From: Eliezer S. Yudkowsky (sentience@pobox.com)
Date: Sun Sep 29 2002 - 16:45:40 MDT
(Here's a little document I once sent a venture capitalist of my acquaintance...
those of you who've met him probably won't have a hard time guessing who.)
-- (On an analogy between evolutionary psychology and venture capital:)
The basis for the analogy is that both evolutionary psychology and venture
capital are about incremental change - not in the sense that revolutions
aren't allowed, but in the sense that any given revolution has to be
achievable in four to six years, maximum, to get venture funding. Both
evolutionary psychology and venture capital try to tear their practioners'
minds away from the beautiful endpoints and make us ask, not whether B is
more wonderful than A, but whether there exists an incremental path from A
to B where each step along the path is an immediate improvement. I think
that evolutionary psychologists as a group have gotten a bit farther in
training people to think incrementally than venture capitalists as a group
(present company excepted), but this may just be because I know more about
evolutionary psychology than about venture capital. Anyway, I'm going to
take the EP way of seeing incremental change and apply it to venture capital
in one specific instance, and you can tell me whether the analogy breaks
down.
One of the really major developments I think we could see in the next few
years is the convergence of wireless devices, wearable computing and Segway
Ginger mindreading technology. The Segway Ginger reads tiny, subtle shifts
in your balance, reconstructs what your mind is telling your body, and makes
the Ginger behave like a part of your body. When you're riding a Ginger, it
feels like a subjective extension of your body; they reverse-engineered the
human sense of balance and extended it. (In my opinion, the Ginger is the
greatest innovation in user interface design since Xerox PARC created the
GUI. We live in a strange world where an invention like Ginger opens up an
entire new area of design space for human-computer interaction and everyone
thinks it's a revolution in urban transportation.)
Anyway, one consequence of a Ginger-wireless-wearable revolution could be a
world where your personal wearables track your gaze, read tiny muscular
shifts in your facial muscles, and deduce what you're looking at and what
you want from it. For example, if you look at a chair while walking toward
it, this should provide enough information for a convergent
Ginger-wearable-wireless world to command the chair to slide out from the
table and let you sit down. This is possible in a purely wearable-wireless
world, but you'd trigger a lot of chairs accidentally, the same way that we
currently set off automatic doors if we walk too close to them.
Ginger-wearable-wireless convergence will enable a wide variety of
applications that are dependent on reading the intentions of the user, maybe
in ways that approach the total subjective integration of a Ginger.
Now it seems to me that "Business 2.0" or "Wired" would stop at this point,
having presented the amazing vision of a world that magically responds to
our desires (though they'd probably go on for a lot longer about the vision
as such). But I think the Ginger-wireless-wearable convergence is not only
sexy, but actually plausible as a real-world outcome once wearable wireless
gets going; Ginger-wireless-wearable is not only "adaptive" but
"evolvable". Why? Well, let me beg your indulgence as I construct an
extended analogy between evolution and venture capital...
One of the things I didn't get a chance to discuss explicitly in the current
draft of "Levels of Organization in General Intelligence" is the way that
interdependent subsystems filter themselves out evolutionarily from confused
mechanisms that try and do everything at once. For example, "Levels of
Organization" distinguishes between hardwired sensory modalities and learned
perceptual concepts. But how did we evolve the ability to form learned
perceptual concepts? For learned concepts to be adaptive, you need to link
acquired perceptual categories to optimization of behavior, which takes
reinforcement learning. But what good is reinforcement learning before the
system can acquire perceptual categories? And what good are perceptual
categories without reinforcement learning?
One solution is that behavioral reinforcement can be adaptive even in the
absence of learned perceptual categories - the evolutionary precursor of
reinforcement could have been a circuit that "tuned" an instinct based on
environmental input. This circuit is adaptive if the environmental
condition is variable over evolutionary time, but more or less persistent
over the life of organism. Once a single reinforcement circuit of this type
existed, it would open up the design space for "tunable instincts" - the
incremental elaboration of more finely tunable instincts would be both
possible and adaptive. Eventually, the process would cross the line into
the creation of de novo "instincts" - acquired perceptual categories linked
to optimization of behavior by reinforcement learning. Once learned
complexity is systemically distinguished from hardwired complexity, there is
design pressure to sort out the sensory modaliites that reflect
environmental invariants from concept-formation brainware that reflects
classes of environmental regularities. (If you haven't read "Levels of
Organization" yet and that all sounded like complete gibberish, I
apologize.)
In summary, the first beginnings of neural circuitry probably conflated
feature structure (sensory modalities), category structure (learned
concepts), and event structure (goal-oriented behavior), into circuitry that
served all three functions at once. Evolution can't create systems that
aren't incrementally adaptive; each neural circuit has to be *immediately*
useful to the organism or the circuit won't evolve. Evolution can't create
a visual cortex in anticipation of adding visual categories and behavioral
reinforcement later on. The first neural circuits must sense, categorize,
and react all at once. But, once these first circuits exist, the
evolutionary elaboration of the circuitry tends to differentiate the
functionality into modular subsystems.
Now, what does this have in common with venture capital? To start with,
there's a loose analogy between evolution having to come up with an
adaptation that works *right away*, and a startup having to come up with a
product that can be manufactured *now* and has an existing market. As far
as I can tell, these forms of thinking are both very hard for humans to
learn because of the innate human tendency to be seduced by the goal instead
of the pathway. Unlike venture capital, though, evolutionary psychology has
a precise population-genetics theory of incremental change, with plenty of
specific examples, to hammer the student's mind into the shape required; and
moreover, evolutionary psychology doesn't make you learn the ropes in the
middle of a marketing frenzy with millions of dollars at stake. This is why
it seems to me - from my admittedly loose acquaintance with venture capital
- that evolutionary psychologists have mastered incremental thinking at a
higher level of precision than prevails in venture-capital culture today.
(Present company again excepted.)
So it's possible that learning evolutionary psychology and neo-Darwinian
population genetics might help venture capitalists or startup leaders learn
how to focus on incremental paths and distentangle themselves from seductive
end-goals - at least in those cases where the venture capitalists or startup
leaders are techies who can easily learn the science involved, and the
important point is hammering away at that emotional attachment to end-goal
thinking. But I think the analogy runs deeper than this; I think
evolutionary psychology's way of seeing incremental pathways can be directly
applied to venture capital.
Let's go back to the alleged convergence of Ginger mindreading, wireless
communication, and wearable gadgetry. Looking at this futuristic scenario
through the special lens of evolutionary psychology, it seems to me not only
sexy, but also evolvable. In the beginning of the wearable revolution,
manufacturers will build wearable gadgets that combine the functions of
accepting user input and carrying out some useful behavior. But the
wearable revolution is also the wireless revolution. Once you build a
gadget M that accepts user input and performs a behavior, someone else can
build a gadget N which relies on wireless information from M's user input
and carries out N's special behavior, or someone can build a gadget O that
provides a different user interface commanding M's behavior. The next step
is to dispense with M and buy N and O separately.
In the beginning of the wearables era, people will need to build gadgets
that both accept user input and do something useful - but once everyone has
those combined gadgets, if it's a wearable *wireless* revolution, the
gadgets will start to sort themselves out into "accepters of user input" and
"implementers of functions". Once you're wearing any gadget, anywhere, that
accepts a general kind of user input, there's a market for a pure
function-implementer that relies on an existing gadget for user input. Once
you're wearing any gadget, anywhere, that implements a function, there's a
market for a specialized gadget that takes over the job of getting user
inputs. Eventually, the user-input-accepters and function-implementers may
become completely different gadgets.
I think the integration of Ginger mindreading functionality into wireless
wearables will be the incremental result of accumulating more and more
specialized gadgets that read more and more information from the user. In
turn, these mindreading gadgets open up the design space for more and more
useful functions that depend on mindreading. There is a fairly precise
analogy between the way I'm proposing that all-in-one wireless wearables
will open up the design space for specialized interdependent wireless
wearables, and the example I gave earlier for how all-in-one neural circuits
open up the design space for interdependent brain subsytems. One might even
argue that the shift to wireless communication in devices is analogous to
the shift to nervous systems in body plans, but this doesn't feel to me like
a strong analogy, although it might be a good marketing point as long as
you're marketing to evolutionary biologists.
Okay, so there's an interesting evolutionary analogy for one possible
scenario in the development of the wireless wearable industry. Is this
analogy actually useful to venture capitalists? Does it help someone build
a profitable company? I'm going to try and take the argument a bit farther,
but at this point, I'm taking leave of my area of expertise completely, so
feel free to zap me if I say anything really silly.
One prediction of the evolutionary analogy is that the wireless wearable
revolution is more likely to actually happen, and that it is more likely to
be exciting when it happens. It says that this revolution has more than a
gloriously sexy vision of the endpoint; it also has an incremental path for
getting there, where each success by one company will open the door to
further successes. I'm not sure we'll ever see a repeat of the dot-com
frenzy, since I don't think people are going to let valuations get that
fluffy again, but there could still be a more restrained wireless frenzy in
a few years. And if wireless devices begin to interoperate the way I
expect, it suggests to me that that there will be a lot of acquisitions
going on in this field. So if you fund a wireless wearable company today,
and their gadget does something that could make a nice component in someone
else's integrated wearable system, then there's a decent chance the startup
will mature during an industry frenzy and get snapped up by another company
even if it isn't strong enough to go IPO.
Of course, evolutionary psychology also teaches us that there can easily be
an appealing incremental path that just doesn't get taken - I can argue that
the potential exists for a certain kind of wireless frenzy, but I don't know
whether the industry will actually go there. Another thing that the
evolutionary ecology and the business ecology have in common is that their
histories are composed of what actually did happen, not what should have
happened or could have happened (another way of thinking that's hard for
humans to learn). As I'm not actually joining a wireless wearables startup,
I have no reason to get stuck on this particular vision of the future.
But let's try and extend the analogy even further. Supposing that the
wireless frenzy occurs as described above, let's apply incremental thinking
to ask what past or present industry the wireless wearable industry will
most resemble. Someone marketing a wireless startup would undoubtedly make
suggestive analogies between wireless computing and the meteoric rise of the
Internet, but is this a strong analogy? Does building one website make it
easier for someone else to make another website because of functional
interdependence? This happens occasionally, but not very often. This was
the big flaw in the "Internet portal" theory - the portal fad was based on a
fundamental misunderstanding of Internet ontology. Packaging a lot of
stores into a mall only makes things easier if you're building a *physical*
mall and people have to travel to a physical location. Everything in
Internet space is exactly one URL away from wherever you are. In Internet
space, the inherent pressure isn't toward putting everything in one
convenient place, but toward differentiation into a lot of specialized sites
that do one thing very well. This trend only reverses when there's a rare
synergy between two sites. The really big portals like Yahoo! have
accumulated enough coincidental synergy to keep going, but most portals
added no value and flopped. They were built to appeal to venture
capitalists and stock buyers, not Internet users. Similarly, giving away
product to build "market share" again rests on a misunderstanding of
Internet ontology - "market share" is a kind of inertia, and the Internet
tends toward inertialessness. When I want to buy something online, I use
MySimon or Yahoo! Shopping or NexTag to find the current lowest price; I
have no reason to shop at the places I've shopped before.
So what's the ontology of wireless gadgets? Two other obvious analogies,
aside from the Internet, would be (a) the consumer software market and (b)
Apple's failed attempt at OpenDoc applications built from combinatorial
components. The reason OpenDoc never took off, IMHO, was that developers
just found it easier to develop their own software than to use someone
else's - the "Not Invented Here" syndrome combined with the programmatic
difficulties of reusable components. The same holds true with the consumer
software market; there are a few major applications that need to
communicate, but the overall tendency is to integrate more and more
functions into the same piece of software, and there's nothing to stop
developers from doing so.
The same thing could happen to the wireless wearable industry. Sharing
functionality could just be too hard on a software level (a la OpenDoc).
Manufacturers could find it easier to build everything into one gadget, or
sell a suite of integrated gadgets that work well together but don't work
with anything else (a la consumer applications). These are both ways that
the wireless revolution could turn out very differently from the picture I
sketched earlier. But since the wireless revolution is also the wearable
revolution, another possibility is that conserving space, conserving weight,
and conserving the manufacturing cost of physical gadgetry will create a
strong pressure toward reusing the functionality you're already wearing.
There's another industry where pressure to reuse existing functionality is
strong, and it's not the dot-com industry or the consumer software industry;
it's the enterprise business software industry. It looks to me like those
forces which I expect to operate in the wireless industry are the ones
already operating in the enterprise software industry. There's the same
balance between the real need to be interoperable with other players'
products as a selling point, and the temptation to be interoperable only
with your own products to force people to buy whole suites. There are
standards issues and compatibility issues, but in enterprise software,
companies sometimes manage to build products that work together despite it
all.
So to me this suggests that you could look at the really successful players
in the enterprise software market, ask how they got that way, then invest in
wireless companies that look poised to follow a similar path today.
Unfortunately I have no idea who today's big players in enterprise software
are, or what they looked like as startups, but that's the result I get when
I pursue the line of reasoning to its conclusion. One could also suggest to
wireless startup leaders that they study up on the past history and present
rules of the enterprise software market, or better yet, you could look for a
wireless startup with management experience in marketing enterprise
software. (And no, I don't know offhand of any wireless startups like that.)
Of course, the complete analogy is a long line of reasoning involving
several chained dependencies. There is a known (and studied) human tendency
to overestimate conjunctive probabilities; even if each step in a line of
reasoning is highly probable, the net probability of everything happening as
described may still be relatively small; humans, even experts, are likely to
intuitively overestimate the probability in such a case. A strong argument
would list more than one possible incremental path to the same end result,
and would try to demonstrate that the paths are convergent attractors, where
the industry will settle back into the groove even if perturbed.
Unfortunately, there's no reason why the future should be that predictable,
even in retrospect. If there's a "strong argument" for
Ginger-wearable-wireless, as opposed to a "plausibility argument", my
present knowledge of wireless wearables and venture capital is too meager to
find it. Then again, even the entire chain of logic from wireless gadget
differentiation to the enterprise software analogy still seems less tenuous
than a lot of arguments I heard during the dot-com era - though of course it
may be unfair to judge the startup industry by its frothiest moments.
What I'd really like to do at this point is apply a seed AI perspective and
figure out some way that venture capital can take larger strides through the
fitness landscape instead of being limited to incremental steps of four to
six years. Unfortunately, the main thing I can think of offhand would be
some way to move faster from collecting the venture fund to making the
actual investments, or some way of not collecting the committed funds until
the investments are ready - freeing up the venture fund's lifespan so that
more lifespan is available to startups. But that's all I can think of at
the moment and it seems likely that venture funds are doing that already, so
I guess the long-term stuff will go on being handled by individual
investors.
-- 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 - 04:00:40 MDT