67 karma · joined May 1, 2020
Edit: Only a fraction of the qubits of the 127 qubit system were calibrated when I looked.
The circuit topology, a heater coupled to a superconducting wire to sense the local tempreature, is the same as in a transition-edge sensor (TES), and functionally identical to a supercondicting nanowire single-photon detector (SNSPD). [SNSPD does not have a separate heater.] These cryogenic detectors are typically used when a good electrical amplifier does not exist for the input radiation. A textbook example would be optical or x-ray photons.
I believe the device will be terribly inefficient as a transistor. The root cause is that the electrical signal gets converted into heat and back: Heater current -> Electron heating -> Phonon (lattice) heating -> Breaking of Cooper pairs in the channel -> Suppression of (super)current. Once the heat is in the phonons (lattice vibrations), it can propagate anywhere in the chip substrate.
Also, the active area needs to be continuously heated to maintain the resistive ("off") state.
Since HN is mostly a computing-oriented forum: This transistor will not be used for general-purpose logic cricuits.
1) To anyone who's studied algebra, it is clear that identities of the form LHS = RHS can be obtained by a nested application of transformations and substitutions in a consistent manner.
2) Of course, arriving at a new, insightful result often involves taking mundane steps. However, in this case, the new mathematical discoveries based on the output tableaus of your algorithm are hypothetical. Whereas the manuscript (and the authors) have already pocketed one of the premium accolades in sciences in the form of a Nature publication.
3) To drive the point above home, do you think the resulting mathematical insights themselves, without riding on the "AI" novelty aspect, would clear the bar for a Nature (or similar high-impact) publication? To be clear, I'm not a mathematican, but I believe the answer would be no. Contrast this with another AI/ML advance published in Nature quite recently: AlphaGo. Note how the gist of their paper, superhuman performance in Go, is a self-standing achievement that merely makes use of machine learning techniques.
I think my take-home from this discussion is that the honor code can be made to work in the right circumstances that exist, at least, at Caltech, Rice, etc.
At the same time, I believe it is impossible to induce the requisite "cohesion" in other contexts such as 100% remote learning or high-stakes mass testing (entrance exams etc.), even if the student body stayed the same.
However, I'd assume that doing either during a test would be against the honor code.
Clearly, any society collectively benefits from honesty, whereas an honest act is, at least in a strict game-theoretic view, a loss to the individual in the short term.
The question "Are you honest?" does not necessarily filter out dishonest people. (Incidentally, I always felt like I would not have been admitted to Caltech as a student.)
I'd still point out that the system is not robust against: - Working in groups. (Bad if tests are supposed to assess individual performance.) - Asking outsiders for help.
Also, your downplaying of "rote" learning feels misguided, no matter how advanced/abstract/high-level the domain in question is. Cue the Bruce Lee quote about 10,000 kicks..
The shortcomings that seem obvious to me are: - Penalizes honesty - Implied notion that Caltech students are "honourable" and honest. How is this achieved in practice?
For algorithms like HHL that have superclassical performance, a complex superposition encoding the data needs to be created first. This state is subsequently "consumed" by the algorithm. The no-cloning theorem forbids creating copies of the encoded state, and hence the encoding step needs to be repeated every time the algorithm is run.
For another example, consider Grover's search that is sub-linear in calls to an oracle function. If the oracle references a linear array of data, for example, it needs to work on superpositions of array indices. In other words, the entire dataset needs to fit in "quantum" memory.
Using a quantum cpu can only be sensible for computationally difficult problems where the hard problem instances can be specified by a relatively small number of bits.
What separates a coherent "quantum" superposition, say, |0> + |1>, from a probabilistic "non-quantum" 50:50 mixture is that I can choose a measurement basis in which the coherent state always yields a definite result, say "1", whereas measuring the mixed state always yields a 50:50 mixture of "0"s and "1"s.
A continuous sweep of the angle of the measurement basis generally results in an interference pattern, the amplitude of which can be used to assess the fidelity of the quantum state.
(I get paid to work on quantum communication and related experiments.)
I used to be puzzled by how anyone could ever drive off with the nozzle still in the car. I had only seen the non-latching dispenser, and service stations don't exist in my country. I was probably (shamefully) thinking something along the lines of "dumb Americans."
Once in the US, I realized it's a mixture of full service and self-service stations with latching or non-latching dispensers, depending on the region. And I do need to make an extra glance at the gas tank door before driving off.
Problematically, to secure funding today, one is essentially expected to frame every condensed-matter experiment as the next transistor. Not only in grant applications, but increasingly also in the abstract and opening paragraphs of research articles. There's a marked contrast with older research articles in physics, which usually go straight to disseminating the results. (Needless to say that I prefer the old style.)
As a result, a great deal of funding and attention is allocated towards projects that simultaneously 1) Will not improve the quality of life of anyone, even in the long term. 2) Are "de-risked" to such extent that no new scientific insights can come out of them.
Foot note: I acknowledge Quantum Computing is just a subfield of QIS, but the abstract only talks about IBM Q and quantum algorithms.
Q: Can I make use of the 60-day grace period to stay in US after I resign? I don't need work authorization for this period.
Google (or anyone else) hasn't shown an implementation of an error correcting code, so we do not have data points for a model-free "ruler extrapolation" of logical error rate vs. lattice size.
In fact, I think the Sycamore qubits were "pre-threshold", i.e. no error correction gain possible even in theory. I wonder if someone will correct/confirm me. I remember the readout fidelity was particularly poor.
Furthermore, I would argue that the large readout errors make the observed scaling of total error slightly less impactful.
But don't get me wrong, it's still a monumental achievement.
As an experimentalist, I would agree with the sentiment that the potential future applications of Quantum computing probably receive too much media attention, given the maturity level of existing technology.
The technical arguments as to why building a useful QC will be impossible are a bit more shaky.
First, the post seems to imply that the calibration effort scales with the number of gates in the algorithm. This is false. In reality, the number of interactions that need individual calibration is basically the number of qubit-qubit pairs exposed by the gate library. The most mainstream approach to error-corrected QC only uses nearest-neighbor interactions, and hence the scaling is linear.
Second, it is not clear to me why the number of computational basis states (2^N) is relevant to the engineering at all. Following this line of thinking, the recent Quantum supremacy result by Google already amounted to a mastery of 2^53 ~ 10^16 degrees of freedom.
Third, an argument is made that large-scale error correction will not work because of correlated errors, of unspecified nature. I think a claim like this should come with a mention of at least one concrete source of such errors, so that we could estimate its magnitude and potential severity. Note that such calculations are almost the essence of day-to-day work of a physicist.
In the absence of more detail, I can make a generic counterargument: Any phenomenon causing such correlated errors by definition affects multiple physical qubits at once, and will tend to be more macroscopic in nature. This is in contrast with the processes that limit the fidelity of one- and two-qubit operations, which is what the error-correcting code will take care of. Macroscopic disturbances are exactly the ones that we can attempt to shield against with clever engineering.
Quantum computing (the integer-factoring kind) is the focus of only the UC Berkeley-led consortium. This effort accounts for 1/3 of the announced funding.
The other two centers will work on Quantum networking (UIUC) and Quantum sensing (U Colorado).
Original NSF announcement:
https://www.nsf.gov/news/special_reports/announcements/07212...
In brief, the irreversible heat dissipation, calculated from the change in the entropy (S) of some sub-system of the quantum computer, does not yield a quantity that has much practical relevance. But it does lead to interesting considerations, some of which I've written out below.
Following the basic thinking behind Landauer's principle, if one considers the state of the qubits themselves, the dissipation is always zero or even negative. In the usual universal gate-based model, any computation starts with a known pure quantum state for which S = 0. The gates are unitary transformations and do not change the entropy. Non-coherent interactions generally lead to a state with nonzero positive entropy. Amusingly, this could be interpeted as using the quantum register to cool its surroundings by a tiny amount, at the cost of randomizing the output of the "computation".
A more fruitful approach (which you also alluded to) relates to the "cost" of the unitary transformations themselves. Implementing high-fielity gates, necessary for useful quantum computing, requires very precise time-varying external control fields. It is correct to think that there is a thermodynamic cost to ensuring that the noise in the fields experienced by the qubits are small.
For a concrete example, consider the fact that superconducting qubits are controlled with microwave signals. The output of a room-temperature microwave source has thermal noise supreimposed with the desired, sythesized signal. For simplicity, we can take the noise temperature to be 300K. (It is much higher in practice.) To use this output to drive a superconducting qubit at T = 30 mK, the power needs to be attenuated by at least a factor of 10^4. (Again, the real factor is much higher.) Hence, a lot of power is seemingly wasted in the synthesis of the control signals. I think a similar argument can be made about the lasers used to control some other kinds of qubits.
This type of dissipation is relevant for hypothesized large-scale quantum computers. It doesn't lead to new deep "quantum" insights, however. First, the dissipation is independent of the state of the qubits, and simply scales linearly with the number of operations. Second, the dissipation takes place outside of the delicate qubit system. Removing any amount of entropy from (i.e., cooling) the environment surrounding the qubits is merely a difficult classical engineering problem.
The thing is, the claims made here, taken at face value, are rather extraordinary!
Under usual conditions, in order to observe quantum phenomena with microwaves, sub-Kelvin temperatures are required to prevent the quantum signal from being swamped by blackbody radiation. Theoretically, there could be a few ways around this limitation (e.g. N00N states), but all prior experiments (to my knowledge) have been performed firmly within the confines of cryogenic refigerators.
In this experiment, too, the non-classical state begins at a brisk temperature of 7 mK. To get a strong enough signal that can be brought out to room temperature, bounced off a target, and finally digitized by room-temperature electronics, the experiment includes a "classical" transistor amplifier in the signal path.
Now, another elementary result of quantum theory forbids amplifiers of this type from boosting quantum signals without significant added noise.
How did the experimentalists get around this limitation? It appears that they have chosen to present some of their results as a function of the number of signal photons, leaving out the added noise by the amplifier.
The problematic part for the purpoted quantum advantage is that the classical noise at the amplifier output is much larger than the quantum component. Hence, a target illuminated by this kind of "radar" will simply pick up the amplifier noise, and no "true" quantum advantage was demonstrated.
Note that the above conclusions can be found in the research article itself, starting at the end of second-to-last paragraph on page 3. Also note that, carefully reading the abstract and introduction, no claims of demonstrated beyond-classical performance are made.
Take-home lesson? At least when it comes to quantum physics, it's really really hard to accurately evaluate the significance of cutting-edge research. I'd be willing to assume the same holds for other branches of science, as well.
A great deal of the recent progress has been on honing, parallelizing, and automating the same processes one used to do by hand/at small scale in 2013. You'll probably instantly recognize most of the things going on in the lab, and could become productive on a short schedule.
However, I'd be wary of the medium-term prospects of the field: The discord between the perceived and actual capabilities of the hardware remains as big as ever. You could consider re-entering the field if you're happy contributing to incremental technological advances relevant to other researchers. Progress with meaningful societal or economic impact is probably much further off.