pennachin, to be
published by Springer-Verlag in 2005. (The three-year delay was for the
entire book, not my own paper; *I* turned in my homework on time.)
The fact that none of the people plaguing me to write papers have even
*noticed* "Levels of Organization in General Intelligence", speaking
instead as if I haven't written *any* papers, is indeed related to the
reason I am not more involved with academia.
I've come a long way over the eight years since 1996. People said to me:
Write up your ideas about AI in a web page. In 1998 I did. Then new
people came along and they said: You'll never get anywhere with this, no
one will be interested enough to pay you to do this. In 2000, thanks to
Brian Atkins, the Singularity Institute started up. Possibly that
impressed a few people who never thought I'd get that far. Then new people
came along, to whom Eliezer had *always been* a part of the Singularity
Institute, so it wasn't impressive, and they said: No one will ever pay
attention to you unless you do as we say and write some kind of paper
targeted at academia and get it published. In 2002 I did. I didn't expect
anyone to notice, and no one did, but the effort of writing the LOGI paper
served to help me unify my ideas and force me to read relevant literature
and therefore I account it a partial success. And lo, the people said:
What you really need, Eliezer, is to write some kind of paper targeted at
academia.
Someone always thinks there's just one more thing you need to do. *That*
never changes, no matter how many times you fulfill the request. They just
find something else for you to do. Often it's something you've already
done. I wasn't puzzled by this. I expected it. Thus the particular
things that I did were selected strictly on the basis of their needing
doing, rather than to one-up naysayers.
Case in point: Dr. Eric Drexler and _Nanosystems_.
Before: Eric Drexler has no PhD and hasn't written up his ideas in great
gory technical detail. People tell him: Eric, no one will pay attention
to you if you don't have a PhD. People tell him: Eric, you need to write
up your technical ideas in great gory detail in a way that a wide audience
can understand.
Eric spends six years writing _Nanosystems_ and making it presentable to
any technical reader without demanding a specific background in chemistry,
physics, or computer science. Eric defends _Nanosystems_ as his thesis and
receives the world's first PhD in nanotechnology from MIT.
Afterward: None of the naysayers read _Nanosystems_ or even mention it
exists. No one pays any more attention to Drexler than before. They just
shift their criterion to something else Eric hasn't done yet. Often they
indignantly proclaim that Drexler hasn't given any technical presentation
of his ideas - complete indifference to the work already accomplished. The
same people who liked Drexler before still like him. The kind of people
who objected to Drexler before find something different to which to object.
I suspect those who objected to nanotechnology did not say: "Hm... I have
no idea whether I like this or not... but wait! Drexler doesn't have a
PhD! Okay, now I've decided that nanotechnology is impossible and Drexler
is scaring our children." The causal sequence of events is more like,
"Eek! Too weird! Hm, it seems that I disbelieve in nanotechnology. I
wonder why I disbelieve in nanotechnology? (Searches for reason.) It must
be because Drexler doesn't have a PhD, hey, yeah, that's it." After
Drexler got a PhD, exactly the same process took place, only the
rationalization search terminated elsewhere.
Drexler has a personality far better suited to academia than I'll ever be.
He's humble. He did everything by the book, the way he was supposed to.
Academia... to put it bluntly, they spit in his face. And Drexler had a
vastly easier problem to explain, in a field with all the underlying
physical equations established and agreed upon. If Drexler didn't make it
in academia there's no chance in hell that I could do so. Friendly AI
would be two orders of magnitude harder to sell to academia than molecular
nanotechnology. I pointed out that last part to Drexler, by the way; he
agreed. And come to think, while he didn't say a word to me against
academia or the academic system, Dr. Eric Drexler is *not* on the list of
people whose advice to me included getting a PhD.
I don't want to sound like I'm criticizing Drexler's intelligence. Drexler
did not have Drexler's case to warn him. Drexler's choices were different;
he may have had nothing better to try than getting a PhD and spending six
years writing a technical book.
But people seem to be absurdly optimistic about how easy it is for the
actors on stage to carry out the helpful advice shouted from the audience.
Then again, as plenty of studies show, people are also absurdly
optimistic about the course of their own lives - except for the severely
depressed, who are sometimes properly calibrated with respect to outcomes,
a phenomenon known as "depressive realism". (I am not making this up.)
