1,891 karma · joined September 14, 2019
Edit: back from watching the video, and a few observations:
- Swift does not suffer from Rust's refusal to add things into the standard library. gasp regex support in stdlib? Async/await runtimes?
- Default testing framework is a godsend, especially when integrated into the build system
- Rust but better for everything that's not low-level systems development is the impression I get
Think of it this way: in 5-10 years, ObjC will be the python2 of Apple development.
My opinion is the opposite: I think the old paradigm of distros managing a giant set of system libraries is a bad one, and is how we ended up in the land of docker. Go and Rust made the right decisions here: vendor all the dependencies, and distros can't mess with them. Makes it easier for upstream and the distro.
Modern languages are not like C/C++: a single non-trivial rust program can easily depend on 100+ crates, and re-creating crates.io in your package manager is just a bad idea, even putting aside that there's probably major version incompatibilities the moment you go beyond a handful of programs. Look at the disaster that's python package management, that's not where you want to end up.
Then people figured out that since the government went through the trouble of assigning every* [1] american* [2] a unique* [3] number* [4], they might as well use that as their own ID too for inter-operability between different financial firms, and that's how we ended up in this shitty situation.
Every time somebody proposes 'hey we should come up with a better national ID number' you get people yelling about mumble mumble big government mumble mumble privacy mumble mumble feds tracking you around, and the proposal dies in committee.
[1] originally only issued to those who worked, and was not exempted from the SS system, notably children. Of course these days children have SS numbers because they're a de-facto ID number, and it's required to claim dependents, and with varying degrees of resistance from weird groups like the Amish. Additionally, non-eligible people are issued tax IDs instead, which have the same format as SSNs because it was easier that way
[2] big(?) political fights over who exactly must get numbers, of course, from religious groups, illegal immigrants, and many more
[3] except of course it's not unique, there were errors where people got duplicate numbers, and people can get reassigned numbers, thereby having more than one
[4] did you know they used to assign SSNs sequentially? The guy who was next in line got your number + 1
Interestingly, Fidelity lets you put your account into "lockdown" mode there they'll reject all ACATS (and ACH too I believe) transfers automatically without your intervention. I've always wondered what kind of legal magic they're pulling to make this compliant with FINRA rules.
What eventually convinced me over that it's not a problem is, as Patrick says, that these institutions will with high likelihood make you whole for fraudulent transfers, the same way they will for fraudulent ACH. The financial system generally works.
My observation is, and this may be more philosophical than technical: this process of "decomposing" middle-layer activations with a sparse autoencoder -- is it capturing accurately underlying features in the latent space of the network, or are we drawing order from chaos, imposing monosemanticity where there aren't any? Or to put it another way, were the features always there, learnt by training, or are we doing post-hoc rationalisations -- where the features exist because that's how we defined the autoencoders' dictionaries, and we learn only what we wanted to learn? Are the alien minds of LLMs truly also operating on a similar semantic space as ours, or are we reading tea leaves and seeing what we want to see?
Maybe this distinction doesn't even make sense to begin with; concepts are made by man, if clamping one of these features modifies outputs in a way that is understandable to humans, it doesn't matter if it's capturing some kind of underlying cluster in the latent space of the model. But I do think it's an interesting idea to ponder.
My view is that, fundamentally, a lot of "professional" workflows is just as much about shuffling data around as it is working on said data. Just as data science taught us that data cleaning is 80% of the job, the same applies to video editing, programming, etc. iPadOS is uniquely bad at moving data around due to its onerous requirements around sandboxing and its poor interfaces around inter-app communication, when these are precisely the things that are most needed.
And then there's the multitasking, which remains terrible to this day. Can they improve it? Definitely, but I wouldn't want to be the guy trying to figure out how to do it without ruining the normal user experience.
This only works for small projects. Once your project grows beyond a handful of files and dependencies this breaks down real fast.
I've been on this train more than once: most non-trivial projects reach a point where you have to build for multiple targets, support multiple configurations, test in all those weird configurations, manage CI, manage external dependencies, build said dependencies, etc. A makefile that supports all of the above becomes a monstrosity long before you get there and at some point you give up and pull in a real build system to take care of it (they're all bad in their own way, pick your poison really with cmake/meson/bazel/<new shiny thing>, just less bad than make).
Sure you can manage a giant pile of target-specific variables, platform-dependent scripts, and what not in make, but really it's much less painful to just bite the bullet and migrate to a build system that thought of these problems before. Things like "this dependency is mandatory is this configuration, optional in these other configurations, and is liked differently depending on whether we're building static or shared objects and whether we're using the platform library or falling back to a vendored version" is basically impossible in my experience.
Makefiles are great when your project is a dozen `.c` files. The moment you grow beyond that it's time to move on and use something else.
If anybody has an explanation for this I'd love to hear it.
That’s … not true? Most Chinese software and websites are utf8 by default, and it’s been that way for a while. GBK and her sisters might linger around in legacy formats, but UTF-8 has certainly reached the point of mass adoption from what I can tell.
What is, in your experience, the minimum that gets you a decent shot at a NIW/O1? You previously mentioned that published papers are helpful but not mandatory, but for PhD students/graduates, if publications are not the determining factor, what is?
Looking through the Dhanasar's three prongs, it does seem like papers are the easiest and most consistent way of fulfilling the second prong (that "The Person is Well Positioned to Advance the Proposed Endeavor").
The third prong is the most mysterious to me (that "On balance, it would be beneficial to the United States to waive the job offer and thus the permanent labor certification requirements"), how do people typically meet the requirements for this prong?
Also quite funny that black friday was translated as a sale on black things in at least one instance. When I grew up a decades ago in east/SE Asia, black friday was not really a thing, but I guess times change.
In this case, and my lack of understanding in both bio and stats is showing here, we're trying to develop a test for breast cancer. An ideal test will have high sensitivity and specificity, and a tight confidence interval for both of those numbers. This way we can be confident that a positive/negative test actually moves our priors meaningfully in either direction.
I guess my question/comment is more on the fact that I don't see how any of the results shown actually translate, as the headline suggests, into a cancer test of high power. The priors for any kind of cancer is pretty low from what I can find, so we need high power tests in both the positive and negative direction to meaningfully effect health outcomes. I can't find any CI numbers in the paper, which may just be me not reading it closely enough, but it doesn't help my confidence.
Actual publication: https://pubs.aip.org/avs/jvb/article/42/2/023202/3262988/Hig...
Experiment size is literally N=21, with 4 healthy participants, 3 in-situ breast cancers, and 14 invasive breast cancers.
N=21 might as well be useless in my opinion. You can't draw any meaningful conclusions about statistical power of this test; if your priors were 10% for breast cancer, after taking this test, your posterior probably remains unchanged.
As with all technologies though, the cost first starts out high as it's introduced, is reduced as it becomes popular, and climbs back up as it's obsoleted and nobody makes them anymore but some people are stuck using them.