Semi-Daily Journal
Brad DeLong's Thoughts of the Moment on Economics, and on Other Topics as Well
DeLong's Home Page
November 01, 2002
Investment in Computers and Peripherals Once Again
Investment in Computers and Peripherals Once Again
This time with the "advance" GDP data from the third quarter of 2002 added in...
This is the chart that leads me to have a very hard time with the "we invested too much in computers in the 1990s and now have this large overhang of excess unused capital equipment in place" argument. We have a lot of excess capacity in telecommunications, yes. But elsewhere? If we have a lot of excess computer capital hanging around, why are we adding to our real computer capital stock more rapidly than ever before?
Posted by DeLong at November 01, 2002 07:39 PM
| Trackback
It is interesting (if not surprising) indeed. The nice thing with computer equipment is that it is designed to "expire" very quickly. We always want the latest model with the latest printer and hard drive blah blah blah. So, in a way, there is a bit of Say's Law at work.
The newest products have stayed so attractive both in terms of price and features that we can't help but buy them (and the same goes for firms' network managers). I can't wait to treat myself with one of those tiny laptops or one of those new gen PDAs.
I think the key is that computers answer a need that is more real (and more widespread) than one would be tempted to think. There is simply a lot of demand out there. And so long as new softwares (and OS'es) keep being more demanding in memory and computing power, consumers will have to amortize their material.
I also think ADSL and cable have to be thanked too for making the net experience so much more enviable than under a 56Kb. And it's good for demand because once you've solved the network bottleneck, you have to fix other bottlenecks to really see a difference. As long as I keep logging in with my 56kb modem at home, there is no point for me to upgrade my x86 with 164Mb of RAM.
I can't wait to see what we will do with fiber optics (when it comes). Television on demand? We will sure pirate whole movies (DVDs) but what else? It's hard to know in advance. Kids will probably be playing amazing network distributed 3D animation games with characters that move according to the movements of electrodes pathed on different body parts.
Is true AI ever going to take off? Back then people used to argue that genuine AI required computer power that was out of reach. With Moores' Law alive and well, aren't we getting there?
Pardon my abysmal ignorance, but what is the difference between "real" and "nominal" investment? Is it just a matter of whether it is adjusted for inflation, or what?
Very surprising data.
So, it's not going to be re-investment in computer equipment that is going to "lead" us out of our current "slow growth."
So what will? What's the next "internet" that will be the boom for the first decade?
a) Wireless?
b) Internet Services?
c) BioTech?
d) Some completely unexpected new Technology?
e) Something completely unexpected in the non-Tech sector?
Mike, why do you assume we will be led out? Look at Japan---they've been standing still for ten years.
Meanwhile it appears that our leaders believe the way to lead us out is through option (f) the weapons sector, apparently forgetting that the second world war didn't go quite as well for the rest of the world as it did for the US, and assuming that this time around once again the US will escape unscathed.
There appears considerable danger that our slow growth will continue for several more quarters.
Fed rate reductions with rates already quite low may have less impact than reductions of recent recessions. Fiscal stimulus is likely a couple of quarters away.
Then, the problem for high tech companies is not that demand for computers is lagging. Rather, the problem is that hard ware production is very competitive and prices are falling as fast or faster than demand is growing. Explosive profits in hard ware, which were once expected, are very hard to come by.
What of soft ware profits? Are microsoft and IBM so dominant in scope but limited in the ability to raise prices that profit growth there is also limited.
The gains from productivity increases would seem to be going largely to consumers just now.
That's what the nominal series tells us. Note however that even nominal investment is growing anew (perhaps because it is catching up with a big dip).
I'd be interested to take a look at the same graph with log-lienar axes and to know more about how (computer) investment is deflated. How is the price index constructed in this sector? Just the standard way?
Maynard, except for perhaps the 70's (I'm no economist so please correct me if wrong), we've always been "led out" by something into impressive growth. That's been the cycle. Our political and economic systems are far more dynamic and flexible compared to Japan's. That's why I don't think we'll see a Japanese style stagnation. And your second point, about others not making out as well as the US after WWII, I'm not sure what you are getting at. Eventually, all except for Soviet-bloc countries made out just fine. Of course, the battle never took place on US soil, so obviously our economy and morale didn't take the hit England, Europe, Japan and Russia did.
