Hello, new here, got my copy not long ago. Good reads so far.
Just wanted to report two errors in the book. One is very minor: on page 148
, the word "Dakotas" is misspelled as "Daktoas" (I thought it was a new species of bird or something).
The second error is a bit deeper, and involves several calculations in the "Human Computer" question. There appear to be a few flaws with the numbers. First, Hans Moravec doesn't say a human does one instruction per 90 seconds; rather, he says humans have the equivalent of 1.19e-08 MIPS (about 0.0119 IPS). The distinction is crucial: MIPS are not actually "instructions per second" in Moravec's terminology; rather, they are machine-independent computation equivalents. So it seems fishy to say that a human takes 90 seconds to execute an instruction (maybe true for division instructions, but definitely not true for data movement operations, which are some of the most common in programming). Anyway, if we assume those numbers are basically right, then the MIPS for the humans (on page 100) is still exactly 6 orders of magnitude wrong: it is stated that the low end of human capability is 0.01 MIPS, when it should be 0.00000001 (1e-8) MIPS.
On the high end, it is also a bit fishy. The footnote notes that K has 82,944 CPUs, 750M transistors each, and that it took 40 minutes to simulate 1 second of 1% of the synapses of a human brain. That's 6e13 transistors in the CPUs themselves. But now we need to figure out how many transistors would be needed to fully emulate the human brain at full speed -- i.e. K would need 40*60 times more transistors to run at real time, and 100 times more transistors to emulate the whole brain. In other words, we should need 1.5e19 transistors to emulate the human brain, which works out to 5e12 MIPS (assuming 30 transistors per IPS). The text only has 5e10 MIPS for the upper end of human capability.
Combining these two new estimates with a geometric mean yields 223 MIPS, which is a lot lower than the 30000 MIPS initially estimated by Randall. It is also much lower than the 100,000,000 MIPS estimated by Hans Moravec at hxxp://frc.ri.cmu.edu/users/hpm/book97/ch3/
(which he derived from multiplying the retina's apparent MIPS capability across the whole brain).
In terms of concrete numbers, the wrong ones are the two MIPS values at the ends of the spectrum on page 100; they should be 1e-8 and 5e12 respectively (and the transistor count should be upped to 1.5e19 from 1e15).