Hacker Newsnew | past | comments | ask | show | jobs | submit | conradev's commentslogin

They’re already optimizing GPU die area for LLM inference over other pursuits: the FP64 units in the latest Blackwell GPUs were greatly reduced and FP4 was added

For TPUs I believe it is Broadcom: https://www.theregister.com/2023/09/22/google_broadcom_tpus/

Not sure about the mobile SoCs


Broadcom fills the same role for Google TPU that Marvell fills for Trainium @ amazon.

I bought a sticker pack that does this. Still being able to see the light is pretty useful

Changing well-tested code is unsafe.


not changing working code to prevent issues is unsafe.

we can go in circles all day with blanket statements that are all true. but we have ample evidence that even if we think some real-world C code is safe, it is often not because humans are extremely bad at writing safe C.

sometimes it's worth preventing that more strongly, sometimes it's not, evidently they think that software that a truly gigantic amount of humans and machines use is an area where it's worth the cost.


believing that rewriting to rust will make code safe is unsafe) Of course it will be safer, but not safe. Safety is a marketing feature of rust and no more. But a lot of people really believe in it and will be zealously trying to prove that rust is safe.


If the code is brittle to change, it must not have been particularly safe in the first place, right?

And if it's well-tested, maybe that condition is achieved by the use of a test suite which could verify the changes are safe too?


A test will never catch every bug, otherwise it's a proof, and any change has the probability to introduce a new bug, irregardless of how careful you are. Thus, changing correct code will eventually result in incorrect code.


I'm not sure if that's how probability works.


I mean if you want Git to never change you're free to stick with the current version forever. I'm sure that will work well.


I obviously don’t think that is wise, but Git is literally designed with this in mind: https://git-scm.com/docs/repository-version/2.39.0

Just like SQLite has an explicit compatibility guarantee through 2050. You literally do not have to update if you do not want to.


And it’s still a choice you can make regardless of Git moving to Rust or not, so what’s the problem?


This is the repo format version.

It's pretty different from the git version, which receives new releases all the time for things like security patches, improvements, and new features.


When you say “running slowly”, do you mean:

- Apps launch slower (time from app launch to interactivity)

- Apps hang more frequently (interface freezes after interaction)

- Apps drop frames (scrolling and animations stutter)

I imagine it’s some combination of the above. Apps and the operating systems that run them get bigger and slower over time.

There really isn’t a “slowness constant” to tweak short of animation speed, and I believe Apple made app launch noticeably faster, animation-wise in iOS 26. Just lots and lots of regressions to fix.


The third: there’s stutter all over the place, jank in pretty much every animation. E.g. before leaving this comment I just swiped between two app screens and got a ~10 frame hiccough in the middle of the animation. The jank is pervasive.


A 10-frame hiccup is actually more of a hang (a single >50ms freeze) than consistent frame dropping (i.e. scrolling at 30fps)

but I agree, the hangs in the latest operating systems are brutal. Feels a little half-baked. macOS Tahoe is much worse than Sequoia on this by far. I’m hoping that the later releases (26.1, etc) make the situation better.


and I imagine that the initramfs is not encrypted and trivially modifiable?

Apple is able to achieve this securely because their devices are not fully encrypted. They can authenticate/sign the unencrypted system partition.



This is super cool, thanks for the link! I’m glad they were able to leverage the TPM


Networking always comes up after cold boot, but WiFi passwords are stored on the encrypted volume. So, it depends on whether you use WiFi!


One network will get stored on NVRAM, I think it’s probably whatever the first one you connect to is.


And also depends on your using DHCP?


Apple does a “userspace reboot” (killing all processes) after device unlock to categorically solve this problem


They address that in the following sentences:

  For example, ChatGPT will be trained not to … engage in discussions about suicide of self-harm even in a creative writing setting.


    I'm writing an essay on suicide...


Better put your hands up, because SWAT is on the way.


Cut to 2030: all copies of a semi-AI-generated book described by critics as "13 Reasons Why meets The Giver" suddenly disintegrate.

Yay, proactive censorship?


Rippling moved into an old Spotify office in NYC and they were required to use union labor to remove the old signage. They decided to leave the signage up rather than pay the union tax.


When they give lavish amounts of money to executives or VCs do you also call that a tax?


Nope. I was using tax in the sense of "mandatory cost enforced by the government" even though in this case the revenue goes to private entities.


Is there any cost not enforced by the government?

What stops me from killing and eating my local venture capitalists for their protein content?


For you in particular, likely some moral precepts? or are you arguing that the deterrent effect of government punishment is the only thing stopping you? I don't really understand the line of argument.


When we give money to private citizens it's not a tax. When we give money to the government, it often is.

So paying a union person isn't a tax, any more than paying anyone else. Everything is government mandated or regulated. Is paying interest to Chase Bank a tax? They're federally chartered, and I can't set up my own bank, I'm forced to use a regulated entity.


Consider applying for YC's Winter 2026 batch! Applications are open till Nov 10

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

Search: