In 1999, Ray Kurzweil, one of the world’s greatest futurists, published a book called The Age of Spiritual Machines. In it, he made the case that artificial intelligence, nanomachines, virtual reality, brain implants, and other technologies would greatly improve during the 21st century, radically altering the world and the human experience. In the final four chapters, titled “2009,” “2019,” “2029,” and “2099,” he made detailed predictions about what the state of key technologies would be in each of those years, and how they would impact everyday life, politics and culture.
Towards the end of 2009, a number of news columnists, bloggers and even Kurzweil himself weighed in on how accurate his predictions from the eponymous chapter turned out. By contrast, no such analysis was done over the past year regarding his 2019 predictions. As such, I’m taking it upon myself to do it.
I started analyzing the accuracy of Kurzweil’s predictions in late 2019 and wanted to publish my full results before the end of that year. However, the task required me to do much more research that I had expected, so I missed that deadline. Really digging into the text of The Age of Spiritual Machines and parsing each sentence made it clear that the number and complexity of the 2019 predictions were greater than a casual reading would suggest. Once I realized how big of a task it would be, I became kind of demoralized and switched to working on easier projects for this blog.
With the end of 2020 on the horizon, I think time is running out to finish this, and I’ve decided to tackle the problem by breaking it into smaller, manageable chunks: My analysis of Kurzweil’s 2019 predictions from The Age of Spiritual Machines will be spread out over three blog entries, the first of which you’re now reading. Except where noted, I will only use sources published before January 1, 2020 to support my conclusions.
“Computers are now largely invisible. They are embedded everywhere–in walls, tables, chairs, desks, clothing, jewelry, and bodies.”
RIGHT
A computer is a device that stores and processes data, and executes its programming. Any machine that meets those criteria counts as a computer, regardless of how fast or how powerful it is (also, it doesn’t even need to run on electricity). This means something as simple as a pocket calculator, programmable thermostat, or a Casio digital watch counts as a computer. These kinds of items were ubiquitous in developed countries in 1998 when Ray Kurzweil wrote the book, so his “futuristic” prediction for 2019 could have just as easily applied to the reality of 1998. This is an excellent example of Kurzweil making a prediction that leaves a certain impression on the casual reader (“Kurzweil says computers will be inside EVERY object in 2019!”) that is unsupported by a careful reading of the prediction.
“People routinely use three-dimensional displays built into their glasses or contact lenses. These ‘direct eye’ displays create highly realistic, virtual visual environments overlaying the ‘real’ environment.”
MOSTLY WRONG
The first attempt to introduce augmented reality glasses in the form of Google Glass was probably the most notorious consumer tech failure of the 2010s. To be fair, I think this was because the technology wasn’t ready yet (e.g. – small visual display, low-res images, short battery life, high price), and not because the device concept is fundamentally unsound. The technological hangups that killed Google Glass will of course vanish in the future thanks to factors like Moore’s Law. Newer AR glasses, like Microsoft’s Hololens, are already superior to Google Glass, and given the pace of improvement, I think AR glasses will be ready for another shot at widespread commercialization by the end of the 2020s, but they will not replace smartphones for a variety of reasons (such as the unwillingness of many people to wear glasses, widespread discomfort with the possibility that anyone wearing AR glasses might be filming the people around them, and durability and battery life advantages of smartphones).
Kurzweil’s prediction that contact lenses would have augmented reality capabilities completely failed. A handful of prototypes were made, but never left the lab, and there’s no indication that any tech company is on the cusp of commercializing them. I doubt it will happen until the 2030s.
However, people DO routinely access augmented reality, but through their smartphones and not through eyewear. Pokemon Go was a worldwide hit among video gamers in 2016, and is an augmented reality game where the player uses his smartphone screen to see virtual monsters overlaid across live footage of the real world. Apps that let people change their appearances during live video calls (often called “face filters”), such as by making themselves appear to have cartoon rabbit ears, are also very popular among young people.
So while Kurzweil got augmented reality technology’s form factor wrong, and overestimated how quickly AR eyewear would improve, he was right that ordinary people would routinely use augmented reality.
The augmented reality glasses will also let you experience virtual reality.
WRONG
Augmented reality glasses and virtual reality goggles remain two separate device categories. I think we will someday see eyewear that merges both functions, but it will take decades to invent glasses that are thin and light enough to be worn all day, untethered, but that also have enough processing power and battery life to provide a respectable virtual reality experience. The best we can hope for by the end of the 2020s will be augmented reality glasses that are good enough to achieve ~10% of the market penetration of smartphones, and virtual reality goggles that have shrunk to the size of ski goggles.
Of note is that Kurzweil’s general sentiment that VR would be widespread by 2019 is close to being right. VR gaming made a resurgence in the 2010s thanks to better technology, and looks poised to go mainstream in the 2020s.
The augmented reality / virtual reality glasses will work by projecting images onto the retinas of the people wearing them.
PARTLY RIGHT
The most popular AR glasses of the 2010s, Google Glass, worked by projecting images onto their wearer’s retinas. The more advanced AR glass models that existed at the end of the decade used a mix of methods to display images, none of which has established dominance.
The “Magic Leap One” AR glasses use the retinal projection technology Kurzweil favored. They are superior to Google Glass since images are displayed to both eyes (Glass only had a projector for the right eye), in higher resolution, and covering a larger fraction of the wearer’s field of view (FOV). Magic Leap One also has advanced sensors that let it map its physical surroundings and movements of its wearer, letting it display images of virtual objects that seem to stay fixed at specific points in space (Kurzweil called this feature “Virtual-reality overlay display”).
Microsoft’s “Hololens” uses a different technology to produce images: the lenses are in fact transparent LCD screens. They display images just like a TV screen or computer monitor would. However, unlike those devices, the Hololens’ LCDs are clear, allowing the wearer to also see the real world in front of them.
The “Vuzix Blade” AR glasses have a small projector that beams images onto the lens in front of the viewer’s right eye. Nothing is directly beamed onto his retina.
It must emphasized again that, at the end of 2019, none of these or any other AR glasses were in widespread or common use, even in rich countries. They were confined to small numbers of hobbyists, technophiles, and software developers. A Magic Leap One headset cost $2,300 – $3,300 depending on options, and a Hololens was $3,000.
And as stated, AR glasses and VR goggles remained two different categories of consumer devices in 2019, with very little crossover in capabilities and uses. The top-selling VR goggles were the Oculus Rift and the HTC Vive. Both devices use tiny OLED screens positioned a few inches in front of the wearer’s eyes to display images, and as a result, are much bulkier than any of the aforementioned AR glasses. In 2019, a new Oculus Rift system cost $400 – $500, and a new HTC Vive was $500 – $800.
“[There] are auditory ‘lenses,’ which place high resolution-sounds in precise locations in a three-dimensional environment. These can be built into eyeglasses, worn as body jewelry, or implanted in the ear canal.”
MOSTLY RIGHT
Humans have the natural ability to tell where sounds are coming from in 3D space because we have “binaural hearing”: our brains can calculate the spatial origin of the sound by analyzing the time delay between that sound reaching each of our ears, as well as the difference in volume. For example, if someone standing to your left is speaking, then the sounds of their words will reach your left ear a split second sooner than they reach your right ear, and their voice will also sound louder in your left ear.
By carefully controlling the timing and loudness of sounds that a person hears through their headphones or through a single speaker in front of them, we can take advantage of the binaural hearing process to trick people into thinking that a recording of a voice or some other sound is coming from a certain direction even though nothing is there. Devices that do this are said to be capable of “binaural audio” or “3D audio.” Kurzweil’s invented term “audio lenses” means the same thing.
Yes, there are eyeglasses with built-in speakers that play binaural audio. The Bose Frames “smart sunglasses” is the best example. Even though the devices are not common, they are commercially available, priced low enough for most people to afford them ($200), and have gotten good user reviews. Kurzweil gets this one right, and not by an eyerolling technicality as would be the case if only a handful of million-dollar prototype devices existed in a tech lab and barely worked.
Wireless earbuds are much more popular, and upper-end devices like the SoundPEATS Truengine 2 have impressive binaural audio capabilities. It’s a stretch, but you could argue that branding, and sleek, aesthetically pleasing design qualifies some higher-end wireless earbud models as “jewelry.”
Sound bars have also improved and have respectable binaural surround sound capabilities, though they’re still inferior to traditional TV entertainment system setups where the sound speakers are placed at different points in the room. Sound bars are examples of single-point devices that can trick people into thinking sounds are originating from different points in space, and in spirit, I think they are a type of technology Kurzweil would cite as proof that his prediction was right.
The last part of Kurzweil’s prediction is wrong, since audio implants into the inner ears are still found only in people with hearing problems, which is the same as it was in 1998. More generally, people have shown themselves more reluctant to surgically implant technology in their bodies than Kurzweil seems to have predicted, but they’re happy to externally wear it or to carry it in a pocket.
“Keyboards are rare, although they still exist. Most interaction with computing is through gestures using hands, fingers, and facial expressions and through two-way natural-language spoken communication. “
MOSTLY WRONG
Rumors of the keyboard’s demise have been greatly exaggerated. Consider that, in 2018, people across the world bought 259 million new desktop computers, laptops, and “ultramobile” devices (higher-end tablets that have large, detachable keyboards [the Microsoft Surface dominates this category]). These machines are meant to be accessed with traditional keyboard and mouse inputs.
The research I’ve done suggests that the typical desktop, laptop, and ultramobile computer has a lifespan of four years. If we accept this, and also assume that the worldwide computer sales figures for 2015, 2016, and 2017 were the same as 2018’s, then it means there are 1.036 billion fully functional desktops, laptops, and ultramobile computers on the planet (about one for every seven people). By extension, that means there are at least 1.036 billion keyboards. No one could reasonably say that Kurzweil’s prediction that keyboards would be “rare” by 2019 is correct.
The second sentence in Kurzweil’s prediction is harder to analyze since the meaning of “interaction with computing” is vague and hence subjective. As I wrote before, a Casio digital watch counts as a computer, so if it’s nighttime and I press one of its buttons to illuminate the display so I can see the time, does that count as an “interaction with computing”? Maybe.
If I swipe my thumb across my smartphone’s screen to unlock the device, does that count as an “interaction with computing” accomplished via a finger gesture? It could be argued so. If I then use my index finger to touch the Facebook icon on my smartphone screen to open the app, and then use a flicking motion of my thumb to scroll down over my News Feed, does that count as two discrete operations in which I used finger gestures to interact with computing?
You see where this is going…
Being able to set the bar that low makes it possible that this part of Kurzweil’s prediction is right, as unsatisfying as that conclusion may be.
Virtual reality gaming makes use of hand-held and hand-worn controllers that monitor the player’s hand positions and finger movements so he can grasp and use objects in the virtual environment, like weapons and steering wheels. Such actions count as interactions with computing. The technology will only get more refined, and I can see them replacing older types of handheld game controllers.
Hand gestures, along with speech, are also the natural means to interface with augmented reality glasses since the devices have tiny surfaces available for physical contact, meaning you can’t fit a keyboard on a sunglass frame. Future AR glasses will have front-facing cameras that watch the wearer’s hands and fingers, allowing them to interact with virtual objects like buttons and computer menus floating in midair, and to issue direct commands to the glasses through specific hand motions. Thus, as AR glasses get more popular in the 2020s, so will the prevalence of this mode of interface with computers.
“Two-way natural-language spoken communication” is now a common and reliable means of interacting with computers, as anyone with a smart speaker like an Amazon Echo can attest. In fact, virtual assistants like Alexa, Siri, and Cortana can be accessed via any modern smartphone, putting this within reach of billions of people.
The last part of Kurzweil’s prediction, that people would be using “facial expressions” to communicate with their personal devices, is wrong. For what it’s worth, machines are gaining the ability to read human emotions through our facial expressions (including “microexpressions”) and speech. This area of research, called “affective computing,” is still stuck in the lab, but it will doubtless improve and find future commercial applications. Someday, you will be able to convey important information to machines through your facial expressions, tone of voice, and word choice just as you do to other humans now, enlarging your mode of interacting with “computing” to encompass those domains.
“Significant attention is paid to the personality of computer-based personal assistants, with many choices available. Users can model the personality of their intelligent assistants on actual persons, including themselves…”
WRONG
The most widely used computer-based personal assistants–Alexa, Siri, and Cortana–don’t have “personalities” or simulated emotions. They always speak in neutral or slightly upbeat tones. Users can customize some aspects of their speech and responses (i.e. – talking speed, gender, regional accent, language), and Alexa has limited “skill personalization” abilities that allow it to tailor some of its responses to the known preferences of the user interacting with it, but this is too primitive to count as a “personality adjustment” feature.
My research didn’t find any commercially available AI personal assistant that has something resembling a “human personality,” or that is capable of changing that personality. However, given current trends in AI research and natural language understanding, and growing consumer pressure on Silicon Valley’s to make products that better cater to the needs of nonwhite people, it is likely this will change by the end of this decade.
“Typically, people do not own just one specific ‘personal computer’…”
RIGHT
A 2019 Pew survey showed that 75% of American adults owned at least one desktop or laptop PC. Additionally, 81% of them owned a smartphone and 52% had tablets, and both types of devices have all the key attributes of personal computers (advanced data storing and processing capabilities, audiovisual outputs, accepts user inputs and commands).
The data from that and other late-2010s surveys strongly suggest that most of the Americans who don’t own personal computers are people over age 65, and that the 25% of Americans who don’t own traditional PCs are very likely to be part of the 19% that also lack smartphones, and also part of the 48% without tablets. The statistical evidence plus consistent anecdotal observations of mine lead me to conclude that the “typical person” in the U.S. owned at least two personal computers in late 2019, and that it was atypical to own fewer than that.
“Computing and extremely high-bandwidth communication are embedded everywhere.”
MOSTLY RIGHT
This is another prediction whose wording must be carefully parsed. What does it mean for computing and telecommunications to be “embedded” in an object or location? What counts as “extremely high-bandwidth”? Did Kurzweil mean “everywhere” in the literal sense, including the bottom of the Marianas Trench?
First, thinking about my example, it’s clear that “everywhere” was not meant to be taken literally. The term was a shorthand for “at almost all places that people typically visit” or “inside of enough common objects that the average person is almost always near one.”
Second, as discussed in my analysis of Kurzweil’s first 2019 prediction, a machine that is capable of doing “computing” is of course called a “computer,” and they are much more ubiquitous than most people realize. Pocket calculators, programmable thermostats, and even a Casio digital watch count computers. Even 30-year-old cars have computers inside of them. So yes, “computing” is “embedded ‘everywhere'” because computers are inside of many manmade objects we have in our homes and workplaces, and that we encounter in public spaces.
Of course, scoring that part of Kurzweil’s prediction as being correct leaves us feeling hollow since those devices don’t the full range of useful things we associate with “computing.” However, as I noted in the previous prediction, 81% of American adults own smartphones, they keep them in their pockets or near their bodies most of the time, and smartphones have all the capabilities of general-purpose PCs. Smartphones are not “embedded” in our bodies or inside of other objects, but given their ubiquity, they might as well be. Kurzweil was right in spirit.
Third, the Wifi and mobile phone networks we use in 2019 are vastly faster at data transmission than the modems that were in use in 1999, when The Age of Spiritual Machines was published. At that time, the commonest way to access the internet was through a 33.6k dial-up modem, which could upload and download data at a maximum speed of 33,600 bits per second (bps), though upload speeds never got as close to that limit as download speeds. 56k modems had been introduced in 1998, but they were still expensive and less common, as were broadband alternatives like cable TV internet.
In 2019, standard internet service packages in the U.S. typically offered WiFi download speeds of 30,000,000 – 70,000,000 bps (my home WiFi speed is 30-40 Mbps, and I don’t have an expensive service plan). Mean U.S. mobile phone internet speeds were 33,880,000 bps for downloads and 9,750,000 bps for uploads. That’s a 1,000 to 2,000-fold speed increase over 1999, and is all the more remarkable since today’s devices can traffic that much data without having to be physically plugged in to anything, whereas the PCs of 1999 had to be plugged into modems. And thanks to wireless nature of internet data transmissions, “high-bandwidth communication” is available in all but the remotest places in 2019, whereas it was only accessible at fixed-place computer terminals in 1999.
Again, Kurzweil’s use of the term “embedded” is troublesome, since it’s unclear how “high-bandwidth communication” could be embedded in anything. It emanates from and is received by things, and it is accessible in specific places, but it can’t be “embedded.” Given this and the other considerations, I think every part of Kurzweil’s prediction was correct in spirit, but that he was careless with how he worded it, and that it would have been better written as: “Computing and extremely high-bandwidth communication are available and accessible almost everywhere.”
“Cables have largely disappeared.”
MOSTLY RIGHT
Assessing the prediction requires us to deduce which kinds of “cables” Kurzweil was talking about. To my knowledge, he has never been an exponent of wireless power transfer and has never forecast that technology becoming dominant, so it’s safe to say his prediction didn’t pertain to electric cables. Indeed, larger computers like desktop PCs and servers still need to be physically plugged into electrical outlets all the time, and smaller computing devices like smartphones and tablets need to be physically plugged in to routinely recharge their batteries.
That leaves internet cables and data/power cables for peripheral devices like keyboards, mice, joysticks, and printers. On the first count, Kurzweil was clearly right. In 1999, WiFi was a new invention that almost no one had access to, and logging into the internet always meant sitting down at a computer that had some type of data plug connecting it to a wall outlet. Cell phones weren’t able to connect to and exchange data with the internet, except maybe for very limited kinds of data transfers, and it was a pain to use the devices for that. Today, most people access the internet wirelessly.
On the second count, Kurzweil’s prediction is only partly right. Wireless keyboards and mice are widespread, affordable, and are mature technologies, and even lower-cost printers meant for people to use at home usually come with integrated wireless networking capabilities, allowing people in the house to remotely send document files to the devices to be printed. However, wireless keyboards and mice don’t seem about to displace their wired predecessors, nor would it even be fair to say that the older devices are obsolete. Wired keyboards and mice are cheaper (they are still included in the box whenever you buy a new PC), easier to use since users don’t have to change their batteries, and far less vulnerable to hacking. Also, though they’re “lower tech,” wired keyboards and mice impose no handicaps on users when they are part of a traditional desktop PC setup. Wireless keyboards and mice are only helpful when the user is trying to control a display that is relatively far from them, as would be the case if the person were using their living room television as a computer monitor, or if a group of office workers were viewing content on a large screen in a conference room, and one of them was needed to control it or make complex inputs.
No one has found this subject interesting enough to compile statistics on the percentages of computer users who own wired vs. wireless keyboards and mice, but my own observation is that the older devices are still dominant.
And though average computer printers in 2019 have WiFi capabilities, the small “complexity bar” to setting up and using the WiFi capability makes me suspect that most people are still using a computer that is physically plugged into their printer to control the latter. These data cables could disappear if we wanted them to, but I don’t think they have.
This means that Kurzweil’s prediction that cables for peripheral computer devices would have “largely disappeared” by the end of 2019 was wrong. For what it’s worth, the part that he got right vastly outweighs the part he got wrong: The rise of wireless internet access has revolutionized the world by giving ordinary people access to information, services and communication at all but the remotest places. Unshackling people from computer terminals and letting them access the internet from almost anywhere has been extremely empowering, and has spawned wholly new business models and types of games. On the other hand, the world’s failure to fully or even mostly dispense with wired computer peripheral devices has been almost inconsequential. I’m typing this on a wired keyboard and don’t see any way that a more advanced, wireless keyboard would help me.
“The computational capacity of a $4,000 computing device (in 1999 dollars) is approximately equal to the computational capability of the human brain (20 million billion calculations per second).” [Or 20 petaflops]
WRONG
Graphics cards provide the most calculations per second at the lowest cost of any type of computer processor. The NVIDIA GeForce RTX 2080 Ti Graphics Card is one of the fastest computers available to ordinary people in 2019. In “overclocked” mode, where it is operating as fast as possible, it does 16,487 billion calculations per second (called “flops”).
A GeForce RTX 2080 retails for $1,100 and up, but let’s be a little generous to Kurzweil and assume we’re able to get them for $1,000.
$4,000 in 1999 dollars equals $6,164 in 2019 dollars. That means today, we can buy 6.164 GeForce RTX 2080 graphics cards for the amount of money Kurzweil specified.
6.164 cards x 16,487 billion calculations per second per card = 101,625 billion calculations per second for the whole rig.
This computational cost-performance level is two orders of magnitude worse than Kurzweil predicted.
Additionally, according to Top500.org, a website that keeps a running list of the world’s best supercomputers and their performance levels, the “Leibniz Rechenzentrum SuperMUC-NG” is the ninth fastest computer in the world and the fastest in Germany, and straddles Kurzweil’s line since it runs at 19.4 petaflops or 26.8 petaflops depending on method of measurement (“Rmax” or “Rpeak”). A press release said: “The total cost of the project sums up to 96 Million Euro [about $105 million] for 6 years including electricity, maintenance and personnel.” That’s about four orders of magnitude worse than Kurzweil predicted.
I guess the good news is that at least we finally do have computers that have the same (or slightly more) processing power as a single, average, human brain, even if the computers cost tens of millions of dollars apiece.
“Of the total computing capacity of the human species (that is, all human brains), combined with the computing technology the species has created, more than 10 percent is nonhuman.”
WRONG
Kurzweil explains his calculations in the “Notes” section in the back of the book. He first multiplies the computation performed by one human brain by the estimated number of humans who will be alive in 2019 to get the “total computing capacity of the human species.” Confusingly, his math assumes one human brain does 10 petaflops, whereas in his preceding prediction he estimates it is 20 petaflops. He also assumed 10 billion people would be alive in 2019, but the figure fell mercifully short and was ONLY 7.7 billion by the end of the year.
Plugging in the correct figure, we get (7.7 x 109 humans) x 1016 flops = 7.7 x 1025 flops = the actual total computing capacity of all human brains in 2019.
Determining the total computing capacity of all computers in existence in 2019 can only really be guessed at. Kurzweil estimated that at least 1 billion machines would exist in 2019, and he was right. Gartner estimated that 261 million PCs (which includes desktop PCs, notebook computers [seems to include laptops], and “ultramobile premiums”) were sold globally in 2019. The figures for the preceding three years were 260 million (2018), 263 million (2017), and 270 million (2016). Assuming that a newly purchased personal computer survives for four years before being fatally damaged or thrown out, we can estimate that there were 1.05 billion of the machines in the world at the end of 2019.
However, Kurzweil also assumed that the average computer in 2019 would be as powerful as a human brain, and thus capable of 10 petaflops, but reality fell far short of the mark. As I revealed in my analysis of the preceding prediction, a 10 petaflop computer setup would cost somewhere between $606,543 in GeForce RTX 2080 graphics cards, or $52.5 million for half a Leibniz Rechenzentrum SuperMUC-NG supercomputer. None of the people who own the 1.34 billion personal computers in the world spent anywhere near that much money, and their machines are far less powerful than human brains.
Let’s generously assume that all of the world’s 1.05 billion PCs are higher-end (for 2019) desktop computers that cost $900 – $1,200. Everyone’s machine has an Intel Core i7, 8th Generation processor, which offers speeds of a measly 361.3 gigaflops (3.613 x 1011 flops). A 10 petaflop human brain is 27,678 times faster!
Plugging in the computer figures, we get (1.05 x 109 personal computers) x 3.61311 flops = 3.794 x 1020 = the total computing capacity of all personal computers in 2019. That’s five orders of magnitude short. The reality of 2019 computing definitely fell wide of Kurzweil’s expectations.
What if we add the computing power of all the world’s smartphones to the picture? Approximately 3.2 billion people owned a smartphone in 2019. Let’s assume all the devices are higher-end (for 2019) iPhone XR’s, which everyone bought new for at least $500. The iPhone XR’s have A12 Bionic processors, and my research indicates they are capable of 700 – 1,000 gigaflop maximum speeds. Let’s take the higher-end estimate and do the math.
3.2 billion smartphones x 1012 flops = 3.2 x 1021 = the the total computing capacity of all smartphones in 2019.
Adding things up, pretty much all of the world’s personal computing devices (desktops, laptops, smartphones, netbooks) only produce 3.5794 x 1021 flops of computation. That’s still four orders of magnitude short of what Kurzweil predicted. Even if we assume that my calculations were too conservative, and we add in commercial computers (e.g. – servers, supercomputers), and find that the real amount of artificial computation is ten times higher than I thought, at 3.5794 x 1022 flops, this would still only be equivalent to 1/2000th, or 0.05% of the total computing capacity of all human brains (7.7 x 1025 flops). Thus, Kurzweil’s prediction that it would be 10% by 2019 was very wrong.
“Rotating memories and other electromechanical computing devices have been fully replaced with electronic devices.”
WRONG
For those who don’t know much about computers, the prediction says that rotating disk hard drives will be replaced with solid-state hard drives that don’t rotate. A thumbdrive has a solid-state hard drive, as do all smartphones and tablet computers.
I gauged the accuracy of this prediction through a highly sophisticated and ingenious method: I went to the nearest Wal-Mart and looked at the computers they had for sale. Two of the mid-priced desktop PCs had rotating disk hard drives, and they also had DVD disc drives, which was surprising, and which probably makes the “other electromechanical computing devices” part of the prediction false.
If the world’s biggest brick-and-mortar retailer is still selling brand new computers with rotating hard disk drives and rotating DVD disc drives, then it can’t be said that solid state memory storage has “fully replaced” the older technology.
“Three-dimensional nanotube lattices are now a prevalent form of computing circuitry.”
MOSTLY WRONG
Many solid-state computer memory chips, such as common thumbdrives and MicroSD cards, have 3D circuitry, and it is accurate to call them “prevalent.” However, 3D circuitry has not found routine use in computer processors thanks to unsolved problems with high manufacturing costs, unacceptably high defect rates, and overheating.
In late 2018, Intel claimed it had overcome those problems thanks to a proprietary chip manufacturing process, and that it would start selling the resulting “Lakefield” line of processors soon. These processors have four, vertically stacked layers, so they meet the requirement for being “3D.” Intel hasn’t sold any yet, and it remains to be seen whether they will be commercially successful.
Silicon is still the dominant computer chip substrate, and carbon-based nanotubes haven’t been incorporated into chips because Intel and AMD couldn’t figure out how to cheaply and reliably fashion them into chip features. Nanotube computers are still experimental devices confined to labs, and they are grossly inferior to traditional silicon-based computers when it comes to doing useful tasks. Nanotube computer chips that are also 3D will not be practical anytime soon.
It’s clear that, in 1999, Kurzweil simply overestimated how much computer hardware would improve over the next 20 years.
“The majority of ‘computes’ of computers are now devoted to massively parallel neural nets and genetic algorithms.”
UNCLEAR
Assessing this prediction is hard because it’s unclear what the term “computes” means. It is probably shorthand for “compute cycles,” which is a term that describes the sequence of steps to fetch a CPU instruction, decode it, access any operands, perform the operation, and write back any result. It is a process that is more complex than doing a calculation, but that is still very basic. (I imagine that computer scientists are the only people who know, offhand, what “compute cycle” means.)
Assuming “computes” means “compute cycles,” I have no idea how to quantify the number of compute cycles that happened, worldwide, in 2019. It’s an even bigger mystery to me how to determine which of those compute cycles were “devoted to massively parallel neural nets and genetic algorithms.” Kurzweil doesn’t describe a methodology that I can copy.
Also, what counts as a “massively parallel neural net”? How many processor cores does a neutral net need to have to be “massively parallel”? What are some examples of non-massively parallel neural nets? Again, an ambiguity with the wording of the prediction frustrates an analysis. I’d love to see Kurzweil assess the accuracy of this prediction himself and to explain his answer.
“Significant progress has been made in the scanning-based reverse engineering of the human brain. It is now fully recognized that the brain comprises many specialized regions, each with its own topology and architecture of interneuronal connections. The massively parallel algorithms are beginning to be understood, and these results have been applied to the design of machine-based neural nets.”
PARTLY RIGHT
The use of the ambiguous adjective “significant” gives Kurzweil an escape hatch for the first part of this prediction. Since 1999, brain scanning technology has improved, and the body of scientific literature about how brain activity correlates with brain function has grown. Additionally, much has been learned by studying the brain at a macro-level rather than at a cellular level. For example, in a 2019 experiment, scientists were able to accurately reconstruct the words a person was speaking by analyzing data from the person’s brain implant, which was positioned over their auditory cortex. Earlier experiments showed that brain-computer-interface “hats” could do the same, albeit with less accuracy. It’s fair to say that these and other brain-scanning studies represent “significant progress” in understanding how parts of the human brain work, and that the machines were gathering data at the level of “brain regions” rather than at the finer level of individual brain cells.
Yet in spite of many tantalizing experimental results like those, an understanding of how the brain produces cognition has remained frustratingly elusive, and we have not extracted any new algorithms for intelligence from the human brain in the last 20 years that we’ve been able to incorporate into software to make machines smarter. The recent advances in deep learning and neural network computers–exemplified by machines like AlphaZero–use algorithms invented in the 1980s or earlier, just running on much faster computer hardware (specifically, on graphics processing units originally developed for video games).
If anything, since 1999, researchers who studied the human brain to gain insights that would let them build artificial intelligences have come to realize how much more complicated the brain was than they first suspected, and how much harder of a problem it would be to solve. We might have to accurately model the brain down the the intracellular level (e.g. – not just neurons simulated, but their surface receptors and ion channels simulated) to finally grasp how it works and produces intelligent thought. Considering that the best we have done up to this point is mapping the connections of a fruit fly brain and that a human brain is 600,000 times bigger, we won’t have detailed human brain simulation for many decades.
“It is recognized that the human genetic code does not specify the precise interneuronal wiring of any of these regions, but rather sets up a rapid evolutionary process in which connections are established and fight for survival. The standard process for wiring machine-based neural nets uses a similar genetic evolutionary algorithm.”
RIGHT
This prediction is right, but it’s not noteworthy since it merely re-states things that were widely accepted and understood to be true when the book was published in 1999. It’s akin to predicting that “A thing we think is true today will still be considered true in 20 years.”
The prediction’s first statement is an odd one to make since it implies that there was ever serious debate among brain scientists and geneticists over whether the human genome encoded every detail of how the human brain is wired. As Kurzweil points out earlier in the book, the human genome is only about 3 billion base-pairs long, and the genetic information it contains could be as low as 23 megabytes, but a developed human brain has 100 billion neurons and 1015 connections (synapses) between those neurons. Even if Kurzweil is underestimating the amount of information the human genome stores by several orders of magnitude, it clearly isn’t big enough to contain instructions for every aspect of brain wiring, and therefore, it must merely lay down more general rules for brain development.
I also don’t understand why Kurzweil wrote the second part of the statement. It’s commonly recognized that part of childhood brain development involves the rapid paring of interneuronal connections that, based on interactions with the child’s environment, prove less useful, and the strengthening of connections that prove more useful. It would be apt to describe this as “a rapid evolutionary process” since the child’s brain is rewiring to adapt to child to its surroundings. This mechanism of strengthening brain connection pathways that are rewarded or frequently used, and weakening pathways that result in some kind of misfortune or that are seldom used, continues until the end of a person’s life (though it gets less effective as they age). This paradigm was “recognized” in 1999 and has never been challenged.
Machine-based neural nets are, in a very general way, structured like the human brain, they also rewire themselves in response to stimuli, and some of them use genetic algorithms to guide the rewiring process (see this article for more info: https://news.mit.edu/2017/explained-neural-networks-deep-learning-0414). However, all of this was also true in 1999.
“A new computer-controlled optical-imaging technology using quantum-based diffraction devices has replaced most lenses with tiny devices that can detect light waves from any angle. These pinhead-sized cameras are everywhere.”
WRONG
Devices that harness the principle of quantum entanglement to create images of distant objects do exist and are better than devices from 1999, but they aren’t good enough to exit the R&D labs. They also have not been shrunk to pinhead sizes. Kurzweil overestimated how fast this technology would develop.
Virtually all cameras still have lenses, and still operate by the old method of focusing incoming light onto a physical medium that captures the patterns and colors of that light to form a stored image. The physical medium used to be film, but now it is a digital image sensor.
Digital cameras were expensive, clunky, and could only take low-quality images in 1999, so most people didn’t think they were worth buying. Today, all of those deficiencies have been corrected, and a typical digital camera sensor plus its integrated lens is the size of a small coin. As a result, the devices are very widespread: 3.2 billion people owned a smartphone in 2019, and all of them probably had integral digital cameras. Laptops and tablet computers also typically have integral cameras. Small standalone devices, like pocket cameras, webcams, car dashcams, and home security doorbell cameras, are also cheap and very common. And as any perusal of YouTube.com will attest, people are using their cameras to record events of all kinds, all the time, and are sharing them with the world.
This prediction stands out as one that was wrong in specifics, but kind of right in spirit. Yes, since 1999, cameras have gotten much smaller, cheaper, and higher-quality, and as a result, they are “everywhere” in the figurative sense, with major consequences (good and bad) for the world. Unfortunately, Kurzweil needlessly stuck his neck out by saying that the cameras would use an exotic new technology, and that they would be “pinhead-sized” (he hurt himself the same way by saying that the augmented reality glasses of 2019 would specifically use retinal projection). For those reasons, his prediction must be judged as “wrong.”
“Autonomous nanoengineered machines can control their own mobility and include significant computational engines. These microscopic machines are beginning to be applied to commercial applications, particularly in manufacturing and process control, but are not yet in the mainstream.”
WRONG
While there has been significant progress in nano- and micromachine technology since 1999 (the 2016 Nobel Prize in Chemistry was awarded to scientists who had invented nanomachines), the devices have not gotten nearly as advanced as Kurzweil predicted. Some microscopic machines can move around, but the movement is guided externally rather than autonomously. For example, turtle-like micromachines invented by Dr. Marc Miskin in 2019 can move by twirling their tiny “flippers,” but the motion is powered by shining laser beams on them to expand and contract the metal in the flippers. The micromachines lack their own power packs, lack computers that tell the flippers to move, and therefore aren’t autonomous.
In 2003, UCLA scientists invented “nano-elevators,” which were also capable of movement and still stand as some of the most sophisticated types of nanomachines. However, they also lacked onboard computers and power packs, and were entirely dependent on external control (the addition of acidic or basic liquids to make their molecules change shape, resulting in motion). The nano-elevators were not autonomous.
Similarly, a “nano-car” was built in 2005, and it can drive around a flat plate made of gold. However, the movement is uncontrolled and only happens when an external stimulus–an input of high heat into the system–is applied. The nano-car isn’t autonomous or capable of doing useful work. This and all the other microscopic machines created up to 2019 are just “proof of concept” machines that demonstrate mechanical principles that will someday be incorporated into much more advanced machines.
Significant progress has been made since 1999 building working “molecular motors,” which are an important class of nanomachine, and building other nanomachine subcomponents. However, this work is still in the R&D phase, and we are many years (probably decades) from being able to put it all together to make a microscopic machine that can move around under its own power and will, and perform other operations. The kinds of microscopic machines Kurzweil envisioned don’t exist in 2019, and by extension are not being used for any “commercial applications.”
Whew! That’s it for now. I’ll try to publish PART 2 of this analysis next month. Until then, please share this blog entry with any friends who might be interested. And if you have any comments or recommendations about how I’ve done my analysis, feel free to comment.
Links:
- Ray Kurzweil’s self-analysis of how accurate his 2009 predictions were: https://kurzweilai.net/images/How-My-Predictions-Are-Faring.pdf
- The inventor of the first augmented reality contact lenses predicted in 2015 that commercially viable versions of the devices wouldn’t exist for at least 20 more years. (https://www.inverse.com/article/31034-augmented-reality-contact-lenses)
- In late 2019, a Magic Leap One cost $2,300 – $3,300 and a Hololens was $3,000. https://www.cnn.com/2019/12/10/tech/magic-leap-ar-for-companies/index.html
- In 2019, a new Oculus Rift system cost $400 – $500, and a new HTC Vive was $500 – $800. (https://www.theverge.com/2019/5/16/18625238/vr-virtual-reality-headsets-oculus-quest-valve-index-htc-vive-nintendo-labo-vr-2019)
- In 2018, people across the world bought 259 million new desktop computers, laptops, and “ultramobile” devices (higher-end tablets that have large, detachable keyboards [the Microsoft Surface dominates this category]). These machines are meant to be accessed with traditional keyboard and mouse inputs. Keyboards aren’t dead.
