Microsoft Word Kurzweil, Ray The Singularity Is Near doc

The Scalability of Human Intelligence.
In response to Hofstadter's concern as to whether human intelligence is just 
above or below the threshold necessary for "self-understanding," the accelerating pace of brain reverse engineering 
makes it clear that there are no limits to our ability to understand ourselves—or anything else, for that matter. The key 
to the scalability of human intelligence is our ability to build models of reality in our mind. These models can be 
recursive, meaning that one model can include other models, which can include yet finer models, without limit. For 
example, a model of a biological cell can include models of the nucleus, ribosomes, and other cellular systems. In turn, 
the model of the ribosome may include models of its submolecular components, and then down to the atoms and 
subatomic particles and forces that it comprises. 
Our ability to understand complex systems is not necessarily hierarchical. A complex system like a cell or the 
human brain cannot be understood simply by breaking it down into constituent subsystems and their components. We 
have increasingly sophisticated mathematical tools for understanding systems that combine both order and chaos—and 
there is plenty of both in a cell and in the brain—and for understanding the complex interactions that defy logical 
breakdown. 
Our computers, which are themselves accelerating, have been a critical tool in enabling us to handle increasingly 
complex models, which we would otherwise be unable to envision with our brains alone. Clearly, Hofstadter's concern 
would be correct if we were limited just to models that we could keep in our minds without technology to assist us. 
That our intelligence is just above the threshold necessary to understand itself results from our native ability, combined 
with the tools of our own making, to envision, refine, extend, and alter abstract—and increasingly subtle—models of 
our own observations. 
Uploading the Human Brain 
To become a figment of your computer's imagination. 
—D
AVID 
V
ICTOR DE 
T
RANSCEND
,
G
ODLING
'
S 
G
LOSSARY
,
DEFINITION OF 
"
UPLOAD
" 
A more controversial application than the scanning-the-brain-to-understand-it scenario is 
scanning the brain to upload 
it
. Uploading a human brain means scanning all of its salient details and then reinstantiating those details into a 
suitably powerful computational substrate. This process would capture a person's entire personality, memory, skills, 
and history. 
If we are truly capturing a particular person's mental processes, then the reinstantiated mind will need a body, 
since so much of our thinking is directed toward physical needs and desires. As I will discuss in chapter 5, by the time 
we have the tools to capture and re-create a human brain with all of its subtleties, we will have plenty of options for 
twenty-first-century bodies for both nonbiological humans and biological humans who avail themselves of extensions 
to our intelligence. The human body version 2.0 will include virtual bodies in completely realistic virtual 
environments, nanotechnology-based physical bodies, and more. 
In chapter 3 I discussed my estimates for the memory and computational requirements to simulate the human 
brain. Although I estimated that 10
16
cps of computation and 10
13
bits of memory are sufficient to emulate human 
levels of intelligence, my estimates for the requirements of uploading were higher: 10
19
cps and 10
18
bits, respectively. 
The reason for the higher estimates is that the lower ones are based on the requirements to re-create regions of the 
brain at human levels of performance, whereas the higher ones are based on capturing the salient details of each of our 

---

approximately 10
11
neurons and 10
14
interneuronal connections. Once uploading is feasible, we are likely to find that 
hybrid solutions are adequate. For example, we will probably find that it is sufficient to simulate certain basic support 
functions such as the signal processing of sensory data on a functional basis (by plugging in standard modules) and 
reserve the capture of subneuron details only for those regions that are truly responsible for individual personality and 
skills. Nonetheless, we will use our higher estimates for this discussion. 
The basic computational resources (10
19
cps and 10
18
bits) will be available for one thousand dollars in the early 
2030s, about a decade later than the resources needed for functional simulation. The scanning requirements for 
uploading are also more daunting than for "merely" re-creating the overall powers of human intelligence. In theory one 
could upload a human brain by capturing all the necessary details without necessarily comprehending the brain's 
overall plan. In practice, however, this is unlikely to work. Understanding the principles of operation of the human 
brain will reveal which details are essential and which details are intended to be disordered. We need to know, for 
example, which molecules in the neurotransmitters are critical, and whether we need to capture overall levels, position 
and location, and/or molecular shape. As I discussed above, we are just learning, for example, that it is the position of 
actin molecules and the shape of ePEB molecules in the synapse that are key for memory. It will not be possible to 
confirm which details are crucial without having confirmed our understanding of the theory of operation. That 
confirmation will be in the form of a functional simulation of human intelligence that passes the Turing test, which I 
believe will take place by 2029.
119
To capture this level of detail will require scanning from within the brain using nanobots, the technology for 
which will be available by the late 2020s. Thus, the early 2030s is a reasonable time frame for the computational 
performance, memory, and brain-scanning prerequisites of uploading. Like any other technology, it will take some 
iterative refinement to perfect this capability, so the end of the 2030s is a conservative projection for successful 
uploading. 
We should point out that a person's personality and skills do not reside only in the brain, although that is their 
principal location. Our nervous system extends throughout the body, and the endocrine (hormonal) system has an 
influence, as well. The vast majority of the complexity, however, resides in the brain, which is the location of the bulk 
of the nervous system. The bandwidth of information from the endocrine system is quite low, because the determining 
factor is overall levels of hormones, not the precise location of each hormone molecule. 
Confirmation of the uploading milestone will be in the form of a "Ray Kurzweil" or "Jane Smith" Turing test, in 
other words convincing a human judge that the uploaded re-creation is indistinguishable from the original specific 
person. By that time we'll face some complications in devising the rules of any Turing test. Since nonbiological 
intelligence will have passed the original Turing test years earlier (around 2029), should we allow a nonbiological 
human equivalent to be a judge? How about an enhanced human? Unenhanced humans may become increasingly hard 
to find. In any event, it will be a slippery slope to define enhancement, as many different levels of extending biological 
intelligence will be available by the time we have purported uploads. Another issue will be that the humans we seek to 
upload will not be limited to their biological intelligence. However, uploading the nonbiological portion of intelligence 
will be relatively straightforward, since the ease of copying computer intelligence has always represented one of the 
strengths of computers. 
One question that arises is, How quickly do we need to scan a person's nervous system? It clearly cannot be done 
instantaneously, and even if we did provide a nanobot for each neuron, it would take time to gather the data. One 
might therefore object that because a person's state is changing during the data-gathering process, the upload 
information does not accurately reflect that person at an instant in time but rather over a period of time, even if only a 
fraction of a second.
120
Consider, however, that this issue will not interfere with an upload's passing a "Jane Smith" 
Turing test. When we encounter one another on a day-to-day basis, we are recognized as ourselves even though it may 
have been days or weeks since the last such encounter. If an upload is sufficiently accurate to re-create a person's state 
within the amount of natural change that a person undergoes in a fraction of a second or even a few minutes, that will 
be sufficient for any conceivable purpose. Some observers have interpreted Roger Penrose's theory of the link between 
quantum computing and consciousness (see chapter 9) to mean that uploading is impossible because a person's 

---

"quantum state" will have changed many times during the scanning period. But I would point out that my quantum 
state has changed many times in the time it took me to write this sentence, and I still consider myself to be the same 
person (and no one seems to be objecting). 
Nobel Prize winner Gerald Edelman points out that there is a difference between a capability and a description of 
that capability. A photograph of a person is different from the person herself, even if the "photograph" is very high 
resolution and three-dimensional. However, the concept of uploading goes beyond the extremely high-resolution scan, 
which we can consider the "photograph" in Edelman's analogy. The scan does need to capture all of the salient details, 
but it also needs to be instantiated into a working computational medium that has the capabilities of the original (albeit 
that the new nonbiological platforms are certain to be far more capable). The neural details need to interact with one 
another (and with the outside world) in the same ways that they do in the original. A comparable analogy is the 
comparison between a computer program that resides on a computer disk (a static picture) and a program that is 
actively running on a suitable computer (a dynamic, interacting entity). Both the data capture and the reinstantiation of 
a dynamic entity constitute the uploading scenario. 
Perhaps the most important question will be whether or not an uploaded human brain is really you. Even if the 
upload passes a personalized Turing test and is deemed indistinguishable from you, one could still reasonably ask 
whether the upload is the same person or a new person. After all, the original person may still exist. I'll defer these 
essential questions until chapter 7. 
In my view the most important element in uploading will be our gradual transfer of our intelligence, personality, 
and skills to the nonbiological portion of our intelligence. We already have a variety of neural implants. In the 2020s 
we will use nanobots to begin augmenting our brains with nonbiological intelligence, starting with the "routine" 
functions of sensory processing and memory, moving on to skill formation, pattern recognition, and logical analysis. 
By the 2030s the nonbiological portion of our intelligence will predominate, and by the 2040s, as I pointed out in 
chapter 3, the nonbiological portion will be billions of times more capable. Although we are likely to retain the 
biological portion for a period of time, it will become of increasingly little consequence. So we will have effectively 
uploaded ourselves, albeit gradually, never quite noticing the transfer. There will be no "old Ray" and "new Ray," just 
an increasingly capable Ray. Although I believe that uploading as in the sudden scan-and-transfer scenario discussed 
in this section will be a feature of our future world, it is this gradual but inexorable progression to vastly superior 
nonbiological thinking that will profoundly transform human civilization. 
S
IGMUND 
F
REUD
: