X-Message-Number: 0019.3
Subject: The Technical Feasibility of Cryonics; Part #3
Newsgroups: sci.cryonics
From: (Ralph Merkle)
Subject: The Technical Feasibility of Cryonics; Part #3
Date: 22 Nov 92 21:15:56 GMT
The Technical Feasibility of Cryonics
PART 3 of 5.
by
Ralph C. Merkle
Xerox PARC
3333 Coyote Hill Road
Palo Alto, CA 94304
A shorter version of this article appeared in:
Medical Hypotheses (1992) 39, pages 6-16.
----------------------------------------------------------
TECHNICAL OVERVIEW
Even if information theoretic death has not occurred, a frozen brain is
not a healthy structure. While repair might be feasible in principle,
it would be comforting to have at least some idea about how such repairs
might be done in practice. As long as we assume that the laws of
physics, chemistry, and biochemistry with which we are familiar today
will still form the basic framework within which repair will take place
in the future, we can draw well founded conclusions about the
capabilities and limits of any such repair technology.
The Nature of This Proposal
To decide whether or not to pursue cryonic suspension we must answer one
question: will restoration of frozen tissue to a healthy and functional
state ever prove feasible? If the answer is "yes," then cryonics will
save lives. If the answer is "no," then it can be ignored. As
discussed earlier, the most that we can usefully learn about frozen
tissue is the type, location and orientation of each molecule. If this
information is sufficient to permit inference of the healthy state with
memory and personality intact, then repair is in principle feasible.
The most that future technology could offer, therefore, is the ability
to restore the structure whenever such restoration was feasible in
principle. We propose that just this limit will be closely approached
by future advances in technology.
It is unreasonable to think that the current proposal will in fact form
the basis for future repair methods for two reasons:
First, better technologies and approaches are likely to be developed.
Necessarily, we must restrict ourselves to methods and techniques that
can be analyzed and understood using the currently understood laws of
physics and chemistry. Future scientific advances, not anticipated at
this time, are likely to result in cheaper, simpler or more reliable
methods. Given the history of science and technology to date, the
probability of future unanticipated advances is good.
Second, this proposal was selected because of its conceptual simplicity
and its obvious power to restore virtually any structure where
restoration is in principle feasible. These are unlikely to be design
objectives of future systems. Conceptual simplicity is advantageous
when the resources available for the design process are limited. Future
design capabilities can reasonably be expected to outstrip current
capabilities, and the efforts of a large group can reasonably be
expected to allow analysis of much more complex proposals than
considered here.
Further, future systems will be designed to restore specific individuals
suffering from specific types of damage, and can therefore use specific
methods that are less general but which are more efficient or less
costly for the particular type of damage involved. It is easier for a
general-purpose proposal to rely on relatively simple and powerful
methods, even if those methods are less efficient.
Why, then, discuss a powerful, general purpose method that is
inefficient, fails to take advantage of the specific types of damage
involved, and which will almost certainly be superseded by future
technology?
The purpose of this paper is not to lay the groundwork for future
systems, but to answer a question: under what circumstances will
cryonics work? The value of cryonics is clearly and decisively based on
technical capabilities that will not be developed for several decades
(or longer). If some relatively simple proposal appears likely to work,
then the value of cryonics is established. Whether or not that simple
proposal is actually used is irrelevant. The fact that it could be used
in the improbable case that all other technical progress and all other
approaches fail is sufficient to let us decide today whether or not
cryonic suspension is of value.
The philosophical issues involved in this type of long range technical
forecasting and the methodologies appropriate to this area are addressed
by work in "exploratory engineering."[1] The purpose of exploratory
engineering is to provide lower bounds on future technical capabilities
based on currently understood scientific principles. A successful
example is Konstantin Tsiolkovsky's forecast around the turn of the
century that multi-staged rockets could go to the moon. His forecast
was based on well understood principles of Newtonian mechanics. While
it did not predict when such flights would take place, nor who would
develop the technology, nor the details of the Saturn V booster, it did
predict that the technical capability was feasible and would eventually
be developed. In a similar spirit, we will discuss the technical
capabilities that should be feasible and what those capabilities should
make possible.
Conceptually, the approach that we will follow is simple:
1.) Determine the coordinates and orientations of all major molecules,
and store this information in a data base.
2.) Analyze the information stored in the data base with a computer
program which determines what changes in the existing structure should
be made to restore it to a healthy and functional state.
3.) Take the original molecules and move them, one at a time, back to
their correct locations.
The reader will no doubt agree that this proposal is conceptually simple
and remarkably powerful, but might be concerned about a number of
technical issues. The major issues are addressed in the following
analysis.
An obvious inefficiency of this approach is that it will take apart and
then put back together again structures and whole regions that are in
fact functional or only slightly damaged. Simply leaving a functional
region intact, or using relatively simple special case repair methods
for minor damage would be faster and less costly. Despite these obvious
drawbacks, the general purpose approach demonstrates the principles
involved. As long as the inefficiencies are not so extreme that they
make the approach infeasible or uneconomical in the long run, then this
simpler approach is easier to evaluate.
Overview of the Brain.
The brain has a volume of 1350 cubic centimeters (about one and a half
quarts) and a weight of slightly more than 1400 grams (about three
pounds). The smallest normal human brain weighed 1100 grams, while the
largest weighed 2050 grams [30, page 24]. It is almost 80% water by
weight. The remaining 20% is slightly less than 40% protein, slightly
over 50% lipids, and a few percent of other material[16, page 419].