We still do not know how fast we have to grow to be a healthy economy. Could a fairly steady 3 percent suffice for a considerable period even with quicker productivity growth? What of 2 percent? I am simply not hearing of economic worries, though I think I should be. I am not hearing grumbles about business abuse, though I think I should be.
Are the legends on this chart reversed?
Stephen, the legends are correct because the price of computing power has plunged over the relevant years.
Does investment in computers and peripherals include home personal computers or are these somehow accounted for in consumption? If the different kinds of sale are accounted for seperately, than please disregard most of my previous post...
"Is true AI ever going to take off? Back then people used to argue that genuine AI required computer power that was out of reach. With Moores' Law alive and well, aren't we getting there?"
True AI (I've heard it referred to as "GOFAI", "Good Old-Fashioned AI") has been a spectacular failure. More computing power won't help--the problem is the combinatorial explosion you encounter when you try to solve these problems.
Best,
Stephen, to what extent can it also be explained by the Object Oriented Programming and the resulting lock-in into VB and C++? In other words, is GOFAI truly hopeless or is it that the bulk of the software industry's effort simply hasn't gone into it because there was a much more profitable market in OOP software (and programming environment)? Pardon the non-expert question...
"...why are we adding to our real computer capital stock more rapidly than ever before?"
This notion of "real" computer capital stock strikes me as being plainly incorrect. Certainly faster CPUs represent potentially higher value. "Real" computer capital stock, however, usually just gets used less intensely - our CPUs purr along at 4% CPU utilization instead of 40%.
I hear everyone talking about software upgrades finally gobbling up that additional computing power. Well, yes - if that software gets installed, it is also entered into the balance sheet. Without it, there just is no "real" additional output.
Do the statisticians correct the value of other new - non-computer - capital investment as well? Are new printing presses evaluated with respect to their potential printing output and then marked up accordingly?
To me, this kind of statistical cosmetics is without merit. It seems benign because it does not interfere with efficient market interaction. But it could ultimately lead to delusions about the degree of macroeconomic progress.
(I could devise a variant of the "correctionist" argument to "prove" that the American advantage in food processing productivity is, in fact, a disadvantage. In most southern European countries, the percentage of GDP spent on food is roughly one to two points higher than in the US, whereas the need for and spending on health care happens to be less - typically also about one to two percent of GDP. Should we now conclude that, "in reality", modern American food processing equipment reduces the quality of life and should be repriced by statisticians to reflect detrimental effects on Americansīhealth?
I guess such mischief would stand no chance of being officially adopted.)
I would hope that the calculations of a rapid increase in potential output do not depend in any way on such revaluations of actual investments.
In some ways, the statistical overemphasis on computing power could be likened to an economist who tries to find a metric for the "value" of language and decides that whatever arbitrary number he hit upon should be evenly divided among the letters of the alphabet. After all, language is made up of sounds, isnīt it? Surely, a little syntax and semantics can be added back in later as an afterthought...
To me, that would seem to be a terrible misrepresentation of the digital age - certainly not a recipe for success.
"True AI (I've heard it referred to as "GOFAI", "Good Old-Fashioned AI") has been a spectacular failure. More computing power won't help--the problem is the combinatorial explosion you encounter when you try to solve these problems"
Stephen--
That's an odd thing to say considering you carry around on your shoulders a computing device with approximately (very approximately!)1 million times the computing power of current computers and which functions in a way which would satisfy the goals of GOFAI (or at least I hope it should.... maybe you ought to register to take the Turing Test ;-)
Well, perhaps you believe it relies on some sort of magic and not ordinary computation, or maybe you think it is a quantum computer of some sort and we haven't figured this out yet. I don't think that's the case, but I suppose that would justify your comment.
I'm not sure what it would be about human intelligence that leads to combinatorial explosion. Perhaps incorrect algorithms to model it might lead to trouble. How do you claim that your brain escapes this (magic or simply proper design)?
In any case, the first major visible application of AI will probably be self driving cars (10-15 years is a good guess I think). This will probably involve a boom period, but likely will have little to do with current technology companies.
If real GOFAI comes to pass, the impact on the world economy would be mind boggling. At the current 2x/year increase in raw computing power and 4x/3 years (Moore's law) for memory this could occur in under 30 years based on the estimate of 1 million times current power needed and assuming somebody figures out what the heck to do with the power. Add in 20 years to the 20-50 year estimate to train these things to become doctors, lawyers, etc...
Meanwhile, the current tech companies are in the business of assisting humans, and will become increasingly irrelevant as our fixed abilities to handle data become increasingly saturated by available computing power.
Jean-Philippe,
OOP refers to a programming technique. GOFAI refers to artificial intelligence, as originally construed by its founders.
I'm hardly an expert on AI. From the little I know about computer science, however, it's basically fallen flat.
My impression is that GOFAI itself *is* hopeless. The question remains whether AI, more broadly construed, is hopeless. Which relates to...:
snsterling,
But the architecture of the human brain and the modern digital computer could not be more different. The brain is a massively parallel computing device; the modern computer is the opposite. Yes, there are parallel computers, but for the most part the architecture isn't parallel, nor are the algorithms.
How is my brain 1 million times more powerful? Depends on how you measure it. It isn't when it comes to floating point operations.
Human intelligence avoids the combinatorial explosion, presumably, by relying on entirely different algorithms than GOFAI.
OTOH, AI more generally might lead to breakthroughs. E.g. one might speculate that neural nets might lead to breakthroughs.
As to training AIs to be doctors, lawyers, etc, in many respects there is software which *can* function in that role. That's part of the point I'm making: it's far easier to create a computer program that solves problems that humans find difficult and require "expertise" (say, playing chess) than a program that does things which most of us find easy (such as parsing speech, supplying hidden inferences to text, walking around a room without bumping into objects, etc).
I think it's wisest not to try to predict the future, either positively or negatively.
In the short run, I think that computers are now more powerful for most users than they need be. Who needs 2.0 GHz of CPU? I just bought a used Dell Pentium III, and upgraded the RAM to 0.5 GB, for all of $100. But most things I do are what most people do: e-mail, browse the web, compose text, and the occasional spreadsheet app. I think that's why there's all this idiotic advertising about digital music and photography: short of that, a Pentium III is more than you need. Even for industrial applications, computers (hardware) offer more than is usually needed. I like to laugh at the problems the IRS and FBI supposedly have with their hardware; they could pick up the PCs we were surplussing at NIH (part of HHS) when I was there, and they'd be good to go.
Best,
Stephen,
I was considering network implementations to be included in GOFAI (the term I usually hear is "Strong" AI where the goal would be human intelligence), so I don't disagree with you. I certainly don't believe software in the current sense will lead to anything like a virtual lawyer. Especially not rules based AI.
And in my assumptions is that the proper algorithms will be found, whether it is by examining brains or if it is by mathematical investigation of how networks of neurons operate. If it isn't found the computational power will indeed be wasted.
The one million times estimate is roughly based on the number of neurons, the frequency of neurons (100-1000 cycles/sec is typical I think), and some estimates of what some people think a neuron might be doing (warning--could be higher if more is going on under the surface). I do not assume that the proper place to implement a virtual brain would be a Pentium PC. I use the term computing power very generally just to be able to get an idea of how much it would cost to implement the degree of multiplies and adds and to store the weights. Specialized hardware is always faster than a general purpose computer, and if one wanted to implement for example a network of neurons there would be many many design attributes one could take advantage of (memory is local to a node on the network, connections are localized too). Just to clarify, there's nothing wrong with emulating a parallel network architecture on a PC--it's still the same thing--but speedwise it just wouldn't make sense.
The estimates are very rough, and the continuation of Moore's law is anything but assured, but I think it is worthwhile to estimate how soon this could come rather than be surprised by its arrival or be surprised that it isn't here already.
I think you can assume that Moore's law is easily good for the next thirty years. The current work in nanotechnology is proceeding just fine and the breakthroughs it promises will easily keep our progress at least as rapid as it is today.
Interestingly, IBM is working on some Linux-based massively parallel operations. All the code for these machines is different than from our single processor universe, but we are engaged in the process and getting good at it is more a matter of time and experimentation that any sudden insight.
Finally, apocryphally, much AI research is undoubtedly being shunted into NSA and other government consumers of giant computing resources. The NSA voice processing algorithms and techniques require very sophisticated software. If you happened to go to school with some crackerjack computer scientists that you haven't heard from in a while, they may be busy writing black code.
When considering demand for computing power, it's important to distinguish between constant and surge demand for computing cycles. Do you sit on your couch every second of every day? Does that mean you don't want a nice couch?
Yes, most of the time a home or office computer might be used for word processing, web browsing, email, or other low-power tasks. Then you pull out the latest game, or want to edit a video, or need to run that new statistical model, and another version of marginal utility comes into play. It is the last unit of demand for processing power which has the most influence on system requirements.
