Note that this is about metadata for package requests, so you're downloading something from a server already. The question is what information is in that request.
The plan is to make aggregate usage data available publicly and potentially share more detailed usage data with individual package authors. The exact format is TBD since it'll depend on the quality of the data that we get (this is not active yet, except on the preview build). The raw logs will be accessible to core developers with a reasonable need to access (e.g. they're working on the infrastructure or running the analytics), but will not be public.
While this is true of the feature mentioned, do note that packages are currently hosted on various third party hosting services that can and do track substantially similar information. In 1.5, we're moving to our own infrastructure for serving packages (which should give better performance and allow things like incremental updates). This thread is about what information gets sent along with those requests.
Hi HN, please note that this is an active discussion thread in the Julia community. You are all more than welcome to chime in, but we do try to keep discussions as productive as possible, so if you do decide to comment, I'd ask that
1) You familiarize yourself with the actual proposal and the improvements that are currently underway and
2) Be kind
A number of people have put in an enormous amount of effort to try and get this right - please remember that they are indeed people.
Honestly nothing in this article is even that bad, just some elementary group theory. If you wanted to, you could probably pick it up in a few hours. The math for the HE schemes themselves does rely on some intermediate level number theory though (e.g. Chinese Remainder Theorem in polynomial rings) and of course there's a whole formal language for the security proofs also.
It represents the same plaintext, but the ciphertexts are different and both computationally indistinguishable from random noise (unless you know the key)
If you want to get away from analysis, I've found that cryptography can be quite an engaging subject. If you have the right book, it can have the rigors of more mathematical subjects, while being accessible without extensive background and having visible real-world applications. I unfortunately don't have much experience with books in this area, but I do like https://files.boazbarak.org/crypto/lnotes_book.pdf (plus it's free ;) ).
[EDIT: Previously I recommended Calculus on Manifolds here also, but on further reflection and reading some of the other responses I think I both misremembered the difficulty level of the book and overestimated what early-undergrad level means]
Looks like the devkits are $500 and need to be sent back at the end of the year. Also curious that the name ARM wasn't mentioned even once. I wonder if that's just marketing or whether they're plotting to do their own ISA at some point in the future also.
Yeah, same confusion here. First example, being HN added to the confusion, since HN seems like exactly the kind of site that would render fine on a thirty year old browser (not meant as a swipe in any way - HN does one thing and it does it well).
It can be if you want it to. At Harvard, there's very few requirements on what courses constitute a degree, so if you want to get a degree by just taking the easiest classes possible, you can definitely do that. Of course, most students don't do that since that's not why they went there. I do think professors try to give higher grade averages for the harder classes to avoid punishing students for challenging themselves, but I don't really have a problem with that (if you chose those classes you're probably not in it for the grades anyway). Also, at least I'm the sciences, a good fraction of students start taking classes at the graduate level either freshman or sophomore year, since they'll have completed a good chunk of the undergrad curriculum in high school.
One aspect that's underappreciated is that composability can actually lead to speed. If it's super easy to compose your algorithmic package with speed-improvement packages (parallelism, hardware accelerators, custom compilers), you're much more likely to be actually able to use them.
You are correct. Our technical documentation is mostly aimed at working scientists who want to start using these techniques in their work. That does sometimes lead to funny cases where a document assumes you know what a smooth manifold is but will explain try/catch blocks. We've started trying to put together more introductory-focused material at https://juliaacademy.com/. We don't currently have anything particularly AD focused (outside of the general ML courses), but I think that's a topic that's high on the interest list.
We have so many AD implementations in Julia now that we actually have infrastructure that separates out the definition of primitive derivate rules from the actual AD mechanism itself, so the rules can be shared amongst all of them. Of course it would be better if there was just one AD to rule them all, but there are tradeoffs in the design space that make that hard. I think having all these different implementations has actually helped crystalize what the design space actually is, what choices need to be made and what the interesting classes of applications are. I think that'll help with the next generation of these tools (disclaimer I just started working on one of those next generation tools yesterday ;) ).
Yes, you can just consider it a piecewise differentiable function. That's not usually a problem in practice if you think of the derivatives more as heuristics in a search problem than absolute truth. Obviously how well your optimization works does depend on the differentiability attributes of your underlying function and how well your optimization scheme can handle it.
The OS can just turn off any instruction extensions by setting the XCR0 register to whatever it wants. Unsupported instructions will then cause general protection faults (SIGSEGV on Linux). Whether that's what the OS wants to do is a totally separate question. It would be kinda funny if this thing lead to the ability of a process to request a certain XCR0 that the kernel would then serve, because I've been asking for that feature for unrelated reasons for quite some time and didn't get the warmest response to it on the mailing lists.
I'm of two minds about this. On the one hand JuliaCon is always quite an amazing event and it's the only place I get to see a number of wonderful people that I usually only interact with online. On the other hand I'm told there's been a few thousand registrations already, so maybe this will make JuliaCon more accessible to a much wider audience, which has always been a goal. We've had live streams and recordings in past years, but I'll be interested to see how things are different with an entirely virtual event.
Are board appointments considered hiring? Honest question. I have no idea how the law works here, but I just figured the shareholders could choose whoever they want.
Several terminals (kitty, iTerm2) do support plotting right into the terminal, but this is still sometimes useful if you don't want to load a while graphics stack just to do a simple plot.
Still super expensive though! I guess at least it's available again. I remember doing some liquid helium experiments in class in college just as the helium crunch started. The department didn't seem very happy suddenly having to spend hundreds of dollars extra just to let some undergrads play with it. I wonder if they discontinued that particular experiment.
I assume you're talking about the quantum resistance of the underlying schemes in use? My understanding is that all the current schemes of practical interest rely on the hardness of (variants of) LWE (https://en.wikipedia.org/wiki/Learning_with_errors), which is believed to be hard even if you have a quantum computer.
It's not just about outsourcing compute. There's lots of interesting things you can build on top of an FHE primitive. The linked blog post has the private information retrieval example (the database owner doesn't want to send you the whole database, you don't want the database owner to know what you're looking for). I have a worked example for machine learning here: https://juliacomputing.com/blog/2019/11/22/encrypted-machine... (The model owner doesn't want to share the model, you don't want to share the image you're asking them to process).
Last I looked, HElib was one of the slower implementations. That said, even in the best case, HE libraries are so slow, that I don't think "overhead" really captures the paradigm correctly, because the performance is highly dependent on the application. The best case "overhead" (doing some operation that the HE scheme natively supports, e.g. arithmetic over a very specific finite field) is probably 5-6 orders of magnitude.
Try if the generic binaries just work for you. A lot of work has gone into making sure things just work even without root privileges, so you shouldn't have to wait for IT to get it set up for you.
I grew up in Germany, but have been in the US for close to a decade now and all my technical education has been here. I have tried sending technical emails in German, but it's usually an absolute disaster. I have no Frome of reference for how to actually express things correctly. I also find my non-technical written German to have suffered quite a bit. Spoken word is fine, but writing is a bit tough. I guess I just didn't do any of that for a good ten years, so it got a bit rusty.
Well aware, but I was hoping to say to people "Just use 5.7 and it'll be fine". At the moment I have to say "Don't use anything prior to 5.7, unless you have a recent stable backpoint. In either case, some things are fine others aren't - I'll hope to have it fixed by 5.8". Which is fine, but more complicated ;).
Oh, already? I got a patch in to fix PTRACE_SYSEMU on aarch64, but I still have a few patches queued up to fix a few more issues there. I had hoped to get them all into the same release so there is only broken and not broken releases, rather than broken, semi-broken and non-broken releases. Oh well, ptrace on aarch64 has a few other big issues anyway that wouldn't have been mergeable during the RC window, so I suppose I should just try to get that in for 5.8.