Has anybody tried it?
There is a lot of emphasis on MacBook Pro in this thread, but I would like to use it with an AMD Halo Strix with 128GB of unified RAM.
I just got the rocm branch compiled and running. Starting with one of the common strix halo rocm toolboxes, just needed to install a few more dependencies to get the repo to build. So far just tried the q2-imatrix model and I'm seeing ~7.32tok/s with a locally bound claude code session. It's pretty unusably slow for agentic coding like this - with it being tens of minutes per round of thinking. But it does seem to be working. Suspiciously amdgpu_top is only showing ~16GB of memory being used. Not sure if this is somehow misreading that.
> Nvidia released the first Shield Android TV in 2015
> it took about 18 months to [create] an entirely new security stack [...] Android updates aren’t actually that much work compared to DRM security, and some of its partners weren’t that keen on re-certifying older products.
> In February 2025, Nvidia released Shield Patch 9.2 [...] That was the Tegra X1 [security] bug finally being laid to rest on the 2015 and 2017 Shield boxes.
This is a real engineering marvel. Everybody else would have just given up entirely long time ago. DRM bugs are in most case practically unrecoverable for products that shipped already (and physically in the hands of the adversary). The incentive to tell to consumers "Ditch that product you bought from us 2 years ago, and buy the more recent hardware revision or successor" is extremely strong.
This really feels like a platform that is maintained with pride and love by the nvidia engineering teams (regardless of one's opinion about DRM per se).
> They added auto-playing, full screen video ads to the home screen.
I'm pretty sure this is actually Google's fault (even Sony televisions suffer from this bullcrap). Unlike phone Android, Google TV (yes, that's the official name now) enforces certain "standards", one of them is this bullcrap.
Who cares who delivered the actual bytes or who initiated the change, the matter of fact is that people buy a device from one company, then the company is responsible for the experience they deliver while it's supported. Since they chose Android, they're responsible for the experience you get when using the stuff you buy from them.
I'd never complain to the maker of a compressor when it dies in a fridge, I'll complain to the one I bought the fridge from. Not sure why we're so adamant on thinking differently regarding computers. NVIDIA might blame Google internally, but feels like consumers are right to be pissed off about NVIDIA changing (or being OK with someone else changing) their experience in a product they bought from NVIDIA.
Because you're on a forum called Hacker News and we pride ourselves on being smart enough to understand the details of systems, and using that to our advantage. Nvidia as a corporation isn't some nobody compared to Google, but do you really think their engineering team has the clout to get Google's advertising arm to bend to their will?
I use an old Amazon FireTV Stick on an old LG LED TV (semi-smart), and neither of them bug me with such fullscreen ads, unless I opt to watch MX content on Amazon Prime (MX is basically third-party ads-funded free OTT content; Amazon Prime requires subscription, and even its standard subscription has occasional ads for Prime content, though Amazon Prime also has a premium pricing tier for ads-free content).
I don't face such third-paety ads nonsense on Netflix and Disney+ (yet), at least on this old FireTV and old LG TV.
Unskippable irrelevant annoying ads and privacg concerns are the main reasons I still steer clear of "smart" TVs.
My Samsung Frame TV shows ads in the app bar and you cannot disable/remove them. They can’t even use the Google excuse because the TV runs Samsung’s OWN TizenOS.
No. The Apple TV _service_ does, and you can configure that service to be some kind of weird god service if you want. But you can also treat that service like any other normal service, one that only comes up if you launch it. In that case, the home screen is just a straight icon grid with no kerfuffle.
Yes, one or two, and not annoying (not trying to grab your attention). No ads for toothpaste or cars.
Apple TV is not the solution for purists who cannot handle anything that can be construed as an ad. It’s a great solution for those who just want to browse and watch content without distracting ads everywhere.
The Apps in the home row on Apple TV will have fullscreen promotions when the home row is along the bottom of the screen. If you set your home row apps with care, the fullscreen previews will not be ads (i.e. Photos will do a slideshow of your photos, Jellyfin just pulls random images from its/your own movie library metadata, etc.).
The only customers who care about DRM are the suppliers; not the users. Force the user to not be able to play DRM content, and they'll end up pirating.
Furthermore, I never demanded a new Android TV version. All I wanted was security fixes, not Google's new shitty launcher. I'd never have bought the product if it contained the current launcher.
They did this with the switch 1 too they were just less well remembered because that subsequently got re-hacked. They lost the ARM trust zone keys and rebuilt the entire DRM stack on the HDCP keys which had been provisioned but they were not using.
I don't think that a 100% anonymous attestation protocol is what most people need and want.
It would be sufficient to be able to freely choose who you trust as proxy for your attestations *and* the ability to modify that choice at any point later (i.e. there should be some interoperability). That can be your Google/Apple/Samsung ecosystem, your local government, a company operating in whatever jurisdiction you are comfortable with, etc.
As mentioned a few days ago, this post mainly covers a gpg problem not a PGP problem.