In terms of AI driving demand, there isn't really a need for GOFAI to be the relevant factor. As techniques for expert systems of various kinds become more common, they cease to be considered AI at all. Consider Google's news page: Does it think? No, but it probably would have been considered a pretty cool achievement by Marvin Minsky back in the 60's. From systems which model flaws in silicon to automobile recognition systems for traffic control, it is the very diversity of these applications which allow them to be both increasingly commonplace in the economy and influential in the overall demand for new and upgraded systems.
When considering demand for computing power, it's important to distinguish between constant and surge demand for computing cycles. Do you sit on your couch every second of every day? Does that mean you don't want a nice couch?
Yes, most of the time a home or office computer might be used for word processing, web browsing, email, or other low-power tasks. Then you pull out the latest game, or want to edit a video, or need to run that new statistical model, and another version of marginal utility comes into play. It is the last unit of demand for processing power which has the most influence on system requirements.
In terms of AI driving demand, there isn't really a need for GOFAI to be the relevant factor. As techniques for expert systems of various kinds become more common, they cease to be considered AI at all. Consider Google's news page: Does it think? No, but it probably would have been considered a pretty cool achievement by Marvin Minsky back in the 60's. From systems which model flaws in silicon to automobile recognition systems for traffic control, it is the very diversity of these applications which allow them to be both increasingly commonplace in the economy and influential in the overall demand for new and upgraded systems.
When considering demand for computing power, it's important to distinguish between constant and surge demand for computing cycles. Do you sit on your couch every second of every day? Does that mean you don't want a nice couch?
Yes, most of the time a home or office computer might be used for word processing, web browsing, email, or other low-power tasks. Then you pull out the latest game, or want to edit a video, or need to run that new statistical model, and another version of marginal utility comes into play. It is the last unit of demand for processing power which has the most influence on system requirements.
In terms of AI driving demand, there isn't really a need for GOFAI to be the relevant factor. As techniques for expert systems of various kinds become more common, they cease to be considered AI at all. Consider Google's news page: Does it think? No, but it probably would have been considered a pretty cool achievement by Marvin Minsky back in the 60's. From systems which model flaws in silicon to automobile recognition systems for traffic control, it is the very diversity of these applications which allow them to be both increasingly commonplace in the economy and influential in the overall demand for new and upgraded systems.
"In terms of AI driving demand, there isn't really a need for GOFAI to be the relevant factor. As techniques for expert systems of various kinds become more common, they cease to be considered AI at all. Consider Google's news page: Does it think? No, but it probably would have been considered a pretty cool achievement by Marvin Minsky back in the 60's."
My understanding is that a very substantial portion of total stock trades are now "programmed" trades...determined by computers. So computers are already more "artificially intelligent" at selecting stocks than humans.
Ray Kurzweil makes the point that once the algorithms are worked out to do something, we stop thinking about it as actually being "artificially intelligent," even though 40, 30, or even 20 years ago, we might have considered a machine doing the same thing as extremely intelligent. He brings up the point that only a few decades ago, many people thought a computer could never defeat a chess grand master.
"Pardon my abysmal ignorance, but what is the difference between "real" and "nominal" investment? Is it just a matter of whether it is adjusted for inflation, or what?"
I don't know the specifics for this graph, but I would assume it's something like: "How many equivalents of a Pentium I computer, with a 400 Meg hard drive, and 64 Meg of RAM, are being sold?"
In other words, they take what could be purchased with $1000 in 1996, and figure out how much more computing power could be purchased with that same $1000 in later years...and how much less computing power could have been purchased in years prior to 1996.
In that sense, I'm a bit surprised by how *little* improvement has been made. If I eyeball the 1987 value to be $15 billion, and eyeball the 2002 (end of the curve) value to be $310 billion...that's "only" a factor of ~20 in 15 years. In other words, a little over 4 doublings in 15 years...or doubling every 3.7 years.
Having just finished reading Ray Kurzweil, I expected better. ;-) (Kurzweil really has some amazing analyses!)
http://www.kurzweilai.net/meme/frame.html?main=/articles/art0184.html
"Even for industrial applications, computers (hardware) offer more than is usually needed."
Yeah. But just last Friday, I downloaded something from the Internet, at work: 13 Meg, in under a minute. Sure beats the 2 hours it would have taken on my home modem!