Hacker News new | past | comments | ask | show | jobs | submit | winwang's comments login

glad to see another Scala fan :)

keep in mind that "for loops" are really "for comprehensions" and desugae into flatMap/map


Do you have strong evidence for this?


Dogs are highly intelligent, and it makes no sense to say that they get their intelligence by calculating the probabilities between consecutive woofs.


Would you say with equal confidence that they don't exemplify their intelligence by their ability to repeatedly select an often-successful next action from a set of possible next actions, based on a set of input observations?

"Tokens" don't have to be words, or woofs...


It still doesn’t make sense for dogs. It might make some sense given a higher-level goal (hiding a toy under the bed)[1] but it doesn’t make much sense for selecting the goals (“I should hide this toy because the other dog keeps stealing it”). In building an AI dog it doesn’t work to elevate these higher-level goals into individual tokens because real dogs form goals dynamically according to their environment and the set is infinitely large. (Note that LLM agents also badly struggle with this; generating goals token-by-token means their goals have hallucinations.)

[1] It still doesn’t make much sense to view this as a statistical process; dogs can generalize far better than transformers, as perhaps best seen with seeing-eye dogs. I believe dogs’ powers of causal reasoning exceed what is possible from mere surface statistics: e.g. they innately understand object permanence as puppies, whereas transformers still don’t understand it after viewing thousands of dogs’ lifetimes of experience.


I've not been able to find any way to distinguish "mere surface statistics" from the deeper, richer, and more meaningful kind of something that it is meant to be contrasted with, except that "surface statistics" are un-compressed. For example, surface statistics might be the set of output measurements generated by a compact process, such as the positions of planets over time; knowing the laws of gravity means we can generate gigabytes of these statistics correctly and easily, which will accurately match future observations.

But then going the other way, from statistics to a causal model, is just an inverse problem -- just like, say, going from a set of noisy magnetic field measurements at the boundary of a container to a pattern of electric current flow inside a volume, or going from planet positions to orbit shapes and periods to an inverse square law of gravity. Generating a compressed inverse model from surface statistics is exactly the sort of thing that deep learning has proven to be very good at. And by now we've seen no shortage of evidence that LLMs and other deep networks contain stateful world models, which is exactly what you'd expect, because for all their parameters, they aren't nearly big enough to contain an infinitesimal fraction of the statistics they were trained on.

So I think it's overly dismissive to regard LLMs as mere surface statistics.


> So I think it's overly dismissive to regard LLMs as mere surface statistics.

It's literally what they are though.

Yes those probabilities embed human knowledge but that doesn't mean that the LLM itself is intelligent. It's why every LLM today fails at anything that isn't centred around rote learning.


It's what they input and output, but it's not literally what they are. The only way to squeeze that many statistics into a compact model is to curve-fit an approximation of the generating process itself. While it fits stochastic sequences (of any type, but usually text), it's conceptually no different from any other ML model. It's no more surface statistics than a deep neural network trained for machine vision would be.


That only shows that word prediction isn't necessary, not that it's insufficient


makes a lot of sense.


For one, we could just start simulating quantum chemistry, though at that point it's more like "actually running quantum chemistry" rather than simulating.


Will AI take that use case though?


AI isn't magic. It's still subject to the limits of classical computing.


Note there's a caveat: problems in CS can be reduced to other problems in CS. If we solved SAT, well, no one cares about SAT, but traveling salesman obviously reduces to that.

(disclaimer: I don't think that's what is going on here, I'd have to dig into it more)


They didn’t solve any kind of CS problem. As far as I can tell the problem they solved is “what is this complicated quantum system going to do” by building the complicated quantum system and seeing what it did.

Then they claim it would take a gazillion years to simulate on a conventional computer. Which I’m sure is true.


> If we solved SAT, well, no one cares about SAT

Really? SAT is the question "I have a set of constraints. Is it possible to obey all of them at once?"

If it were impossible to use that question to answer any other questions, I'm pretty sure there would be a lot of interest anyway.

It's kind of like how a lot of people care about the determinant of a matrix, which is exactly the same question set against a much more restrictive set of possible constraints.


Edit after skimming arxiv preprint[1]:

Yeah, this is pretty huge. They achieved the result with surface codes, which are general ECCs. The repetition code was used to further probe quantum ECC floor. "Just POC" likely doesn't do it justice.

(Original comment):

Also quantum dabbler (coincidentally dabbled in bitflip quantum error correction research). Skimmed the post/research blog. I believe the key point is the scaling of error correction via repetition codes, would love someone else's viewpoint.

Slightly concerning quote[2]:

"""

By running experiments with repetition codes and ignoring other error types, we achieve lower encoded error rates while employing many of the same error correction principles as the surface code. The repetition code acts as an advance scout for checking whether error correction will work all the way down to the near-perfect encoded error rates we’ll ultimately need.

"""

I'm getting the feeling that this is more about proof-of-concept, rather than near-practicality, but this is certainly one fantastic POC if true.

[1]: https://arxiv.org/abs/2408.13687

[2]: https://research.google/blog/making-quantum-error-correction...