Part of the reason why people are absurdly optimistic is that they think:
I'll just do X, and then everything will be all right! Not: I'll try to
do X, it will take four times as long as I expect, I'll probably fail, and
even if I succeed, only one in ten successes of this kind have as great an
impact as the one I pleasantly imagined.
I remember meeting Chris Phoenix of CRN at a Foresight Gathering, and Chris
Phoenix spoke optimistically of the day when molecular manufacturing is
proved possible, and all the naysayers have to admit it... and I said:
"Yes, Chris, we can look forward to the fine day when the naysayers are
presented with a working example of mechanosynthesis, and they are finally
forced to stand up and say, in unison: 'Oh, but that isn't *really*
nanotechnology.'"
Did I get a Ph.D., nothing would change. I'd just hear: oh, but you aren't
an eminent scientist in the field, go write more papers.
If I were the sort of person who chased all over the map - starting
companies, getting PhDs, whatever - then I wouldn't be here in the first
place. My life would have happened to me while I was making other plans.
Antoine de Saint-Exupéry: "Perfection is achieved, not when there is
nothing left to add, but when there is nothing left to take away." People
overestimate conjunctive probabilities and underestimate disjunctive
probabilities; they overestimate the chance of many things going right in
sequence, underestimate the probability of a single thing going wrong. The
way to success is to remove everything from the plan that doesn't
absolutely *have* to be there. The way to have any chance at all of
finishing on time is to do nothing that is not absolutely necessary.
Is being a part of academia absolutely necessary to success? I don't think
so. No one's told me to get a PhD in something because in-depth technical
mastery of that subject is absolutely necessary to the creation of AI, and
yet that is *supposed* to be what PhDs are about. No one's said a word
about learning or knowledge. It's all about the impressiveness of some
letters after your name. I know I'm far from the first person to point out
the massive failure of the educational system, but it remains just as huge
a problem and just as horribly awry. The failure doesn't go away just
because someone has pointed it out before.
To tackle AI I've had to learn, at one time or another, evolutionary
psychology, evolutionary biology, population genetics, game theory,
information theory, Bayesian probability theory, mathematical logic,
functional neuroanatomy, computational neuroscience, anthropology,
computing in single neurons, cognitive psychology, the cognitive psychology
of categories, heuristics and biases, decision theory, visual neurology,
linguistics, linear algebra, physics, category theory, and probably a dozen
other fields I haven't thought of offhand. Sometimes, as with evolutionary
psychology, I know the field in enough depth to write papers in it. Other
times I know only the absolute barest embarassingly simple basics, as with
category theory, which I picked up less than a month ago because I needed
to read other papers written in the language of category theory. But the
point is that in academia, where crossbreeding two or three fields is
considered daring and interdisciplinary, and where people have to achieve
supreme depth in a single field in order to publish in its journals, that
kind of broad background is pretty rare.
I'm a competent computer programmer with strong C++, Java, and Python, and
I can read a dozen other programming languages.
I accumulated all that (except category theory) before I was twenty-five
years old, which is still young enough to have revolutionary ideas.
That's another thing academia doesn't do very well. By the time people
finish a Ph.D. in *one* field, they might be thirty years old, past their
annus mirabilis years. To do AI you need a dozen backgrounds and you need
them when you're young. Small wonder academia hasn't had much luck on AI.
Academia places an enormous mountain of unnecessary inconveniences and
little drains of time in the way of learning and getting the job done. Do
your homework, teach your classes, publish or perish, compose grant
proposals, write project reviews, suck up to the faculty... I'm not saying
it's all useless. Someone has to teach classes. But it is not absolutely
necessary to solving the problem of Friendly AI.
Nearly all academics are untrained in the way of rationality. Not
surprising; few academics are fifth-dan black belts and there are a lot
more fifth-dan black belts than fifth-dan rationalists. But if I were in
academia I would be subject to the authority of those who were not Bayesian
Masters. In the art of rationality, one seeks to attain the perception
that most of the things that appear to be reasons and arguments are not
Bayesian. Eliminate the distractions, silence the roar of cognitive noise,
and you can finally see the small plain trails of genuine evidence. One of
my academic friends once asked me to look at a paper on decision theory;
the paper described the conventional theory, presented a problem, and then
proposed several different individual patches to the conventional theory
and analyzed the patches individually, concluding that none of the
solutions were satisfactory. I replied by arguing that the conventional
theory actually contained *two* independent foundational errors, which
needed to be simultaneously refactored to solve the problem, and in fact,
he needed to look at this whole problem a different way. And the one said:
But I have to take the traditional theory as a point of departure and
then present changes to it, because that's what the reviewers will expect.