(https://venturebeat.com/2019/01/10/gartner-and-idc-hp-and-lenovo-shipped-the-most-pcs-in-2018-but-total-numbers-fell/) - Survey data from 2018 about the global usage of “digital personal assistants.” Users speak to their smartphones or smart speakers, mostly to obtain simple information (like weather forecasts) or to have their computers do simple tasks. (https://www.business2community.com/infographics/the-growth-in-usage-of-virtual-digital-assistants-infographic-02056086)
- 2019 Pew Survey showing that the overwhelming majority of American adults owned a smartphone or traditional PC. People over age 64 were the least likely to own smartphones. (https://www.pewresearch.org/internet/fact-sheet/mobile/)
- A 2015 American Community Survey revealed that households headed by people over 64 were the least likely to have smartphones, PCs, or internet access. (https://www.census.gov/content/dam/Census/library/publications/2017/acs/acs-37.pdf)
- In 2000, 34% of Americans accessed the internet through dial-up modems, and only 3% did so through “broadband” (a catch-all for cable, DSL, and satellite access). Most U.S. internet users were still using dial-up modems that were at most 56k. The remaining 63% didn’t access it at all. (http://thetechnews.com/2016/01/03/usa-getting-faster-internet-speeds-but-not-at-the-pace-others-are/)
- In 2019, a mid-tier internet service plan in the U.S. granted users download speeds of 30 – 60 Mbps. (https://www.pcmag.com/news/state-by-state-the-fastest-and-slowest-us-internet)
- 2019 U.S. mobile phone network average speeds were 33.88 Mbps for downloads and 9.75 Mbps for uploads (https://www.speedtest.net/reports/united-states/ )
- The Black Friday 2019 circular for Newegg.com featured five models of printers for sale. Only one of them, the Brother HL-L2300D, wasn’t WiFi-capable. (https://bestblackfriday.com/ads/newegg-black-friday/page-12#ad_view)
- Gartner figures for global computer sales in 2015, 2016, 2017, 2018 and 2019.
(https://www.gartner.com/en/newsroom/press-releases/2017-01-11-gartner-says-2016-marked-fifth-consecutive-year-of-worldwide-pc-shipment-decline)
(https://venturebeat.com/2018/01/11/gartner-and-idc-agree-hp-shipped-the-most-pcs-in-2017/)
(https://www.gartner.com/en/newsroom/press-releases/2020-01-13-gartner-says-worldwide-pc-shipments-grew-2-point-3-percent-in-4q19-and-point-6-percent-for-the-year) - Intel’s i7 Generation 8 processor is capable of 361.3 gigaflop speeds. (https://www.pugetsystems.com/labs/hpc/Skylake-X-7800X-vs-Coffee-Lake-8700K-for-compute-AVX512-vs-AVX2-Linpack-benchmark-1068/)
- 3.2 billion people owned a smartphone in 2019. (https://newzoo.com/insights/trend-reports/newzoo-global-mobile-market-report-2019-light-version/)
- In 2019, 3D chips were common in memory storage devices, like MicroSD cards. 3D NAND chips had up to 64 layers. (https://semiengineering.com/what-happened-to-nanoimprint-litho/)
- In 2019, Intel was still working the kinks out of its first 3D computer processor, called “Lakefield,” and it wasn’t ready for commercial sales. (https://www.overclock3d.net/news/cpu_mainboard/intel_details_their_lakefield_processor_design_and_foveros_3d_packaging_tech/1)
- In 2019, computer circuits made of carbon nanotubules were still stuck in research labs, and held back from commercialization by many unsolved problems relating to cost of manufacture and reliability. Silicon was still the dominant computing substrate. (https://www.sciencenews.org/article/chip-carbon-nanotubes-not-silicon-marks-computing-milestone)
- “Compute cycle” has three meanings: #1 (https://www.zdnet.com/article/how-much-is-a-unit-of-cloud-computing/), #2 (https://www.quora.com/What-is-a-Compute-cycle) and #3 (https://www.computerhope.com/jargon/c/compute.htm)
- In a 2019 experiment, researchers were able to decode the words a person was speaking by studying their brain activity. (https://www.biorxiv.org/content/10.1101/350124v2)
- “The current ways of trying to represent the nervous system…[are little better than] what we had 50 years ago.” –Marvin Minsky, 2013 (https://youtu.be/3PdxQbOvAlI)
- “Today’s neural nets use algorithms that were essentially developed in the early 1980s.” (https://futurism.com/cmu-brain-research-grant)
- The inventor of “back-propagation,” which spawned many computer algorithms central to AI research, now believes it will never lead to true intelligence, and that the human brain doesn’t use it. (https://www.axios.com/artificial-intelligence-pioneer-says-we-need-to-start-over-1513305524-f619efbd-9db0-4947-a9b2-7a4c310a28fe.html)
- Henry Markram’s project to create a human brain simulation by 2019 failed. (https://www.theatlantic.com/science/archive/2019/07/ten-years-human-brain-project-simulation-markram-ted-talk/594493/)
- “Like, yes, in particular areas machines have superhuman performance, but in terms of general intelligence we’re not even close to a rat.” –Yann LeCun, 2017 (https://www.theverge.com/2017/10/26/16552056/a-intelligence-terminator-facebook-yann-lecun-interview)
- Machine neural networks are similar to human brains in key ways. (https://news.mit.edu/2017/explained-neural-networks-deep-learning-0414)
- Some machine neural nets use genetic algorithms. (https://blog.coast.ai/lets-evolve-a-neural-network-with-a-genetic-algorithm-code-included-8809bece164)
- Quantum imaging is a real thing. However, devices that can make use of it are still experimental. (https://onlinelibrary.wiley.com/doi/full/10.1002/lpor.201900097)
- The Samsung Galaxy S10 is an upper-end smartphone released in 2019. It has three digital cameras, all of which operate on the same technology principles as the digital cameras of 1999. (https://www.digitalcameraworld.com/reviews/samsung-galaxy-s10-camera-review)
- The 2016 Nobel Prize in Chemistry was given to three scientists who had done pioneering work on nanomachines. (https://www.extremetech.com/extreme/237575-2016-nobel-prize-in-chemistry-awarded-for-nanomachines)
- Dr. Marc Miskin’s micromachines from 2019 are interesting, but a far cry from what Kurzweil thought we’d have by then. (https://www.inquirer.com/health/micro-robots-upenn-cornell-20190307.html)