Relevant quote from preprint (end of section 1, sorry for copy-paste artifacts):

"""

In this work, we realize surface codes operating below threshold on two superconducting processors. Using a 72-qubit processor, we implement a distance-5 surface code operating with an integrated real-time decoder. In addition, using a 105-qubit processor with similar performance, we realize a distance-7 surface code. These processors demonstrate Λ > 2 up to distance-5 and distance7, respectively. Our distance-5 quantum memories are beyond break-even, with distance-7 preserving quantum information for more than twice as long as its best constituent physical qubit. To identify possible logical error f loors, we also implement high-distance repetition codes on the 72-qubit processor, with error rates that are dominated by correlated error events occurring once an hour. These errors, whose origins are not yet understood, set a current error floor of 10−10. Finally, we show that we can maintain below-threshold operation on the 72qubit processor even when decoding in real time, meeting the strict timing requirements imposed by the processor’s fast 1.1µs cycle duration.

"""


You got the main idea, it's a proof-of-concept: that a class of error-correcting code on real physical quantum chips obey the threshold theorem, as is expected based on theory and simulations.

However the main scaling of error correction is via surface codes, not repetition codes. It's an important point as surface codes correct all Pauli errors, not just either bit-flips or phase-flips.

They use repetition codes as a diagnostic method in this paper more than anything, it is not the main result.

In particular, I interpret the quote you used as: "We want to scale surface codes even more, and if we were able to do the same scaling with surface codes as we are able to do with repetition codes, then this is the behaviour we would expect."

Edit: Welp, saw your edit, you came to the same conclusion yourself in the time it took me to write my comment.


Haha, classic race condition, but I appreciate your take nonetheless!


Google could put themselves and everyone else out of business if the algorithms that underpin our ability to do e-commerce and financial transactions can be defeated.

Goodbye not just to Bitcoin, but also Visa, Stripe, Amazon shopping, ...


bitcoin proof of work is not as impacted by quantum computers - grover's algorithm provides a quadratic speedup for unstructured search - so SHA256 ends up with 128 bits of security for pre-image resistance. BTC can easily move to SHA512.

symmetric ciphers would have similar properties (AES, CHACHA20). Asymmetric encryption atm would use ECDH (which breaks) to generate a key for use with symmetric ciphers - Kyber provides a PQC KEM for this.

So, the situation isn't as bad. We're well positioned in cryptography to handle a PQC world.



It seems you can get TLS 1.3 (or atlest slighty modified 1.3) to be quantum secure, but it increases the handshake size by roughly 9x. Cloudflare unfortunately didn't mention much about the other downsides though.

https://blog.cloudflare.com/kemtls-post-quantum-tls-without-...


About one third of traffic with Cloudflare is already using post-quantum encryption. https://x.com/bwesterb/status/1866459174697050145

Signatures still have to be upgraded, but that's more difficult. We're working on it. http://blog.cloudflare.com/pq-2024/#migrating-the-internet-t...


Yes-ish. They're not enabled yet, but post-quantum signatures & KEMs are available in some experimental versions of TLS. None are yet standardized, but I'd expect a final version well before QCs can actually break practical signatures or key exchanges.


One third of all human traffic with Cloudflare is using a post-quantum KEM. I'd say that counts as enabled. We want that to be 100% of course. Chrome (and derivates) enabled PQ by default. https://radar.cloudflare.com/adoption-and-usage


It's currently believed that quantum computers cannot break all forms of public key cryptography. Lattice based cryptography is a proposed replacement to RSA that would let us keeping buying things online no problem.


Why is no one else talking about this? I came here to see a discussion about this and encryption.


Because this result is still very far from anything related to practical decryption.


And if they were, would they tell the world?


If they had a QC that could run Shor's algorithm to factor the number 1000, I'd guarantee you they'd tell the whole world. And it would still be a long, long time from there to having a QC that can factor 2048-bit numbers.


My friends have successfully relied on Medicaid during financial hardships + unemployment. At least in NYC, the Medicaid plans are quite decent.

Also, for those who require plans similar to the one previously provided, COBRA (18 months) is decent -- expensive but presumably less expensive than "equivalent" in the marketplace if we're talking about a good corporate plan.


18 months + 18 months in NY State for a total of 36

Just that COBRA costs quite a lot.



At the physical layer, this makes a ton of sense. Apologies for completely ignoring that part. I was specifically curious on encoding each "digit" of a ternary number independently in a computer. Which.... yeah, that isn't what this article was even saying.


It is also worth noting that Gigabit Ethernet uses PAM5.


I break up my messages, as do many people.


Do you do so with the expectation that they might arrive out of order, or one fails?


Out of order no, failing and having to manually re-send which makes it out of order is acceptable.


I find that unacceptable and frustrating, personally. If a message fails to send, I want the client to hold back any later messages until the failed message is resolved somehow. It should auto-retry and (hopefully) eventually succeed, or I can manually delete it and "release" the following messages.


I do so with the expectation that they should arrive in order if no one fails, but apparently it is "debatable" if this is a reasonable expectation


I actually thought you meant that this was about applying QoS algos to real life, before I saw the link lol.


Using it to run easily-scalable Spark clusters. Previously used it for large distributed builds. It's been pretty great (even if annoying at times).


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: