Comments on Kurzweil’s prediction graphs
Editorial notes sent to Amara Angelica at KurzweilAI.net in May 2010, on a set of trend graphs being prepared for one of Ray Kurzweil’s books. Alyssa, then consulting for KurzweilAI, walks the charts one by one and flags missing data points, sources that don’t support the curve they’re drawn under, and at least one chart (“paradigm shifts” of key historical events) that she thinks should be cut entirely because it conflates information decay with acceleration — a classic Schmidhuber objection. The notes also catch several places where a sigmoid fits the recent data far better than the exponential the chart assumes.
OK, looking at them: Magnetic data storage: I would note that there appears to be a different (longer) doubling time until 1990, when a new (shorter) doubling time takes over.
CPU performance: I think we should include a 2008 data point, available at Tom’s Hardware . We might also want to include 2003 and 2004 data points, since there’s a gap during that time.
RAM prices per dollar: we ought to include more intermediate data points, since we have nothing between 2004 and 2008. There are a zillion such data points in the Pricewatch archives .
Supercomputers: we ought to put something on the graph differentiating “planned computers” (the dates after 2010) from computers currently in operation. I would also recommend removing the 2018 datapoint (exaflop), since it just represents a goal and not a current engineering project.
Total bits shipped: I honestly can’t tell what this is supposed to mean, just looking at the chart. Number of memory bits? Number of magnetic storage bits? Number of data bits, e.g. YouTube videos?
Transistors per chip: for data since 1990, the sigmoidal curve fits a lot better than the linear curve for log(transistors) vs. time (R 2 = 0.9953 compared to 0.9416, more than a full OOM better). Also, the 2010 data point appears not to come from the source listed, but from a different source . If we include GPU chips, we’re now up to 3 billion transistors . Maybe we ought to do more charts involving GPUs?
Charts that we want new data sources for: IT spending as a percentage of GDP: the World Bank figure for 2005 is 8.8% of GDP in the US for “information technology and communication expenditures,” which includes things like phones. Forrester Research gives a figure of 3.9% for 2009 for “technology products and services” in the US. For comparison, the 1998 figure in the original graph is 8.2%.
US phone industry: I’ve actually written a paper about this, available here .
“Paradigm shifts” of key events: this is a highly problematic method of analysis that I recommend not using, because of the way history works. Older events, due to information decay, are harder to get information on; we know more about the Apollo project than about the entire decades-long reign of the Egyptian pharaoh Khufu. Hence, as you move toward the present, each section of the history books will contain more information per year, and so if you graph the events listed in sequence, it will seem like history is speeding up and happening faster and faster, even if it is stagnant. Jürgen Schmidhuber jokes about this in his 2009 Summit talk .
US life expectancy: growth has actually slowed down in recent years, despite exponentially increasing medical spending ( chart ).
Transistor prices: I’m curious as to what kind of devices the figures are for. CPU transistor prices are in the 1.5 × 10 -7 range right now, about an OOM higher than Zieber’s figures.
Cost of gene sequencing: I think we should include more intermediate data points.
This NYT piece cites an effective cost of $0.00016 per base pair in 2007, an OOM more costly than the 2008 figure and a factor of 6 less costly than the 2004 figure. I can probably find more if desired.
DRAM prices: the graphs as they are now include projected prices out to 2022, in addition to current prices. There really ought to be something on the graph which differentiates actual vs. projected. As for more recent data (ITRS only gives out to 2008 actual), ZDNet Asia gives a cost per bit of 0.02095 microcents, as compared to ITRS projections of 0.048 for 2009 and 0.034 for 2010.
Total amount of genetic data since 1990: fitting a sigmoid function to the log of the number of base pairs/sequences fits way better than just a straight line (exponential growth). I get R 2 values of 0.9985 vs. 0.9708 (base pairs) and 0.99917 vs. 0.9715 (sequences).
Internet traffic: our current source (U. Minnesota) now gives a figure of 1,800–2,700 PB/month for 2009 US Internet traffic — note that all of the Minnesota figures, the ones in the current chart, are for US traffic specifically, not world traffic; see MINTS . Global traffic is in the 10,000 PB/month range. As a cross-check, Cisco has their own numbers on global Internet usage, which can be found on Wikipedia at Internet traffic . Unfortunately, the original (Cisco) source does not appear to be publicly indexed; you can contact Cisco about that, if you want, at traffic-inquiries@cisco.com.