PSA: quantum computers are waaay less powerful than the public assumes. Experts know this, but are incentivized not to say it directly. Quantum computers are less like rocket ships and more like teleportation beams that, due to unavoidable constraints in the laws of physics, will only ever be able to transport neutrinos, gerbil skeletons, and small quantities of blue lint.
Quantum computers do not "try all paths at once." If they did, they would be *ungodly powerful* (solve all problems in NP). Quantum computers are not ungodly powerful. Their capabilities are extremely spikey. Four spikes, basically*. If you interpolate between the capability spikes, quantum computers sound like rocket ships. If look at the spikes the spikes directly, you'll see they are miraculous solutions to problems almost nobody has.
The one big exception is for NSA. Quantum computers do legitimately have the potential to give a mind-blowing speedup for breaking certain kinds of encryption. But that's a very, *very* specific usecase, and very unlike of what what quantum computers do in general. Most hard problems computers deal with in practice are "NP-Complete", and we are almost sure that quantum computers can't solve those kinds of problems quickly.
The second more speculative exception, which the interviewee highlights to his credit, is chemistry and physics R&D. Quantum computers could, plausibly, speed up simulations there. But how much of physics and chemistry R&D is bottlenecked on accurate computer simulations? It's pretty speculative. AI algorithms like AlphaFold are making huge advances there but it's not at all clear what an unlock that will be.
*four spikes are: Shor's algorithm (breaks asymmetric encryption, legitimately a big deal for the NSA), Grover's algorithm (general-purpose n^2 speedup for "try all paths at once" [NP] sort of problems, which sounds great but in practice likely won't be practical), quantum annealing (imo basically useless) and simulating other quantum systems like small chemicals (which is cool BUT I my guess is chemical simulations are not much of an economic bottleneck [and to the extent chemical simulations are economically valuable, most of those problems could be solved by modern AI algorithms in the vein of AlphaFold]).
(I'm in Taipei if you want to ask someone to ask questions about this over tea, though LLMs are much better than me at this point)
Worth naming the spikes: Shor (factoring), Grover (quadratic search, so less dramatic than advertised), Hamiltonian simulation (chemistry/materials), and HHL for linear systems. HHL's preconditions are so restrictive it almost never beats classical in practice. Grover's quadratic speedup means doubling qubits barely moves the needle on most search problems. Popular coverage keeps interpolating between spikes instead of looking at them.
Agreed. (Though a bit less certain about whether the four spikes will remain the only ones).
I was pretty disappointed that this is still the talking point. If you listen closely to the audio, you can hear Scott Aaronson wailing in the distance.
Zach is clearly a quantum optimist. Building a quantum computer requires solving the decoherence problem, which is so difficult that it seems extremely unlikely in the foreseeable future.
From the perspective of the US government, it still makes sense to invest since even a small chance of success could be transformative. But it is frustrating to think that this funding will most likely not pay off.
Every time I try to understand quantum mechanics, and feel equally dumb each time, I'm comforted by one of Feynman's other famous quotes: "I think I can safely say that nobody understands quantum mechanics".
A 72-100 hour refresh works for AGI discourse, but worth flagging where it intersects quantum: NIST finalized the post-quantum crypto standards (FIPS 203/204/205) in August 2024. Harvest-now-decrypt-later means the migration deadline isn't when quantum arrives, it's now. Most syntheses still track them as separate stories.
PSA: quantum computers are waaay less powerful than the public assumes. Experts know this, but are incentivized not to say it directly. Quantum computers are less like rocket ships and more like teleportation beams that, due to unavoidable constraints in the laws of physics, will only ever be able to transport neutrinos, gerbil skeletons, and small quantities of blue lint.
Quantum computers do not "try all paths at once." If they did, they would be *ungodly powerful* (solve all problems in NP). Quantum computers are not ungodly powerful. Their capabilities are extremely spikey. Four spikes, basically*. If you interpolate between the capability spikes, quantum computers sound like rocket ships. If look at the spikes the spikes directly, you'll see they are miraculous solutions to problems almost nobody has.
The one big exception is for NSA. Quantum computers do legitimately have the potential to give a mind-blowing speedup for breaking certain kinds of encryption. But that's a very, *very* specific usecase, and very unlike of what what quantum computers do in general. Most hard problems computers deal with in practice are "NP-Complete", and we are almost sure that quantum computers can't solve those kinds of problems quickly.
The second more speculative exception, which the interviewee highlights to his credit, is chemistry and physics R&D. Quantum computers could, plausibly, speed up simulations there. But how much of physics and chemistry R&D is bottlenecked on accurate computer simulations? It's pretty speculative. AI algorithms like AlphaFold are making huge advances there but it's not at all clear what an unlock that will be.
*four spikes are: Shor's algorithm (breaks asymmetric encryption, legitimately a big deal for the NSA), Grover's algorithm (general-purpose n^2 speedup for "try all paths at once" [NP] sort of problems, which sounds great but in practice likely won't be practical), quantum annealing (imo basically useless) and simulating other quantum systems like small chemicals (which is cool BUT I my guess is chemical simulations are not much of an economic bottleneck [and to the extent chemical simulations are economically valuable, most of those problems could be solved by modern AI algorithms in the vein of AlphaFold]).
(I'm in Taipei if you want to ask someone to ask questions about this over tea, though LLMs are much better than me at this point)
Worth naming the spikes: Shor (factoring), Grover (quadratic search, so less dramatic than advertised), Hamiltonian simulation (chemistry/materials), and HHL for linear systems. HHL's preconditions are so restrictive it almost never beats classical in practice. Grover's quadratic speedup means doubling qubits barely moves the needle on most search problems. Popular coverage keeps interpolating between spikes instead of looking at them.
Agreed. (Though a bit less certain about whether the four spikes will remain the only ones).
I was pretty disappointed that this is still the talking point. If you listen closely to the audio, you can hear Scott Aaronson wailing in the distance.
Zach is clearly a quantum optimist. Building a quantum computer requires solving the decoherence problem, which is so difficult that it seems extremely unlikely in the foreseeable future.
From the perspective of the US government, it still makes sense to invest since even a small chance of success could be transformative. But it is frustrating to think that this funding will most likely not pay off.
Every time I try to understand quantum mechanics, and feel equally dumb each time, I'm comforted by one of Feynman's other famous quotes: "I think I can safely say that nobody understands quantum mechanics".
A 72-100 hour refresh works for AGI discourse, but worth flagging where it intersects quantum: NIST finalized the post-quantum crypto standards (FIPS 203/204/205) in August 2024. Harvest-now-decrypt-later means the migration deadline isn't when quantum arrives, it's now. Most syntheses still track them as separate stories.