192
PARTISAN REVIEW
And if we use this model and extend these graphs into the twenty-first
century, what we see is the following.
Currently, a one-thousand-dollar computer provides an amount of
computation that's somewhere between that of an insect and a mouse
brain. Now, insect and mouse brains are very well designed for meeting
the challenges of being an insect or a mouse. Computers are designed
for other purposes, but that's where we are today. They're still at least
a million times simpler than the human brain, which is at least one rea–
son why computers today don't have all of the endearing qualities we
associate with human intelligence. This is in terms of just processing
power, which is only one part of the equation because we're more than
just raw capacity, although it is at least one necessary ingredient. The
human brain has about a hundred billion neurons, and each neuron has
an average of a thousand connections, so the brain has about a hundred
trillion connections. The calculations basically take place in the connec–
tions. They're analogue, not digital, but we can emulate analogue cal–
culations, either by using digital circuitry or just by building analogue
circuits electronically. But that's a technical detail.
There are a hundred trillion things going on in our brain at the same
time. But neurons are very slow. Neurons are very complex entities,
although most of that complexity is devoted to life processes, not fheir
information-handling abilities. Their information processing is not as
complex as the overall complexity of the neurons, but is much more
complex than, say, the simplified mathematical models that we use in the
field of neural nets. But there is an emerging field that is sort of a cross–
section between neural nets and neurobiology, where we're attempting to
build very detailed, realistic models of neurons.
I'll
come back to that.
At any rate, neurons calculate only about two hundred calculations
per second, which is at least ten million times slower than electronic cir–
cuits. But two hundred calculations per second multiplied by one hun–
dred trillionfold parallelism is twenty million billion calculations per
second, or about twenty million MIPs. We will achieve twenty million
MIPs for one thousand dollars in about
2020.
By
2030,
a one-thousand–
dollar computation will be one thousand times more powerful than the
human brain; by
2050
a one-thousand-dollar computation will equal ten
billion human brains. I might be off by a couple of years, but the twenty–
first century won't be wanting for basic computational capacity. How–
ever, this is just one necessary-but not sufficient- ingredient to
recreate human intelligence.
If
all we had was an extremely fast computer
that could do twenty billion MIPs, or more, we could compute your
