4 ms·
Data security okay. But AI? How will that change?
by isoprophlex 2y ago
Data security okay. But AI? How will that change?
- adastra22 2y agoAI is essentially search. Quantum computers are really good at search.
- sheeshkebab 2y agoSearch of what?
- roywiggins 2y agoLatent space? https://en.wikipedia.org/wiki/Latent_space https://en.wikipedia.org/wiki/Latent_space
- adastra22 2y agoAnything. Everything. In domains where the search space is small enough to physically enumerate and store or evaluate every option, search is commonly understood as a process solved by simple algorithms. In domains where the search space is too large to physically realize or index, search becomes "intelligence." E.g. winning at Chess or Go (traditional AI domains) is searching through the space of possible game states to find a most-likely-to-win path. E.g. an LLM chat application is searching through possible responses to find one which best correlates with expected answer to the prompt. With Grover's algorithm, quantum computers let you find an answer in any disordered search space with O(sqrt(N)) operations instead of O(N). That's potentially applicable to many AI domains. But if you're so narrow minded as to only consider connectionist / neural network algorithms as "AI", then you may be interested to know that quantum linear algebra is a thing too: https://en.wikipedia.org/wiki/HHL_algorithm https://en.wikipedia.org/wiki/HHL_algorithm
- cwillu 2y agoO(sqrt(N)) is easily dominated by the relative ease of constructing much bigger classical computers though.
- adastra22 2y agoUh, no? Not for large N. There are about 2^152 possible legal chess states. You cannot build a classical computer large enough to compute that many states. Cryptography is generally considered secure when it involves a search space of only 2^100 states. But you could build a computer to search though sqrt(2^152) = 2^76 states. I mean it'd be big--that's on the order of total global storage capacity. But not "bigger than the universe" big.
- IanCal 2y agoDoing 2^76 iterations is huge. That's a trillion operations a second for two and a half thousand years if I've not slipped up and missed a power of ten.
- hatthew 2y agoMaybe 100 years from now we can do 2^18 quantum ops/sec and solve chess in a day, whereas a classical computer could do 2^36 ops/sec and still take longer than the lifetime of the universe to complete.
- adastra22 2y agoGoogle's SHA-1 collision took 2^63.1 hash operations to find. Given that a single hash operation takes more than 1000 cycles, that's only less than three doublings away. Cryptographers worry about big numbers. 2^80 is not considered secure.
- IanCal 2y agoIt's early so I'm thinking out loud here but I don't think the algorithm scales like this, does it? We're talking about something that can search a list of size N in sqrt(N) iterations. Splitting the problem in two doesn't halve the compute required for each half. If you had to search 100 items on one machine it's taken 10x iterations but split over two it'd take ~7x on each or ~14 in total.
- CJefferson 2y agoGrover's algorithm is useful for very few things in practice, because for most problems we have a better technique than checking sqrt(N) of all possible solutions, at least heuristicly. There is, at present, no quantum algorithm which looks like it would beat the state of the art on Chess, Go, or NP-complete problems in general.
- spencerchubb 2y agotraining an ai model is essentially searching for parameters that can make a function really accurate at making predictions. in the case of LLMs, they predict text.
- cwillu 2y agoUnstructured search is only a √n improvement. You need to find some way for algorithmically significant interference/cancellation of terms in order for a qc to potentially (!) have any benefit.
- numpad0 2y agoAren't quantum computers expected to be like digitally read analog computers for high dimension optimization problems, and AI is like massive high dimension optimization problems?
- Yoric 2y agoGoogle is betting on digital quantum computers. There are, however, analog quantum computers, e.g. by Pasqal, which hope to capitalize on this to optimize AI-like high dimension optimization problems.
- gecko22 2y agoWhy do quantum computers need to be analog to be applied to such problems?
- arlort 2y agoThey don't, but if I interpreted the post above correctly analog QCs aremuch easier to build At least as far as I'm aware by digital they probably mean a generally programmable QC, whereas another approach is to encode a specific class of problems in the physical structure of an analog QC so that it solves those problems much faster than classical. This latter approach is less general (so for instance you won't use it to factor primes) but much more attainable. I think D-wave or someone like that already had commercial application for optimization problems (either traveling salesman or something to do with port organization)
- numpad0 2y agoBut aren't qubits sort of analog? I thought the bits have figurative angles of rotation or something continuous that would resonate together and/or cancel out like neurons in fully-connected parts of neural nets that would be read out at the end of the pipeline.