I recommend people to spend some time and try out sequoia (sq) [0][1], which is a sane, clean room re-implementation of OpenPGP in Rust. For crypto, it uses the backend you prefer (including openssl, no more ligcrypt!) and it isn't just a CLI application but also as a library you can invoke from many other languages.
It does signing and/or encryption, for modern crypto including AEAD, Argon2, PQC.
Sure, it still implements OpenPGP/RFC 9580 (which is not the ideal format most people would define from scratch today) but it throws away the dirty water (SHA1, old cruft) while keeping the baby (interoperability, the fine bits).
But if you use the modern crypto stuff you loose interoperability, right? What is the point of keeping the cruft of the format if you still won't have compatability if you use the modern crypto? The article mentions this:
> Take AEAD ciphers: the Rust-language Sequoia PGP defaulted to the AES-EAX AEAD mode, which is great, and nobody can read those messages because most PGP installs don’t know what EAX mode is, which is not great.
Other implementations also don't support stuff like Argon2.
So it feels like the article is on point when it says
> You can have backwards compatibility with the 1990s or you can have sound cryptography; you can’t have both.
When you encrypt something, you are the one deciding which level of interoperability
you want and you can select the crypto primitives matching capabilities you know you recipient reasonably have. I don't see anything special with this: when you run a web service, you also decide if you want to talk to TLS 1.0 clients (hopefully not).
sequoia's defaults are reasonable as far as I remember. It's also bit strange that the post found it defaulted to using AEAD in 2019 when AEAD was standardized only in 2024 with RFC 9580.
But the elephant in the room is that gpg famously decided to NOT adopt RFC 9580 (which Sequoia and Proton do support) and stick to a variant of the older RFC (LibrePGP), officially because the changes to the crypto were seen as too "ground-breaking".
I think GP’s point isn’t that you don’t have the freedom to decide your own interoperability (you clearly do), but that the primary remaining benefit of PGP as an ecosystem is that interoperability. If you’re throwing that away, then there’s very little reason to shackle yourself to a larger design that the cryptographic community (more or less) unanimously agrees is dangerous and antiquated.
It is not a coincidence that most of the various proposed alternatives to PGP (signal, wormhole, age, minisign, etc) are led by a single golden implementation and neither support nor promote community-driven specifications (e.g., at the IETF).
Over the decades, PGP has already transitioned out of old key formats or old crypto. None of us is expecting to receive messages encrypted with BassOmatic (the original encryption algorithm by Zimmermann) I assume? The process has been slow, arguably way slower than it should have after the advancements in attacks in the past 15 years (and that is exactly the crux behind the schism librepgp/opengpgp). Nonetheless, here we are, pointing at the current gpg as "the" interoperable (yet flawed) standard.
In this age, when implementations are expected (sometimes by law) to be ready to update more quickly, the introduction of new crypto can take into account adoption rates and the specific context one operates in. And still, that happens within the boundaries of a reasonably interoperable protocol.
TLS 1.3 is a case in point - from certain points of view, it has been a total revolution and break with the past. But from many others, it is still remarkably similar to the previous TLS as before, lots of concepts are reused, and it can be deemed as an iteration of the same standard. Nobody is questioning its level of interoperability, and nobody is shocked by the fact that older clients can't connect to a TLS 1.3-only server.
You're right, it's not a coincidence. The track record of standards-body-driven cryptography is wretched. It's why we all use WireGuard and not IPSEC. TLS 1.3 is an actually good protocol, but it took for-ev-er to get there, and part of that process involved basically letting the cryptographers seize the microphones and make decisions by fiat in the 1.2->1.3 shift (TLS 1.3 also follows a professionalization at CFRG). It's the exception that proves the rule. It's contemporaneous sibling is WPA3 and Dragonfly, and look how that went.
I wrote the post and object to the argument that it primarily covers GnuPG issues.
But stipulate that it does, and riddle me this: what's the point? You can use Sequoia set up for "modern crypto including AEAD", yes, but now you're not compatible with the rest of the installed base of PGP.
If you're going to surrender compatibility, why on Earth would you continue to use OpenPGP, a design mired in 1990s decisions that no cryptography engineer on the planet endorses?
You're missing my point. I agree that you can use Sequoia to communicate between peers also using Sequoia. But you're no longer compatible with the overwhelming majority of PGP deployments. So what's the point? Why not just use a modern tool with that same group of peers?
For other security mechanisms, like PKI, things are more complicated (and inefficient).
And one can argue that even if in theory the above gives you better security margin, the whole system becomes more complicated, and it may be practically less secure because of the additional moving parts. That is why there is no unanimous consensus: agencies in Europe recommend it, but the NSA does not.
Finally, note that the the 3rd standard (SLH-DSA), is PQC but it is based on old and well-understood standards (SHA2/SHA3), so it can arguably be used by itself.
I like it, because it is indeed nice to have a NIST-backed construction.
