In a previous article that was about useful-to-know Achilles’ heels of the current ‘second wave of AI’, I mentioned that these Achilles’ heels are all of a single making: the fact that our underlying technology is digital, or more correctly: discrete. Current neural networks for instance are just ‘data driven rule based systems in disguise’ (with hidden rules). In the end, a digital computer is a Turing machine and the fact that I am programming it differently doesn’t change its fundamental capabilities. And these capabilities have limitations, as the kind of intelligence we want is not logical rule based and can also not be emulated with logical rule-based approaches.
That is why there are scientists all over the world hard at work at creating computing with a radically different foundation: the operations known as ‘quantum mechanics’, the weird behaviour that underlies the reality that we experience every day. For instance, the fact that you can rest your hand on a seemingly solid table, while in fact it is mostly made up of ‘nothing’, is the result of the ‘Pauli exclusion principle’, one of the interesting rules of behaviour of nature at the quantum level.
All this scientific work captures the imagination of many, and people may even feel that it will not be long before quantum computing takes over the world. You may find a lot of this on tellingly titled blogs like ‘next big future‘ where for instance in the outlook for 2018-2028 it is simply stated: “artificial Intelligence and quantum computers will transform IT and logistics”.
Will it? Maybe it is a good idea for another dose of realism. Let’s start with the digital computers that are all-pervasive these days and are the machines that work the current wonders (and Achilles’ heels) of AI.
Digital computing
Digital computers are ‘classical logic machines’ or as one of the founders of computing Alan Turing would have it “discrete state machines”. Which are
“machines which move by sudden jumps or clicks from one quite definite state to another. These states are sufficiently different for the possibility of confusion between them to be ignored. Strictly speaking there, are no such machines. Everything really moves continuously. But there are many kinds of machine which can profitably be thought of as being discrete-state machines. For instance in considering the switches for a lighting system it is a convenient fiction that each switch must be definitely on or definitely off. There must be intermediate positions, but for most purposes we can forget about them.”
Digital computers have many, many states, but all these states are clearly separated from each other, the way characters are separated from each other, there is no overlap, something is either an ‘a’ or a ‘b’, there is no ‘halfway a-b’ character or a character that is ‘both a and b concurrently’. (Incidentally: I sometimes ask people how old they think that digital technology is. They never give what in my view the correct answer is: a few thousand years. Any alphabet is discrete and as such digital). This discreteness is what underlies all our digital technology. Everything in digital technology is based on this foundation. And this foundation, again, is just machinery that can operate what we call ‘classical logic’ on ‘discrete states’. True or false everything is, and nothing else. One could argue that writing is more reliable than oral reports because of the fact that the discreteness of alphabets makes transport reliable over space and time. There is more to that line of thought, but I am digressing, as usual. Anyway, digital computers handle only discrete states, they just handle integers using classical logic and nothing else.
You might be forgiven to think that that is not true. After all, won’t computers easily handle ‘real numbers’? Can’t I work with numbers like 2.47 and 3.1788835552 in my spreadsheet, and doesn’t the computer handle these perfectly? But dig deep and it turns out that digital computers play a trick on you. They calculate with integers. They fake real numbers with representing them as ‘exponentiation’ using integers as a base and integers as an exponent. And they’re so good at it that we tend to assume that they are handling real numbers. But they are not. They may represent a lot of numbers exactly, but not all, which means that whatever the mechanism, you will see rounding errors (some older background here).
There were analog computers before there were digital computers and they have even somewhat resurged. Where a digital computer has bits that are either 0 or 1 (making it discrete), analog computers have values that can be anything between two values (e.g. 0 and 1). (Technically, digital computers also have states between 0 and 1, for instance while they are switching from one to the other, but the digital computer has means (the ‘clock’) to ignore values during these intermediary states. As Turing already said, physically, digital computers exist because we ‘forget’ about the states where the situation is between 0 and 1.). But the hope of many is directed not at analog computers but quantum computers, computers employing the quantum physics that underlies our reality at the smallest scale.
Quantum computing
Quantum computers as they are envisioned now are a very special kind of digital computer. Instead of each bit begin either 0 or 1 and nothing else, a qubit (the ‘bit’ in quantum computing) can during operation be 0 and 1 at the same time (a quantum-specific state called ‘superposition’). In such a state, each qubit has a probability of being 0 and a probability of being 1, the sum of which is 1. The moment a qubit is ‘measured’ it becomes either a 0 or 1. Quantum computing is about not ignoring what happens when the bit is not 0 or 1, in combination with the weird ‘being two things at the same time’ behaviour of the smallest scale of our everyday reality.
The intuitive but false story about QM computing goes like this:
Being in two states at once provides interesting possibilities for parallelism. Suppose you want to guess a 8-bit number, say an encryption key, that is a value of between 0 and 255. In a 8-bit digital computer, you have to try them all one by one. So, you start with the number 0 (8 zero-bits, or 00000000), then if that fails with the number 1 (7 zero bits and one one bit, or 00000001) etc. If the number you have to guess is 255 you have to do your operation 256 times, on average you have to do it 128 times.
Now, suppose the number to guess (e.g. an encryption key, or the best route of a possible 256 routes for a delivery car to deliver all its packets) is 153, or “10011001” in binary notation. In a digital 8-bit computer you have to make 154 tries before you hit on the right answer. But if your computer is a quantum computer, you start with one 8-qubit ‘superposition’, or “{0 or 1}{0 or 1}{0 or 1}{0 or 1}{0 or 1}{0 or 1}{0 or 1}{0 or 1}” and in one step, the computer finds the answer {0 or 1}{0 or 1}{0 or 1}{0 or 1}{0 or 1}{0 or 1}{0 or 1}{0 or 1}. On average, your 8-bit quantum computer will be 128 times as fast as an 8-bit digital computer. In real world terms, where computers are 64bit, a 64-bit quantum computer will be 2 to the power of 63 times as fast, or 4,611,686,018,427,387,904 times as fast in this kind of operation. So, the more bits, the more impressive the gain, because of the massive parallelism that superposition brings. And that is only for one single operation. What if the outcome of the first operation is input for the second operation? The potential is — in theory — almost infinite.
But the story is wrong. There are several huge flies in the ointment.
The true story about QM computing goes like this (inspired by and sometimes paraphrasing writings by Scott Aaronson, see here, here, here and here):
First the basics:
I am simplifying the story, so it is technically not perfectly correct, but it does give the right idea.
An, at first seemingly small, fly in the ointment is that qubits are very hard to make. Most technologies will have to work at near absolute-zero temperatures. So, making lots of them is a problem. This is called the width of the quantum computer. Scientists can now make roughly 70 qubits working together. The second fly in the ointment is that our qubits are fundamentally unstable. We need to work on our qubits in isolation, but this isolation cannot be perfect. So, we only have a limited number of quantum operations we can perform before the whole thing crashes. The number of operations you can perform before the thing crashes is called the depth of a quantum computer. And the larger the width, the harder depth becomes. We don’t know if this problem is even solvable. There might be a hard limit around the corner.
To make matters worse, there is also fundamentally a non-negligible chance to read (measure) a qubit wrong or for quantum operations (the steps in a quantum algorithm) to operate incorrectly, which brings us to another fly: the error rates: Quantum computing is fundamentally uncertain and outcomes cannot be trusted perfectly. The way to get around this is for instance to calculate many times (formally, you need to have something that is more than two thirds of the time correct to do it this way). Another way is to have a digital computer to check the result. The latter is for instance possible when you are after answers that are hard to find but easy to check, e.g. an encryption key. So, what we preferably need is not just a quantum computer, but an error-correcting quantum computer. To have enough room for quantum error correction we need lots and lots of qubits, which brings us back to fly number one.
But these two, potentially even unsolvable, issues are not even the real problem. For the real problem, we need to dive into the actual working of a quantum computer and why the wrong explanation above is actually wrong.
A true part of the ‘wrong story’ is this: while a 64-bit number in a digital computer can have 2 to the power of 64, or 18,446,744,073,709,551,616 different values, it can only have one such a value at any time. A 64-qubit computer can — while it is calculating — have all these values simultaneously ‘in a single 64-qubit memory’. A comparable 64-bit computer would have to have not one single 64-bit number to operate on, but it would have to have 288,230,376,151,711,744 64-bit values to operate on. Simultaneously. A bit like having 288,230,376,151,711,744 CPUs in your computer. With current digital technology, such a computer would weigh more than 11,529,215,046,068 kilograms and use 288,23,037,615,171 Megawatt of electrical power. That’s 14 billion times the total average power that humanity currently uses. For a single computer. That’ll definitely put the final nail in the climate coffin.
So, the 64-qubit computer is extremely efficient while it is operating on all the values in parallel. But then comes reading the calculation result, what in a QM computing is the measurement of the state of those qubits. And then we run into quantum mechanics’ fundamental property: the measurement result is random. In quantum mechanics, individual events cannot be predicted, only statistical averages can. So, when we read a 64-bit quantum computer’s 64-bit value at the end of operation, we get a random number, not necessarily the answer we are looking for.
Here is why: suppose we have a qubit that can be in a mixed state of UP and DOWN. Each state can be thought of as a wave. The amplitudes of these waves, when squared, give the probability of that state. And these probabilities (the amplitudes squared) add up to 100%. At all times, a qubit has a probability for each of its two possible states, all chances together adding up to 1 (or 100%). But when you measure it, it randomly (but with a chance given by those squared amplitudes) becomes one of the states. In other words, when the qubit is in a state where it has a UP amplitude of 0.3, it has a 0.09 (or 9%) chance of becoming UP when measured. It thus has a 91% (or 0.91) chance of becoming DOWN when measured. And that means the DOWN amplitude is roughly plus or minus 0.95. 0.3 squared is 0.09. Roughly, 0.95 squared is 0.91. Which means that even with a DOWN amplitude of 0.95, measuring it has a 9% chance of producing UP when it is read. (compare to this: a bit on digital media that physically produces a 95% of the amplitude of a 1 bit reads as a 1 with perfect certainty).
So, when you have a 64-qubit QM computer, you will measure the computation result as a pretty random 64-bit digital number, unless your quantum algorithm is such that the probability of reading the wrong answer is zero. The operation of the steps of the algorithm must lead to a state of all those qubits combined, where every qubit that should be 0 for the correct answer must be so with 100% certainty and every qubit that should read as 1 for the correct answer must be so with 100% certainty: unless your QM algorithm makes sure all the wrong answers have a zero probability of being the read result, measuring the state of your quantum computing (reading the answer) will likely give you some wrong random result. This is the main reason why the ‘intuitive but false’ explanation above, and the number 18,446,744,073,709,551,616 is wrong. Your quantum algorithm that is made up of a successive number of quantum operations on those qubits needs to make sure that the chance of reading (measuring) a wrong answer is (almost) zero.
So, what we need is an algorithm that does precisely this: have quantum operations on our qubits such that — because during operation all the wrong states for qubits have canceled out — only a right answer can be measured (almost) deterministically at the end. This is possible because the waves that define the qubit may interfere with each other based on operations on the qubit. These interferences may be destructive (canceling out) or constructive (boosting), just as any wave can interfere with another wave. That way, operations on qubit can drive it to either a 0% or a 100% probability to being measured in a certain state. A qubit interferes with itself and with the operations performed on it. An operation on a qubit, by the way, is something like shooting a laser or a radio wave at it.
And this need to have algorithms that ‘cancel out’ wrong answers really limits what you can do with quantum computers, because there are only a few specific problems for which such algorithms exist. One such an algorithm is famous. It is called Shor’s Algorithm and it is a quantum algorithm that can factor integers in reasonable time. Given that much of our encryption and thus security is based on the assumption that it takes an impossible amount of time to factor very large integers (say, the 1300 digits of modern day encryption values), this quantum algorithm has of course spiked a lot if interest. But there are not many others and it is unclear if there will be many.
To drive the message home: a quantum computer is not some sort of super fast general computer. It is a computer for a very limited and specific set of problems, and we only have a very few algorithms that may do something useful. What problems? Which algorithms? Much of that is still up in the air. The most likely use is probably emulation (calculation) of quantum physics itself, as those calculations often stymie digital computers.
Where we are today
I’m ignoring the D-WAVE ‘quantum annealing‘ system in this story. Mainly because it is pretty much unclear if such a system actually delivers any speedup from true quantum mechanical effects. Researchers so far have come up empty. It is questionable if it is really a quantum computer, even if it might use some quantum mechanical effects. It might be better seen as dedicated hardware (potentially QM in nature) to solve a useful but very limited type of calculation. On me, but I’m no expert, with it’s optimum-seeking behaviour it evokes more the image of an analog computer.
So, where do we stand in terms of actual real quantum computing? Well, in 2011 Shor’s Algorithm has been able to factor 21 into 3 times 7 using 2 qubits. Wow, Bob, wow. With a mechanism that is better than Shor’s Algorithm, a 4-qubit quantum computer has been able to factor the number 143 (11×13). Later researchers proved that that algorithm could actually factor larger numbers as long as those factors only differed in two bits (what researchers then call ‘a special class of numbers’). 56153 was in 2014 the largest number factored, though with another quantum computing technology than the one used for Shor’s. We’re still really far from factoring today’s 1300-digit numbers at the root of encryption in reasonable time.
IBM recently published research that there is a class of problems for which you do not need as much depth (number of steps) on a quantum computer as you would need for the same problem on a digital computer. This has widely been reported on the net as “IBM proves a quantum computing advantage over digital computing” and you may walk away with the idea that quantum computing is around the corner. But in fact what they have found is — again — a small set of problems that may not require as much depth (number of steps) on quantum computing as it does on digital computing. Yes, less steps is an advantage. But small set (of what already is a small set) is a definitive disadvantage. And is there actually a problem in that set that has a practical application? We don’t know. Maybe. If we are lucky.
Most of quantum algorithms still run on simulations of actual quantum computing hardware, simulations that run on digital computers who do not have the actual capability of truly simulating the underlying hardware because of digital’s limitations with respect to both quantum’s superpositions (so the simulations are slow) as well as with respect to the actual real math that is required to model the underlying physics.
And what about actual quantum computing hardware? Well, Google is testing a 72 qubit quantum processor with a 1% readout error rate, and it hopes to be able to keep it coherent for long enough to get over 40 steps from it (depth). That may get them a quantum computing that for a very specific type of problem would outperform a digital computer for the first time. Sounds, great, doesn’t it?
Not so fast. Have a look at the following image from Google’s AI blog:
The green area is where the actual useful error-corrected quantum computing is. We need about 100,000 as many qubits per processor as we have today, and in a coherent state for an extended time period, so algorithms can have some depth. We need one to two orders of improvement on the error rate. The blue area might be in reach, however, so some very specific problems may see quantum computing outperform digital computing sooner. Again, the graph is what Google intends, not what it has done. A bit like still not having solved that ‘labeling dark people as gorilla’s‘ problem in three years.
So, not only are what we see now mostly intentions, there are several other hurdles and important notes.
- As mentioned above: we may be up against fundamental physical limitations, comparable with the hard limit of light speed. I.e.: can we overcome decoherence at larger scales? In the words of a 2012 paper: “Is controlling large-scale quantum systems merely really, really hard, or is it ridiculously hard?” and follows up with “In the former case we might succeed in building large-scale quantum computers after a few decades of very hard work. In the latter case we might not succeed for centuries, if ever”
- What kind of problems are actually feasible to calculate using quantum algorithms (and better than digital)? The set is interesting, but also limited. A quantum computer will not be a general purpose computer, it will have specific advantages for specific mathematical problems. From that same article on one such class of problems: “It seems that […] speedups are possible only for problems with special structure well matched to the power of a quantum computer”. In fact, what quantum computers may be best at is doing calculations about quantum physics. A bit like the analog computers that were electronic analogs for mechanical problems (such as target calculation for large guns), and definitely not what is in the mind of hype-peddlers.
- And then there are the problems where we cannot check correct operation of a quantum computer at all other than running it on another quantum computer or performing a real world experiment. That kind of sounds like the ‘hidden proxy’ problem of the previous article. Yes, the answer is 42, but we have no idea what it means.
- The necessity of error correction is one of those areas where we quickly descend into the edge of knowledge or even unknown physics. To create states that both show quantum coherence and stability, theorists dive into worlds of two-dimensional quasi-particles or even worlds with four spatial dimensions. Um.
- And even without that, assumptions that have been made based on approximating mathematical models (physicists almost always have to throw out ‘negligible’ terms or they aren’t able to calculate anything at all) may not hold in reality where these ‘negligible’ terms interact and are not independent. This may have detrimental effects on the possibility of real quantum error correction. Things may be much harder in reality than approximate calculations so far have shown (from a Jan 2019 article by some very smart Finns and Germans). Another example of how using approximations (just like the ones used by digital computers to approximate real numbers) can be fraught with dangers.
As it turns out, some of the open questions require physics that we don’t have yet and trying to build quantum computing may in a way be seen as an experiment to get at new underlying fundamental physics regarding the reality we are part of, only formulated in a way that large companies such as Microsoft, Google, and IBM are willing to put a lot of money in it. You have to admire the physicists for their capability of getting money for physics. Or maybe it is just that in the end every physical subject (including computing machinery) becomes a physics subject.
So, before believing statements like “[before 2028] […] quantum computers will transform IT and logistics“, it’s best not to put up our hopes too high for that foreseeable time period. I’m not a big fan of trying to predict the future (you’re lying, even if you’re telling the truth) but for me: this is probably going to take decades of hard work and success is far from certain. That statement is pure hype. Really.
[This article appeared before in a slightly different form on Architecture for Real Enterprises on InfoWorld. It is also available on R&A with an appendix on the continuum hypothesis]