And I said: Okay, but for myself I don't have to give a damn about
reviewers, and so I plan to go on using the solution with two simultaneous
corrections. That bias against two simultaneous changes, owing to the need
to take the conventional theory as a point of departure, was justified as
necessary by pointing to social forces instead of Bayesian forces. That
makes it a distraction.
I refuse to accept that entire class of distractions. As an independent
scholar, I never have to give any reason for saying something or thinking
something that points to social forces instead of the facts of the matter.
I have the freedom to do the right thing, without the faintest bias
toward the academically acceptable thing except insofar as the academically
acceptable thing happens to be right. I have the luxury of giving no more
credence to an idea than the weight of Bayesian evidence calls for, even if
the idea has become fixed in academia through any of the non-Bayesian
processes that prevail there.
Now, most of the time, I don't second-guess academia - certainly not in
established fields with great weights of evidence, nor after learning just
the basics of something. Like I said, it's dangerous to be half a
rationalist; if you learn the skill of challenging conventional ideas,
you'd damn well better learn the skill of accepting conventional ideas, or
end up worse off than before. But sometimes, on the fringes, AI for
example, people just make stuff up that sounds cool, and it becomes fixed
because everyone repeats it. Look at Freudian analysis: not one scrap of
experimental evidence. It was a major academic field, with peer-reviewed
journals and everything, but not the faintest hint of science. If that's
the academic standard then academia's standards are too damn low. Or
sometimes the people in one field don't know about the results in another
field, and they say things that are silly and get past the reviewers,
because the people who could catch the mistake work in a different building
of the college campus. That likewise happens, *a lot*, in AI.
It seems to me that the secret of effectiveness is refusing to be
distracted. At one point in my life I did permit myself to be
distracted... by writing freelance programs, by planning to start a
company... eventually I noticed that the only projects in my life that had
ever done the slightest bit of good were the ones that were *directly* on
track to the Singularity. *Not* the distraction projects that I thought
would provide resources or whatever, but the projects that were directly
part of the critical path to AI. In 1998 I took one month out of my
all-important plots to accumulate Singularity resources to write "Coding a
Transhuman AI", and in the end CaTAI that was the only thing I did that
year that actually mattered. And that was very much the story of my life,
until the day I finally snapped and decided to concentrate solely on the
Singularity. Today I refuse to be distracted. Not by academia, not by
technology companies, not by anything. All I ask of myself is that I do
this one thing, solve this one challenge of Friendly AI.
> 2. If lacking a PhD really becomes a problem... well, why not get
> one? PhD students get living stipends and support for their research.
> So even being a PhD student may well put Eliezer in a better position
> to pursue his research than the current scenario allows. Further, if
> he (or anyone else involved) doesn't like the idea of being forced to
> take 2 years of coursework for the PhD, he could always pursue the PhD
> outside of the U.S., where PhDs are pure research degrees with no
> course requirements.
>
> 3. There are a couple of schools (e.g., The University of Technology,
> Sydney, in Australasia) that award PhDs by prior publication. After
> applying, you put together a portfolio of your research, and write an
> overarching paper that illustrates your contribution to the field of
> study, and if deemed PhD-level, you are granted a PhD. I've come
> across at least a few professors in the UK and elsewhere who have
> received their doctorates in this manner. (I've also seen quite a few
> professors with just Master's, but this falls back to point 2 above).
>
> Of course, most people aren't aware of some or all of the above
I wasn't aware. Thanks. If I don't need to spend eight years, that does
shift the cost/benefit ratio. But not far enough, I'm afraid.
-- Eliezer S. Yudkowsky http://intelligence.org/ Research Fellow, Singularity Institute for Artificial Intelligence
This archive was generated by hypermail 2.1.5 : Tue Feb 21 2006 - 04:22:47 MST