Thus, an average brain has slightly over 100 grams of protein, about 175
grams of lipids, and some 30 to 40 grams of "other stuff".
How Many Molecules
If we are considering restoration down to the molecular level, an
obvious question is: how many molecules are there? We can easily
approximate the answer, starting with the proteins. An "average"
protein molecule has a molecular weight of about 50,000 amu. One mole
of "average" protein is 50,000 grams (by definition), so the 100 grams
of protein in the brain is 100/50,000 or .002 moles. One mole is 6.02 x
10^23 molecules, so .002 moles is 1.2 x 10^21 molecules.
We proceed in the same way for the lipids (lipids are most often used to
make cell membranes) - a "typical" lipid might have a molecular weight
of 500 amu, which is 100 times less than the molecular weight of a
protein. This implies the brain has about 175/500 x 6.02 x 10^23 or
about 2 x 10^23 lipid molecules.
Finally, water has a molecular weight of 18, so there will be about 1400
x 0.8/18 x 6.02 x 10^23 or about 4 x 10^25 water molecules in the brain.
In many cases a substantial percentage of water will have been replaced
with cryoprotectant during the process of suspension; glycerol at a
concentration of 4 molar or more, for example. Both water and glycerol
will be treated in bulk, and so the change from water molecules to
glycerol (or other cryoprotectants) should not have a significant impact
on the calculations that follow.
These numbers are fundamental. Repair of the brain down to the
molecular level will require that we cope with them in some fashion.
How Much Time
Another parameter whose value we must decide is the amount of repair
time per molecule. We assume that such repair time includes the time
required to determine the location of the molecule in the frozen tissue
and the time required to restore the molecule to its correct location,
as well as the time to diagnose and repair any structural defects in the
molecule. The computational power required to analyze larger-scale
structural damage - e.g., this mitochondria has suffered damage to its
internal membrane structure (so called "flocculant densities") - should
be less than the power required to analyze each individual molecule. An
analysis at the level of sub-cellular organelles involves several orders
of magnitude fewer components and will therefore require correspondingly
less computational power. Analysis at the cellular level involves even
fewer components. We therefore neglect the time required for these
additional computational burdens. The total time required for repair is
just the sum over all molecules of the time required by one repair
device to repair that molecule divided by the number of repair devices.
The more repair devices there are, the faster the repair will be. The
more molecules there are, and the more time it takes to repair each
molecule, the slower repair will be.
The time required for a ribosome to manufacture a protein molecule of
400 amino acids is about 10 seconds[14, page 393], or about 25
milliseconds to add each amino acid. DNA polymerase III can add an
additional base to a replicating DNA strand in about 7 milliseconds[14,
page 289]. In both cases, synthesis takes place in solution and
involves significant delays while the needed components diffuse to the
reactive sites. The speed of assembler-directed reactions is likely to
prove faster than current biological systems. The arm of an assembler
should be capable of making a complete motion and causing a single
chemical transformation in about a microsecond[85]. However, we will
conservatively base our computations on the speed of synthesis already
demonstrated by biological systems, and in particular on the slower
speed of protein synthesis.
We must do more than synthesize the required molecules - we must analyze
the existing molecules, possibly repair them, and also move them from
their original location to the desired final location. Existing
antibodies can identify specific molecular species by selectively
binding to them, so identifying individual molecules is feasible in
principle. Even assuming that the actual technology employed is
different it seems unlikely that such analysis will require
substantially longer than the synthesis time involved, so it seems
reasonable to multiply the synthesis time by a factor of a few to
provide an estimate of time spent per molecule. This should, in
principle, allow time for the complete disassembly and reassembly of the
selected molecule using methods no faster than those employed in
biological systems. While the precise size of this multiplicative
factor can reasonably be debated, a factor of 10 should be sufficient.
The total time required to simply move a molecule from its original
location to its correct final location in the repaired structure should
be smaller than the time required to disassemble and reassemble it, so
we will assume that the total time required for analysis, repair and
movement is 100 seconds per protein molecule.
Temperature of Analysis
Warming the tissue before determining its molecular structure creates
definite problems: everything will move around. A simple solution to
this problem is to keep the tissue frozen until after all the desired
structural information is recovered. In this case the analysis will
take place at a low temperature. Whether or not subsequent operations
should be performed at the same low temperature is left open. A later
section considers the various approaches that can be taken to restore
the structure after it has been analyzed.
Repair or Replace?
In practice, most molecules will probably be intact - they would not
have to be either disassembled or reassembled. This should greatly
reduce repair time. On a more philosophical note, existing biological
systems generally do not bother to repair macromolecules (a notable
exception is DNA - a host of molecular mechanisms for the repair of this
molecule are used in most organisms). Most molecules are generally used
for a period of time and then broken down and replaced. There is a slow
and steady turnover of molecular structure - the atoms in the roast beef
sandwich eaten yesterday are used today to repair and replace muscles,
skin, nerve cells, etc. If we adopted nature's philosophy we would
simply discard and replace any damaged molecules, greatly simplifying
molecular "repair".
Carried to its logical conclusion, we would discard and replace all the
molecules in the structure. Having once determined the type, location
and orientation of a molecule in the original (frozen) structure, we
would simply throw that molecule out without further examination and
replace it. This requires only that we be able to identify the
location and type of individual molecules. It would not be necessary to
determine if the molecule was damaged, nor would it be necessary to
correct any damage found. By definition, the replacement molecule would
be taken from a stock-pile of structurally correct molecules that had
been previously synthesized, in bulk, by the simplest and most
economical method available.
Discarding and replacing even a few atoms might disturb some people.
This can be avoided by analyzing and repairing any damaged molecules.
However, for those who view the simpler removal and replacement of
damaged molecules as acceptable, the repair process can be significantly
simplified. For purposes of this paper, however, we will continue to
use the longer time estimate based on the premise that full repair of
every molecule is required. This appears to be conservative. (Those
who feel that replacing their atoms will change their identity should
think carefully before eating their next meal!)
Total Repair Machine Seconds
We shall assume that the repair time for other molecules is similar per
unit mass. That is, we shall assume that the repair time for the lipids
(which each weigh about 500 amu, 100 times less than a protein) is about
100 times less than the repair time for a protein. The repair time for
one lipid molecule is assumed to be 1 second. We will neglect water
molecules in this analysis, assuming that they can be handled in bulk.
We have assumed that the time required to analyze and synthesize an
individual molecule will dominate the time required to determine its
present location, the time required to determine the appropriate
location it should occupy in the repaired structure, and the time
required to put it in this position. These assumptions are plausible
but will be considered further when the methods of gaining access to and
of moving molecules during the repair process are considered.
This analysis accounts for the bulk of the molecules - it seems unlikely
that other molecular species will add significant additional repair
time.
Based on these assumptions, we find that we require 100 seconds x 1.2 x
10^21 protein molecules + 1 second times 2 x 10^23 lipids, or 3.2 x
10^23 repair-machine-seconds. This number is not as fundamental as the
number of molecules in the brain. It is based on the (probably
conservative) assumption that repair of 50,000 amu requires 100 seconds.
Faster repair would imply repair could be done with fewer repair
machines, or in less time.
How Many Repair Machines
If we now fix the total time required for repair, we can determine the
number of repair devices that must function in parallel. We shall
rather arbitrarily adopt 10^8 seconds, which is very close to three
years, as the total time in which we wish to complete repairs.
If the total repair time is 10^8 seconds, and we require 3.2 x 10^23
repair-machine-seconds, then we require 3.2 x 10^15 repair machines for
complete repair of the brain. This corresponds to 3.2 x 10^15 / (6.02
x 10^23) or 5.3 x 10^-9 moles, or 5.3 nanomoles of repair machines. If
each repair device weighs 10^9 to 10^10 amu, then the total weight of
all the repair devices is 53 to 530 grams: a few ounces to just over a
pound.
Thus, the weight of repair devices required to repair each and every
molecule in the brain, assuming the repair devices operate no faster
than current biological methods, is about 4% to 40% of the total mass of
the brain.
By way of comparision, there are about 10^14 cells[44, page 3] in the
human body and each cell has about 10^7 ribosomes[14, page 652] giving
10^21 ribosomes. Thus, there are about six orders of magnitude more
ribosomes in the human body than the number of repair machines we
estimate are required to repair the human brain.
It seems unlikely that either more or larger repair devices are
inherently required. However, it is comforting to know that errors in
these estimates of even several orders of magnitude can be easily
tolerated. A requirement for 530 kilograms of repair devices (1,000 to
10,000 times more than we calculate is needed) would have little
practical impact on feasibility. Although repair scenarios that involve
deployment of the repair devices within the volume of the brain could
not be used if we required 530 kilograms of repair devices, a number of
other repair scenarios would still work - one such approach is discussed
in this paper. Given that nanotechnology is feasible, manufacturing
costs for repair devices will be small. The cost of even 530 kilograms
of repair devices should eventually be significantly less than a few
hundred dollars. The feasibility of repair down to the molecular level
is insensitive to even large errors in the projections given here.
THE REPAIR PROCESS
We now turn to the physical deployment of these repair devices. That
is, although the raw number of repair devices is sufficient, we must
devise an orderly method of deploying these repair devices so they can
carry out the needed repairs.
Other Proposals: On-board Repair
We shall broadly divide repair scenarios into two classes: on-board and
off-board. In the on-board scenarios, the repair devices are deployed
within the volume of the brain. Existing structures are disassembled in
place, their component molecules examined and repaired, and rebuilt on
the spot. (We here class as "on-board" those scenarios in which the
repair devices operate within the physical volume of the brain, even
though there might be substantial off-board support. That is, there
might be a very large computer outside the tissue directing the repair
process, but we would still refer to the overall repair approach as "on-
board"). The on-board repair scenario has been considered in some
detail by Drexler[18]. We will give a brief outline of the on-board
repair scenario here, but will not consider it in any depth. For
various reasons, it is quite plausible that on-board repair scenarios
will be developed before off-board repair scenarios.
The first advantage of on-board repair is an easier evolutionary path
from partial repair systems deployed in living human beings to the total
repair systems required for repair of the more extensive damage found in
the person who has been cryonically suspended. That is, a simple repair
device for finding and removing fatty deposits blocking the circulatory
system could be developed and deployed in living humans[2], and need not
deal with all the problems involved in total repair. A more complex
device, developed as an incremental improvement, might then repair more
complex damage (perhaps identifying and killing cancer cells) again
within a living human. Once developed, there will be continued pressure
for evolutionary improvements in on-board repair capabilities which
should ultimately lead to repair of virtually arbitrary damage. This
evolutionary path should eventually produce a device capable of
repairing frozen tissue.
It is interesting to note that "At the end of this month [August 1990],
MITI's Agency of Industrial Science and Technology (AIST) will submit a
budget request for 430 million ($200,000) to launch a 'microrobot'
project next year, with the aim of developing tiny robots for the
internal medical treatment and repair of human beings. ... MITI is
planning to pour 425,000 million ($170 million) into the microrobot
project over the next ten years..."[86]. Iwao Fujimasa said their
objective is a robot less than .04 inches in size that will be able to
travel through veins and inside organs[17, 20]. While substantially
larger than the proposals considered here, the direction of future
evolutionary improvements should be clear.
A second advantage of on-board repair is emotional. In on-board repair,
the original structure (you) is left intact at the macroscopic and even
light microscopic level. The disassembly and reassembly of the
component molecules is done at a level smaller than can be seen, and
might therefore prove less troubling than other forms of repair in which
the disassembly and reassembly processes are more visible. Ultimately,
though, correct restoration of the structure is the overriding concern.
A third advantage of on-board repair is the ability to leave functional
structures intact. That is, in on-board repair we can focus on those
structures that are damaged, while leaving working structures alone. If
minor damage has occured, then an on-board repair system need make only
minor repairs.
The major drawback of on-board repair is the increased complexity of the
system. As discussed earlier, this is only a drawback when the design
tools and the resources available for the design are limited. We can
reasonably presume that future design tools and future resources will
greatly exceed present efforts. Developments in computer aided design
of complex systems will put the design of remarkably complex systems
within easy grasp.
In on-board repair, we might first logically partition the volume of the
brain into a matrix of cubes, and then deploy each repair device in its
own cube. Repair devices would first get as close as possible to their
assigned cube by moving through the circulatory system (we presume it
would be cleared out as a first step) and would then disassemble the
tissue between them and their destination. Once in position, each
repair device would analyse the tissue in its assigned volume and peform
any repairs required.
The Current Proposal: Off-Board Repair
The second class of repair scenarios, the off-board scenarios, allow the
total volume of repair devices to greatly exceed the volume of the human
brain.
The primary advantage of off-board repair is conceptual simplicity. It
employees simple brute force to insure that a solution is feasible and
to avoid complex design issues. As discussed earlier, these are
virtures in thinking about the problem today but are unlikely to carry
much weight in the future when an actual system is being designed.
The other advantages of this approach are fairly obvious. Lingering
concerns about volume and heat dissipation can be eliminated. If a ton
of repair devices should prove necessary, then a ton can be provided.
Concerns about design complexity can be greatly reduced. Off-board
repair scenarios do not require that the repair devices be mobile -
simplifying communications and power distribution, and eliminating the
need for locomotor capabilities and navigational abilities. The only
previous paper on off-board repair scenarios was by Merkle[101].
Off-board repair scenarios can be naturally divided into three phases.
In the first phase, we must analyze the structure to determine its
state. The primary purpose of this phase is simply to gather
information about the structure, although in the process the disassembly
of the structure into its component molecules will also take place.
Various methods of gaining access to and analyzing the overall structure
are feasible - in this paper we shall primarily consider one approach.
We shall presume that the analysis phase takes place while the tissue is
still frozen. While the exact temperature is left open, it seems
preferable to perform analysis prior to warming. The thawing process
itself causes damage and, once thawed, continued deterioration will
proceed unchecked by the mechanisms present in healthy tissue. This
cannot be tolerated during a repair time of several years. Either
faster analysis or some means of blocking deterioration would have to be
used if analysis were to take place after warming. We will not explore
these possibilities here (although this is worthwhile). The temperature
at which other phases takes place is left open.
The second phase of off-board repair is determination of the healthy
state. In this phase, the structural information derived from the
analysis phase is used to determine what the healthy state of the tissue
had been prior to suspension and any preceding illness. This phase
involves only computation based on the information provided by the
analysis phase.
The third phase is repair. In this phase, we must restore the structure
in accordance with the blue-print provided by the second phase, the
determination of the healthy state.
Intermediate States During Off-Board Repair
Repair methods in general start with frozen tissue, and end with healthy
tissue. The nature of the intermediate states characterizes the
different repair approaches. In off-board repair the tissue undergoing
repair must pass through three highly characteristic states, described
in the following three paragraphs.
The first state is the starting state, prior to any repair efforts. The
tissue is frozen (unrepaired).
In the second state, immediately following the analysis phase, the
tissue has been disassembled into its individual molecules. A detailed
structural data base has been built which provides a description of the
location, orientation, and type of each molecule, as discussed earlier.
For those who are concerned that their identity or "self" is dependent
in some fundamental way on the specific atoms which compose their
molecules, the original molecules can be retained in a molecular "filing
cabinet." While keeping physical track of the original molecules is
more difficult technically, it is feasible and does not alter off-board
repair in any fundamental fashion.
In the third state, the tissue is restored and fully functional.
By characterizing the intermediate state which must be achieved during
the repair process, we reduce the problem from "Start with frozen tissue
and generate healthy tissue" to "Start with frozen tissue and generate a
structural data base and a molecular filing cabinet. Take the
structural data base and the molecular filing cabinet and generate
healthy tissue." It is characteristic of off-board repair that we
disassemble the molecular structure into its component pieces prior to
attempting repair.
As an example, suppose we wish to repair a car. Rather than try and
diagnose exactly what's wrong, we decide to take the car apart into its
component pieces. Once the pieces are spread out in front of us, we can
easily clean each piece, and then reassemble the car. Of course, we'll
have to keep track of where all the pieces go so we can reassemble the
structure, but in exchange for this bookkeeping task we gain a
conceptually simple method of insuring that we actually can get access
to everything and repair it. While this is a rather extreme method of
repairing a broken carburetor, it certainly is a good argument that we
should be able to repair even rather badly damaged cars. So, too, with
off-board repair. While it might be an extreme method of fixing any
particular form of damage, it provides a good argument that damage can
be repaired under a wide range of circumstances.
Off-Board Repair is the Best that can be Achieved
Regardless of the initial level of damage, regardless of the functional
integrity or lack thereof of any or all of the frozen structure,
regardless of whether easier and less exhaustive techniques might or
might not work, we can take any frozen structure and convert it into the
canonical state described above. Further, this is the best that we can
do. Knowing the type, location and orientation of every molecule in the
frozen structure under repair and retaining the actual physical
molecules (thus avoiding any philosophical objections that replacing the
original molecules might somehow diminish or negate the individuality of
the person undergoing repair) is the best that we can hope to achieve.
We have reached some sort of limit with this approach, a limit that will
make repair feasible under circumstances which would astonish most
people today.
One particular approach to off-board repair is divide-and-conquer. This
method is one of the technically simplest approaches. We discuss this
method in the following section.
Divide-and-Conquer
Divide-and-conquer is a general purpose problem-solving method
frequently used in computer science and elsewhere. In this method, if a
problem proves too difficult to solve it is first divided into sub-
problems, each of which is solved in turn. Should the sub-problems
prove too difficult to solve, they are in turn divided into sub-sub-
problems. This process is continued until the original problem is
divided into pieces that are small enough to be solved by direct
methods.
If we apply divide-and-conquer to the analysis of a physical object -
such as the brain - then we must be able to physically divide the object
of analysis into two pieces and recursively apply the same method to the
two pieces. This means that we must be able to divide a piece of
frozen tissue, whether it be the entire brain or some smaller part, into
roughly equal halves. Given that tissue at liquid nitrogen temperatures
is already prone to fracturing, it should require only modest effort to
deliberately induce a fracture that would divide such a piece into two
roughly equal parts. Fractures made at low temperatures (when the
material is below the glass transition temperature) are extremely clean,
and result in little or no loss of structural information. Indeed,
freeze fracture techniques are used for the study of synaptic
structures. Hayat [40, page 398] says "Membranes split during freeze-
fracturing along their central hydrophobic plane, exposing
intramembranous surfaces. ... The fracture plane often follows the
contours of membranes and leaves bumps or depressions where it passes
around vesicles and other cell organelles. ... The fracturing process
provides more accurate insight into the molecular architecture of
membranes than any other ultrastructural method." It seems unlikely
that the fracture itself will result in any significant loss of
structural information.
The freshly exposed faces can now be analyzed by various surface
analysis techniques. A review article in Science, "The Children of the
STM," supports the idea that such surface analysis techniques can
recover remarkably detailed information. For example, optical
absorption microscopy "...generates an absorption spectrum of the
surface with a resolution of 1 nanometer [a few atomic diameters]."
Science quotes Kumar Wickramasinghe of IBM's T. J. Watson Research
Center as saying: "We should be able to record the spectrum of a single
molecule" on a surface. Williams and Wickramasinghe said [51] "The
ability to measure variations in chemical potential also allows the
possibility of selectively identifying subunits of biological
macromolecules either through a direct measurement of their chemical-
potential gradients or by decorating them with different metals. This
suggest a potentially simple method for sequencing DNA." Several other
techniques are discussed in the Science article. While current devices
are large, the fundamental physical principles on which they rely do not
require large size. Many of the devices depend primarily on the
interaction between a single atom at the tip of the STM probe and the
atoms on the surface of the specimen under analysis. Clearly,
substantial reductions in size in such devices are feasible[ft. 18].
High resolution optical techniques can also be employed. Near field
microscopy, employing light with a wavelength of hundreds of nanometers,
has achieved a resolution of 12 nanometers (much smaller than a
wavelength of light). To quote the abstract of a recent review article
on the subject: "The near-field optical interaction between a sharp
probe and a sample of interest can be exploited to image,
spectroscopically probe, or modify surfaces at a resolution (down to ~12
nm) inaccessible by traditional far-field techniques. Many of the
attractive features of conventional optics are retained, including
noninvasiveness, reliability, and low cost. In addition, most optical
contrast mechanisms can be extended to the near-field regime, resulting
in a technique of considerable versatility. This versatility is
demonstrated by several examples, such as the imaging of nanometric-
scale features in mammalian tissue sections and the creation of
ultrasmall, magneto-optic domains having implications for high-density
data storage. Although the technique may find uses in many diverse
fields, two of the most exciting possibilities are localized optical
spectroscopy of semiconductors and the flourescence imaging of living
cells."[111]. Another article said: "Our signals are currently of such
magnitude that almost any application originally conceived for far-field
optics can now be extended to the near-field regime, including:
dynamical studies at video rates and beyond; low noise, high resolution
spectroscopy (also aided by the negligible auto-fluorescence of the
probe); minute differential absorption measurements; magnetooptics; and
superresolution lithography."[100].
How Small are the Pieces
The division into halves continues until the pieces are small enough to
allow direct analysis by repair devices. If we presume that division
continues until each repair device is assigned its own piece to repair,
then there will be both 3.2 x 10^15 repair devices and pieces. If the
1350 cubic centimeter volume of the brain is divided into this many
cubes, each such cube would be about .4 microns (422 nanometers) on a
side. Each cube could then be directly analyzed (disassembled into its
component molecules) by a repair device during our three-year repair
period.
One might view these cubes as the pieces of a three-dimensional jig-saw
puzzle, the only difference being that we have cheated and carefully
recorded the position of each piece. Just as the picture on a jig-saw
puzzle is clearly visible despite the fractures between the pieces, so
too the three-dimensional "picture" of the brain is clearly visible
despite its division into pieces[ft. 19].
Moving Pieces
There are a great many possible methods of handling the mechanical
problems involved in dividing and moving the pieces. It seems unlikely
that mechanical movement of the pieces will prove an insurmountable
impediment, and therefore we do not consider it in detail. However, for
the sake of concreteness, we outline one possibility. Human arms are
about 1 meter in length, and can easily handle objects from 1 to 10
centimeters in size (.01 to .1 times the length of the arm). It should
be feasible, therefore, to construct a series of progressively shorter
arms which handle pieces of progressively smaller size. If each set of
arms were ten times shorter than the preceding set, then we would have
devices with arms of: 1 meter, 1 decimeter, 1 centimeter, 1 millimeter,
100 microns, 10 microns, 1 micron, and finally .1 microns or 100
nanometers. (Note that an assembler has arms roughly 100 nanometers
long). Thus, we would need to design 8 different sizes of manipulators.
At each succeeding size the manipulators would be more numerous, and so
would be able to deal with the many more pieces into which the original
object was divided. Transport and mechanical manipulation of an object
would be done by arms of the appropriate size. As objects were divided
into smaller pieces that could no longer be handled by arms of a
particular size, they would be handed to arms of a smaller size.
If it requires about three years to analyze each piece, then the time
required both to divide the brain into pieces and to move each piece to
an immobile repair device can reasonably be neglected. It seems
unlikely that moving the pieces will take a significant fraction of
three years.
Memory Requirements
The information storage requirements for a structural data-base that
holds the detailed description and location of each major molecule in
the brain can be met by projected storage methods. DNA has an
information storage density of about 10^21 bits/cubic centimeter.
Conceptually similar but somewhat higher density molecular "tape"
systems that store 10^22 bits/cubic centimeter [1] should be quite
feasible. If we assume that every lipid molecule is "significant" but
that water molecules, simple ions and the like are not, then the number
of significant molecules is roughly the same as the number of lipid
molecules[ft. 20] (the number of protein molecules is more than two
orders of magnitude smaller, so we will neglect it in this estimate).
The digital description of these 2 x 10^23 significant molecules
requires 10^25 bits (assuming that 50 bits are required to encode the
location and description of each molecule). This is about 1,000 cubic
centimeters (1 liter, roughly a quart) of "tape" storage. If a storage
system of such capacity strikes the reader as infeasible, consider that
a human being has about 10^14 cells[44, page 3] and that each cell
stores 10^10 bits in its DNA[14]. Thus, every human that you see is a
device which (among other things) has a raw storage capacity of 10^24
bits - and human beings are unlikely to be optimal information storage
devices.
A simple method of reducing storage requirements by several orders of
magnitude would be to analyze and repair only a small amount of tissue
at a time. This would eliminate the need to store the entire 10^25 bit
description at one time. A smaller memory could hold the description of
the tissue actually under repair, and this smaller memory could then be
cleared and re-used during repair of the next section of tissue.
Computational Requirements
The computational power required to analyze a data base with 10^25 bits
is well within known theoretical limits[9,25,32]. It has been seriously
proposed that it might be possible to increase the total computational
power achievable within the universe beyond any fixed bound in the
distant future[52, page 658]. More conservative lower bounds to nearer-
term future computational capabilities can be derived from the
reversible rod-logic molecular model of computation, which dissipates
about 10^-23 joules per gate operation when operating at 100 picoseconds
at room temperature[85]. A wide range of other possibilities exist.
Likharev proposed a computational element based on Josephson junctions
which operates at 4 K and in which energy dissipation per switching
operation is 10^-24 joules with a switching time of 10^-9 seconds[33,
43]. Continued evolutionary reductions in the size and energy
dissipation of properly designed NMOS[113] and CMOS[112] circuits should
eventually produce logic elements that are both very small (though
significantly larger than Drexler's proposals) and which dissipate
extraordinarily small amounts of energy per logic operation.
Extrapolation of current trends suggest that energy dissipations in the
10-23 joule range will be achieved before 2030[31, fig. 1]. There is no
presently known reason to expect the trend to stop or even slow down at
that time[9,32].
Energy costs appear to be the limiting factor in rod logic (rather than
the number of gates, or the speed of operation of the gates). Today,
electric power costs about 10 cents per kilowatt hour. Future costs of
power will almost certainly be much lower. Molecular manufacturing
should eventually sharply reduce the cost of solar cells and increase
their efficiency close to the theoretical limits. With a manufacturing
cost of under 10 cents per kilogram[85] the cost of a one square meter
solar cell will be less than a penny. As a consequence the cost of
solar power will be dominated by other costs, such as the cost of the
land on which the solar cell is placed. While solar cells can be placed
on the roofs of existing structures or in otherwise unused areas, we
will simply use existing real estate prices to estimate costs. Low cost
land in the desert south western United States can be purchased for less
than $1,000 per acre. (This price corresponds to about 25 cents per
square meter, significantly larger than the projected future
manufacturing cost of a one square meter solar cell). Land elsewhere in
the world (arid regions of the Australian outback, for example) is much
cheaper. For simplicity and conservatism, though, we'll simply adopt
the $1,000 per acre price for the following calculations. Renting an
acre of land for a year at an annual price of 10% of the purchase price
will cost $100. Incident sunlight at the earth's surface provides a
maximum of 1,353 watts per square meter, or 5.5 x 10^6 watts per acre.
Making allowances for inefficiencies in the solar cells, atmospheric
losses, and losses caused by the angle of incidence of the incoming
light reduces the actual average power production by perhaps a factor of
15 to about 3.5 x 10^5 watts. Over a year, this produces 1.1 x 10^13
joules or 3.1 x 10^6 kilowatt hours. The land cost $100, so the cost
per joule is 0.9 nanocents and the cost per kilowatt hour is 3.3
millicents. Solar power, once we can make the solar cells cheaply
enough, will be over several thousand times cheaper than electric power
is today. We'll be able to buy over 10^15 joules for under $10,000.
While the energy dissipation per logic operation estimated by
Drexler[85] is about 10^-23 joules, we'll content ourselves with the
higher estimate of 10^-22 joules per logic operation. Our 10^15 joules
will then power 10^37 gate operations: 10^12 gate operations for each
bit in the structural data base or 5 x 10^13 gate operations for each of
the 2 x 10^23 lipid molecules present in the brain.
It should be emphasized that in off-board repair warming of the tissue
is not an issue because the overwhelming bulk of the calculations and
hence almost all of the energy dissipation takes place outside the
tissue. Much of the computation takes place when the original
structure has been entirely disassembled into its component molecules.
How Much Is Enough?
Is this enough computational power? We can get a rough idea of how much
computer power might be required if we draw an analogy from image
recognition. The human retina performs about 100 "operations" per
pixel, and the human brain is perhaps 1,000 to 10,000 times larger than
the retina. This implies that the human image recognition system can
recognize an object after devoting some 10^5 to 10^6 "operations" per
pixel. (This number is also in keeping with informal estimates made by
individuals expert in computer image analysis). Allowing for the fact
that such "retinal operations" are probably more complex than a single
"gate operation" by a factor of 1000 to 10,000, we arrive at 10^8 to
10^10 gate operations per pixel - which is well below our estimate of
10^12 operations per bit or 5 x 10^13 operations per molecule.
To give a feeling for the computational power this represents, it is
useful to compare it to estimates of the raw computational power of the
human brain. The human brain has been variously estimated as being
able to do 10^13[50], 10^15 or 10^16[114] operations a second (where
"operation" has been variously defined but represents some relatively
simple and basic action)[ft. 21]. The 10^37 total logic operations will
support 10^29 logic operations per second for three years, which is the
raw computational power of something like 10^13 human beings (even when
we use the high end of the range for the computational power of the
human brain). This is 10 trillion human beings, or some 2,000 times
more people than currently exist on the earth today. By present
standards, this is a large amount of computational power. Viewed
another way, if we were to divide the human brain into tiny cubes that
were about 5 microns on a side (less than the volume of a typical cell),
each such cube could receive the full and undivided attention of a
dedicated human analyst for a full three years.
The next paragraph analyzes memory costs, and can be skipped without
loss of continuity.
This analysis neglects the memory required to store the complete state
of these computations. Because this estimate of computational abilities
and requirements depends on the capabilities of the human brain, we
might require an amount of memory roughly similar to the amount of
memory required by the human brain as it computes. This might require
about 10^16 bits (10 bits per synapse) to store the "state" of the
computation. (We assume that an exact representation of each synapse
will not be necessary in providing capabilities that are similar to
those of the human brain. At worst, the behavior of small groups of
cells could be analyzed and implemented by the most efficient method,
e.g., a "center surround" operation in the retina could be implemented
as efficiently as possible, and would not require detailed modeling of
each neuron and synapse. In point of fact, it is likely that algorithms
that are significantly different from the algorithms employed in the
human brain will prove to be the most efficient for this rather
specialized type of analysis, and so our use of estimates derived from
low-level parts-counts from the human brain are likely to be very
conservative). For 10^13 programs each equivalent in analytical skills
to a single human being, this would require 10^29 bits. At 100 cubic
nanometers per bit, this gives 10,000 cubic meters. Using the cost
estimates provided by Drexler[85] this would be an uncomfortable
$1,000,000. We can, however, easily reduce this cost by partitioning
the computation to reduce memory requirements. Instead of having 10^13
programs each able to "think" at about the same speed as a human being,
we could have 10^10 programs each able to "think" at a speed 1,000 times
faster than a human being. Instead of having 10 trillion dedicated
human analysts working for 3 years each, we would have 10 billion
dedicated human analysts working for 3,000 virtual years each. The
project would still be completed in 3 calendar years, for each computer
"analyst" would be a computer program running 1,000 times faster than an
equally skilled human analyst. Instead of analyzing the entire brain at
once, we would instead logically divide the brain into 1,000 pieces each
of about 1.4 cubic centimeters in size, and analyze each such piece
fully before moving on to the next piece.
This reduces our memory requirements by a factor of 1,000 and the cost
of that memory to a manageable $1,000.
It should be emphasized that the comparisons with human capabilities are
used only to illustrate the immense capabilities of 10^37 logic
operations. It should not be assumed that the software that will
actually be used will have any resemblance to the behavior of the human
brain.
More Computer Power
In the following paragraphs, we argue that even more computational power
will in fact be available, and so our margins for error are much larger.
Energy loss in rod logic, in Likharev's parametric quantron, in properly
designed NMOS and CMOS circuits, and in many other proposals for
computational devices is related to speed of operation. By slowing down
the operating speed from 100 picoseconds to 100 nanoseconds or even 100
microseconds we should achieve corresponding reductions in energy
dissipation per gate operation. This will allow substantial increases
in computational power for a fixed amount of energy (10^15 joules). We
can both decrease the energy dissipated per gate operation (by operating
at a slower speed) and increase the total number of gate operations (by
using more gates). Because the gates are very small to start with,
increasing their number by a factor of as much as 10^10 (to
approximately 10^27 gates) would still result in a total volume of 100
cubic meters (recall that each gate plus overhead is about 100 cubic
nanometers). This is a cube less than 5 meters on a side. Given that
manufacturing costs will eventually reflect primarily material and
energy costs, such a volume of slowly operating gates should be
economical and would deliver substantially more computational power per
joule.
We will not pursue this approach here for two main reasons. First,
published analyses use the higher 100 picosecond speed of operation and
10^-22 joules of energy dissipation[85]. Second, operating at 10^-22
joules at room temperature implies that most logic operations must be
reversible and that less than one logic operation in 30 can be
irreversible. Irreversible logic operations (which erase information)
must inherently dissipate at least kT x ln(2) for fundamental
thermodynamic reasons. The average thermal energy of a single atom or
molecule at a temperature T (measured in degrees K) is approximately kT
where k is Boltzman's constant. At room temperature, kT is about 4 x
10^-21 joules. Thus, each irreversible operation will dissipate almost
3 x 10^-21 joules. The number of such operations must be limited if we
are to achieve an average energy dissipation of 10^-22 joules per logic
operation.
While it should be feasible to perform computations in which virtually
all logic operations are reversible (and hence need not dissipate any
fixed amount of energy per logic operation)[9,25,32,53], current
computer architectures might require some modification before they could
be adapted to this style of operation. By contrast, it should be
feasible to use current computer architectures while at the same time
performing a major percentage (e.g., more than 99%) of their logic
operations in a reversible fashion.
Various electronic proposals show that almost all of the existing
combinatorial logic in present computers can be replaced with reversible
logic with no change in the instruction set that is executed[112, 113].
Further, while some instructions in current computers are irreversible
and hence must dissipate at least kT x ln(2) joules for each bit of
information erased, other instructions are reversible and need not
dissipate any fixed amount of energy if implemented correctly.
Optimizing compilers could then avoid using the irreversible machine
instructions and favor the use of the reversible instructions. Thus,
without modifying the instruction set of the computer, we can make most
logic operations in the computer reversible.
Further work on reversible computation can only lower the minimum energy
expenditure per basic operation and increase the percentage of
reversible logic operations. A mechanical logic proposal by the
author[105] eliminates most mechanisms of energy dissipation; it might
be possible to reduce energy dissipation to an extraordinary and
unexpected degree in molecular mechanical computers. While it is at
present unclear how far the trend towards lower energy dissipation per
logic operation can go, it is clear that we have not yet reached a limit
and that no particular limit is yet visible.
We can also expect further decreases in energy costs. By placing solar
cells in space the total incident sunlight per square meter can be
greatly increased (particularly if the solar cell is located closer to
the sun) while at the same time the total mass of the solar cell can be
greatly decreased. Most of the mass in earth-bound structures is
required not for functional reasons but simply to insure structural
integrity against the forces of gravity and the weather. In space both
these problems are virtually eliminated. As a consequence a very thin
solar cell of relatively modest mass can have a huge surface area and
provide immense power at much lower costs than estimated here.
If we allow for the decreasing future cost of energy and the probability
that future designs will have lower energy dissipation than 10^-22
joules per logic operation, it seems likely that we will have a great
deal more computational power than required. Even ignoring these more
than likely developments, we will have adequate computational power for
repair of the brain down to the molecular level.
Chemical Energy of the Brain
Another issue is the energy involved in the complete disassembly and
reassembly of every molecule in the brain. The total chemical energy
stored in the proteins and lipids of the human brain is quite modest in
comparison with 10^15 joules. When lipids are burned, they release
about 9 kilocalories per gram. (Calorie conscious dieters are actually
counting "kilocalories" - so a "300 Calorie Diet Dinner" really has
300,000 calories or 1,254,000 joules). When protein is burned, it
releases about 4 kilocalories per gram. Given that there are 100 grams
of protein and 175 grams of lipid in the brain, this means there is
almost 2,000 kilocalories of chemical energy stored in the structure of
the brain, or about 8 x 10^6 joules. This much chemical energy is over
10^8 times less than the 10^15 joules that one person can reasonably
purchase in the future. It seems unlikely that the construction of the
human brain must inherently require substantially more than 10^7 joules
and even more unlikely that it could require over 10^15 joules. The
major energy cost in repair down to the molecular level appears to be in
the computations required to "think" about each major molecule in the
brain.
Rate This Message: http://www.cryonet.org/cgi-bin/rate.cgi?msg=0019.3