But at the same time, it is disappointing that you get locked out of several niceties of NIST KDFs, such as label and context. I get that they are sacrificed to minimize the number of AES calls, but still I would prioritize strong cryptographic separation over just a few saved AES calls, especially for messages longer than a few hundred bytes.
Finally, *random* GCM nonces longer than 96 bits are definitely misunderstood and bring better guarantees than 96 bits nonces [1]. But of course, if you can derive a fresh key for every message, that's definitely to prefer.
You assume that boolean operations are constant time, and whether that holds depends on the uarchitecture and how sophisticated the optimization layers are (e.g. nothing prevents the compiler or even the CPU from short-circuiting the OR as soon as the first XOR is non-zero).
Computing a MAC of the two input values under a once-off random key is actually much stronger.
As a matter of that, this highlights that the goal is to lower the SNR for the attacker, and constant time computation is only one of the two non-exclusive ways to achieve that, the other being adding sufficient noise to their measurements.
As it is written in the parent article, the verification of the code generated by the compiler is mandatory whenever a constant-time algorithm is written in a high-level language.
The short-circuiting could be prevented e.g. by storing the result in a volatile variable before testing if it is null, because the compiler cannot assume that the tested value is the same that has been written previously. Nevertheless, it is better to just check the generated assembly code and disable optimizations for that function or use inline assembly, if necessary.
The binary Boolean operations have been executed in constant-time in all electronic computers, since those made with vacuum tubes until now.
It is impossible to make them execute in variable time (because they are independent for all bit pairs and they must be executed for all of them), unless you do this on purpose, by inserting unnecessary delays that are dependent on the operands.
For no other operations implemented in hardware it is as certain that they are done in constant time. Even for word additions, it would be possible to make an implementation where the time to propagate the carry until the most significant bit would be variable, depending on the pattern of "1" bits in the operands, but no mainstream CPU has used such adders during the last half of century. Such adders have been used only in some early computers with vacuum tubes, which used serial adders instead of the parallel adders used in all modern CPUs.
Your argument boils down to "all hardware implementations so far in history never optimized word boolean operations so future implementations will keep doing so".
I think that is just an assumption and I would not take that risk for high security applications, unless for specific CPU models where the behavior is measured in practice (certainly not by just auditing the assembly, which is by itself already too high-level). After all, never in history we had such a level of sophistication.
Right now, all zero register values can lead to speed gains, so they could be obversable. But both ARM and Intel latest ISA have introduced flags that permit the future CPUs to perform operations with data-dependent timing. Boolean operations are officially marked as potentially affected by that flag.
No, my argument is that it is impossible to optimize the binary Boolean operations in a way in which they will have variable execution time.
Moreover, they are the only commonly encountered operations that are implemented in hardware and for which this is unconditionally true.
Therefore they are the first choice for the implementation of any constant-time algorithm.
Most modern CPUs have many other operations that are executed in constant time, like additions and subtractions, but for those, variable-time implementations are also possible.
As I have already said, for any bitwise binary boolean operation, the elementary operations having a pair of bits as operands are independent, so there exists no way of performing the complete operation without doing all the sub-operations, unlike for other simple operations, like additions, shifts or rotations, where it is possible to detect conditions that allow an earlier completion of the operation.
The hardware implementation can do the sub-operations sequentially or concurrently, in any order, but it always must execute all of them, regardless of the values of the input operands, so they will take the same time.
Only in an asynchronous CPU and only for certain kinds of logic gate implementations it is possible to have a completion time for a binary Boolean operation that varies with the values of the input bits, but for the most common asynchronous circuit synthesis methods, which use dual rail encoding for bits, i.e. each bit is encoded as a pair of complementary bits, the binary boolean operations remain constant-time, like in the normal synchronous CPUs.
Let's assume that you have a simple XOR between two registers.
If the CPU can pre-label a register as having no bits set (and they can or speculate on it), during scheduling, it could theoretically simply drop the XOR, transfer or rename the relevant register which may lead to a tiny but different timing that can be measured and exploited.
That is just one simple counter-example that shows how the assumptions you present are not necessarily valid on modern complex CPUs. Many more are examples are possible, which is of course not to say that they are implemented today. But they could be implemented in the future (without us knowing, and again, both ARM and Intel are explicit on that), therefore the security of a security-sensitive piece of code should not solely rely on that.
But then, the vast majority of the affected libraries in that page don't use GMP at all, but their own custom implementation (including openssl).
In reality, RSA signing with blinding will make any implementation (including those based on GMP) resistant to side channel attacks, targeted at the private key.
What most of these library tripped over in that case, is the treatment of the plaintext in a side channel-safe way after the private key operation. For instance, just the simple conversion of an integer to a byte string can be targeted.
Has anybody tried it? There is a lot of emphasis on MacBook Pro in this thread, but I would like to use it with an AMD Halo Strix with 128GB of unified RAM.